THESIS

Z.

MICHIGAN STATE UNIVERSITY LIBRARIES

I I (If!!!Ill/ll!!!)Ill/N/IIHWHHIIIll/II

3 1293 01592 8561

I!

 

 

 

 

 

 

 

 

 

 

 

 

 

 

LIBRARY
'- Michigan State
3 University

 

 

This is to certify that the

dissertation entitled

INCIDENT DETECTION ALGORITHM DEVELOPMENT ON SIGNALIZED
URBAN ARTERIAL STREETS

presented by

Jung-Taek Lee

has been accepted towards fulfillment
of the requirements for

Ph . D . degree in Civil Engineering

 

 

Mm 0-73/47

Major professor

Date 4/3/97

MS U i: an Affirmatiw Action/Equal Opportunity Institution 0-12771

 

PLACE II RETURN BOX to romanthb checkout from your record.
TO AVOID FINES retum on or before dd. duo.

DATE DUE DATE DUE DATE DUE

      

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

L.

 

 

 

Um

 

-i

E

 

 

 

 

 

 

MSU In An Afflrmutm ActioNEan Opportunlty Intuition
W

F’T’T’]

m1

INCIDENT DETECTION ALGORITHM DEVELOPMENT
ON SIGNALIZED URBAN ARTERIAL STREETS

By

J Ling-Tack Lee

A DISSERTATION
Submitted to
Michigan State University

in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Civil and Environmean Engineering

1997

ABSTRACT

INCIDENT DETECTION ALGORITHM DEVELOPMENT
ON SIGNALIZED URBAN ARTERIAL STREETS

By

J ung-Taek Lee

Congestion due to incidents on urban freeways and arterial streets is an important
issue in urban mobility. It is also an area with potential for successful implementation of
ITS (Intelligent Transportation Systems) technology. For many years, traffic engineers
have studied techniques to detect an incident on urban freeways. However, since the
traffic flow pattern on arterial streets is more dynamic than on freeways, algorithms to
detect incidents on arterials have been more difficult to define.

Existing freeway and signalized arterial street incident detection algorithms have been
investigated to determine the merits of each in their potential use on urban arterial streets.
Based on the literature review, a Kalman Filtering Predictor Algorithm was modified and
used to recursively filter and update aggregate traffic control variables and to eventually
determine an incident state.

A test using non-incident arterial street data showed good tracking ability of selected
traffic variables over time. A second test, using data from an incident on an arterial street,
confirmed that the algorithm developed in this study has good potential for arterial street

incident detection.

A simulation using TRAF-NETSIM was then conducted to test the algorithm for
incident-free and various incident states. The algorithm performance measures were a
false alarm rate (FAR) of 0.319 percent per hour per station, a detection rate (DR) of 100
percent with a mean detection time (MDT) of 4.93 minutes for one-lane blockages. For
the high traffic volume direction, the FAR was reduced to 0.111 percent per hour per
station. The MDT varied from 4.8 minutes for long links to 2.7 minutes for short links.

A discussion of the algorithm and recommendations for further research to improve

incident detection on urban arterial streets is also included in this report.

ACKNOWLEDGMENTS

I would like to express my sincere appreciation to my advisor, Dr. William C. Taylor,
for his continuous advice, guidance and support. He is not only a teacher who helped me
realize what “engineering” is but also an example of my life.

I want to thank my guidance committee: Dr. Thomas Maleck, Dr. Thomas Wolff and
Dr. Martin F ox for their invaluable advice and suggestions that helped me complete this
work. I would like to express my special gratitude to Dr. Thomas Maleck. This work
would not be possible without his continuous support.

I would like to show my gratitude to my parents for their concerns and continuous
support. Their support was a great help through my school years. My thanks to my
brothers and sisters who encouraged and supported me to reach my goals. My mother-in-
law and brothers-in-law are ones that I can not thank enough for their support.

My wife, Jeongrim, and my daughter, Jean, are a reason for me to have been able to
finish this work. It was great luck for me to meet and have Jeongn'm in my life.

My thanks also go to my colleagues at Michigan State University who helped and

continuously supported me in my school years although I can not refer to them all here.

TABLE OF CONTENTS

LIST OF TABLES ......................................................................................................... viii
LIST OF FIGURES .......................................................................................................... x
CHAPTER 1 INTRODUCTION ...................................................................................... 1
1.1 Statement of the Problem ............................................................................................... 1
1.2 Characteristics of an Incident ......................................................................................... 3
1.2.] Definition of an Incident ................................................................................. 3
1.2.2 Impact of Incident ........................................................................................... 4
1.2.3 Trafiic Variable Changes During a Freeway Incident .................................... 5
1.3 Objective and Scope of the Research ............................................................................. 8
1.4 Research Methodology ................................................................................................... 8
CHAPTER 2 EXISTING INCIDENT DETECTION ALGORITHM REVIEW ...... 10
2.1 Existing Freeway Incident Detection Algorithms ........................................................ 10
2.1.1 Comparative Algorithms (Pattern Recognition Algorithms) ........................ 12
2.1.2 Statistical Algorithms .................................................................................... 14
2.1.3 Dynamic Model-Based Technique Algorithm .............................................. 22
2.1.4 High Occupancy (HIOCC) Algorithm .......................................................... 23
2.1.5 McMaster Algorithm (Catastrophe Theory Algorithm) ................................ 24
2.1.6 Advanced Technology-Based Algorithms .................................................... 25
2.1.7 Video Image Processing Technique-Based Algorithms ................................ 26
2.1.8 VRC Sensor—Based Incident Detection Algorithm ....................................... 28
2.2 Existing Urban Arterial Street Incident Detection Algorithms .................................... 29
2.2.1 Automatic Incident Detection (AID)
within Urban Traffic Control System .................................................................... 29
2.2.2 A Dynamic Real-Time Incident Detection System
for Urban Arterial System ...................................................................................... 30
2.2.3 Neural Networks for Real-Time Data Fusion Method .................................. 32
2.2.4 Video Image Processing Technique-Based Algorithm ................................. 33
2.2.5 Intersection Accident Detection Algorithm .................................................. 34
2.2.6 Smart Corridor Arterial Incident Detection System ...................................... 35
2.3 Summary of Existing Freeway and Urban Arterial Street
Incident Detection Algorithms ........................................................................................... 36

Chapter 3 INCIDENT DETECTION ALGORITHM DEVELOPMENT ................. 40

3.1 Kalman Filtering Theoretical Background ................................................................... 40
3.1.1 Optimal Estimate and Orthogonal Projection ............................................... 40
3.1.2 Linear Optimal Filters ................................................................................... 42
3.1.3 Kalman Filtering Algorithm .......................................................................... 44
3.2 Incident Detection Algorithm Development using the Kalman Filtering .................... 46
3.2.1 Considerations for Signalized Arterial Street Incident Detection ................. 47
3.2.2 Linear State-Space Model ............................................................................. 48
3.2.3 A Recursive Incident Detection Algorithm ................................................... 50
3.2.4 Incident State Determination ........................................................................ 51

Chapter 4 INCIDENT DETECTION ALGORITHM TESTS ON ARTERIAL

STREET FIELD DATA .................................................................................................. 54
4.1 Application to Incident-Free Field Data ...................................................................... 55
4.1.] Linear State-Space Model for the Data ......................................................... 55
4.1.2 Linear Measurement Model for the Data ...................................................... 56
4.1.3 Initialization .................................................................................................. 58
4.1.4 Results of Incident-Free Field Data Application .......................................... 59
4.2 Application to Incident Data Set .................................................................................. 68
Chapter 5 TRAF-NETSIM SIMULATION TEST ....................................................... 74
5.1 Simulation Test for the Algorithm ............................................................................... 74
5.2 Description of the Arterial Street Network .................................................................. 74
5.3 Measures of Effectiveness ........................................................................................... 75
5.4 Incident-Free Data Set Description .............................................................................. 77
5.5 Incident Data Set Description ...................................................................................... 78
5.5.1 Link Selection for Incident Generation ......................................................... 78
5.5.2 Incident Generation in TRAF-NETSIM ....................................................... 84
5.5.3 Various Incident Generation ......................................................................... 85
5.6 Results of the Incident-Free Test ................................................................................. 90
5.7 Results of the Incident Test .......................................................................................... 95
5.7.1 Results of One-Lane Blockage in Link 34 .................................................... 95
5.7.2 Results of F ull-Link Blockage in Link 34 ..................................................... 96
5.7.3 Results of One-Lane Blockage in Link 45 .................................................... 97
5.7.4 Results of F ull-Link Blockage in Link 45 ..................................................... 97
5.8 A Sensitivity Test for Various Confidence Limits ....................................................... 97
5.8.1 False Alarm Changes over Different Limits ................................................. 98
5.8.2 Detection Rate and Mean Detection Time Changes
over Different Limits ............................................................................................. 99
5.8.3 A Detailed Sensitivity Test for Link 34 ...................................................... 102
5.8.4 Determination of Recommended Limits ..................................................... 115
5.9 Parameter Estimation for Different Conditions ......................................................... 119

vi

5.10 Summary of TRAF-NETSIM Simulation Results ................................................... 125

5.10.1 A Test at 95 Confidence Limits (2.0 0 Limit) ......................................... 125
5.10.2 A Sensitivity Test for Various Confidence Limits .................................... 126
5.10.3 Parameter Estimation for Different Conditions ........................................ 127
Chapter 6 CONCLUSIONS AND DISCUSSION ....................................................... 128
6.1 Conclusions ................................................................................................................ 128
6.1.1 Summary of the Research ........................................................................... 128
6.1.2 Comparisons with Previous Studies ........................................................... 132
6.1.3 Conclusions of the Research ....................................................................... 135
6.2 Discussion of the Research and Further Research ..................................................... 135
APPENDICES
APPENDIX A. DESCRIPTIONS OF TRAFFIC VOLUME FOR EACH NODE IN THE
NETWORK ...................................................................................................................... 139
APPENDIX B. DESCRIPTION OF GEOMETRIC CONFIGURATIONS FOR
NETWORK LINKS ......................................................................................................... 140
APPENDIX C. DESCRIPTION OF SURFACE STREET TURNING
MOVEMENTS ................................................................................................................ l4]
APPENDD( D. DESCRIPTION OF PRETIMED SIGNAL INTERVAL FOR EACH
NODE .............................................................................................................................. 143
LIST OF REFERENCES .............................................................................................. 144

vii

LIST OF TABLES

Table 4.1 MSEs for Flow and Speed in Lane 1 and Lane 2 ............................................... 59
Table 5.1 LOS in Node 3, 4, and 5 .................................................................................... 79
Table 5.2 Incidents in Link 34 Description ........................................................................ 87
Table 5.3 Incidents in Link 45 Description ........................................................................ 89
Table 5.4 Results of Incident Free Test for 95 Percent Confidence Limits ....................... 91
Table 5.5 Distribution of False Alarms over Detector Stations ......................................... 92
Table 5.6 One-Hour Traffic Statistics at Detector Station 20 ............................................ 93
Table 5.7 One-Hour Traffic Statistics at Detector Station 5 .............................................. 94
Table 5.8 False Alarm Statistics Changes for 2.0, 2.5, and 3.0 a“ ................................. 98

Table 5.9 Detection Rates of One-Lane Blockage over 2.0, 2.5, and 3.0 on Changes
(percent) ........................................................................................................................... 100

Table 5.10 Detection Rates of Full-Link Blockage over 2.0, 2.5, and 3.0 0‘“ Changes
(percent) ........................................................................................................................... 100

Table 5.11 Mean Detection Time of One-Lane Blockage over 2.0, 2.5, and 3.0 on
Changes (minutes) ........................................................................................................... 101

Table 5.12 Mean Detection Time of Full-Link Blockage over 2.0, 2.5, and 3.0 0'“
Changes (minutes) ........................................................................................................... 101

Table 5.13 False Alarm Changes over Different Intervals ............................................... 105

Table 5.14 Detection Rate Changes for One-Lane Blockage in Link 34 over Different
Intervals (percent) ............................................................................................................ 106

Table 5.15 Mean Detection Time Changes for One-Lane Blockage in Link 34 over
Different Intervals (minutes) ............................................................................................ 106

viii

Table 5.16 Detection Rate Changes for Full-Link Blockage in Link 34 over Different

Intervals (percent) ............................................................................................................ 107
Table 5.17 Mean Detection Time Change for Full-Link Blockage in Link 34 over
Different Intervals (minutes) ............................................................................................ 107
Table 5.18 Slope Changes of False Alarms over Different Intervals ............................... 117
Table 5.19 Slope Changes of DR and MDT for One-Lane Blockages over Different
Intervals (Both Stations) .................................................................................................. 117
Table 5.20 Slope Changes of DR and MDT for Full-Link Blockages over Different
Intervals (Both Stations) .................................................................................................. 118
Table 5.21 Results of Incident Detection Performance at Recommended Intervals ........ 118
Table 5.22 Characteristics of Selected Links ................................................................... 119
Table A Description of Traffic Volume for Each Node in the Network ......................... 139
Table B Description of Geometric Configurations for Network Links ........................... 140
Table C Description of Surface Street Turning Movements ............................................ 141
Table D Description of Pretimed Signal Interval for Each Node .................................... 143

 

ii

F15

LIST OF FIGURES

Figure 1.1 Freeway Traffic Flow Changes due to an Incident ............................................. 7
Figure 3.1 Recursive Kalman Filtering Steps .................................................................... 46

Figure 3.2 Incident Detection Algorithm using the Kalman Filtering Recursive

Procedure ........................................................................................................................... 53
Figure 4.1 Data Fitting for True and Observed Flow and Speed ....................................... 61
Figure 4.2 Predicted and Filtered Estimates of Flow and Speed in Lane 1, and 2 and 95
Confidence Intervals .......................................................................................................... 62
Figure 4.3 Scatter Plots of Residuals for Traffic Flow and Average Speed in Lane 1 and
Lane 2 ................................................................................................................................. 64
Figure 4.4 (D(i, 1') Changes over Time ............................................................................... 66
Figure 4.5 Q(i, i) Changes over Time ................................................................................. 67
Figure 4.6 A Priori Error Covariance Changes over Time ................................................ 67
Figure 4.7 Incident Detection Procedure on Field Data ..................................................... 70
Figure 5.1 Simulation Network .......................................................................................... 80
Figure 5.2 Node 3 Geometric and Hourly Traffic Volume ................................................ 81
Figure 5.3 Node 4 Geometric and Hourly Traffic Volume ................................................ 82
Figure 5.4 Node 5 Geometric and Hourly Traffic Volume ................................................ 83
Figure 5.5 Number of False Alarms for the Network over 032/, Changes ....................... 108
Figure 5.6 Number of False Alarms per Station for the Network over 0'“ Changes ..... 108
Figure 5.7 False Alarm Rate per Station for the Network over 0'“ Changes ................ 109

Figure 5.8 Number of False Alarms for High Traffic Direction over 0'“ Changes ....... 109

Figure 5.9 Number of False Alarms per Station for High Traffic Direction over 0'”
Changes ............................................................................................................................ 110

Figure 5.10 Number of False Alarms per Station for High Traffic Direction over 032/.
Changes ............................................................................................................................ 110

Figure 5.1] Detection Rate Changes for One-Lane Blockage in Link 34 over Interval
Changes ............................................................................................................................ 111

Figure 5.12 Mean Detection Time (minutes) for One-Lane Blockage in Link 34 with Bars
of One Standard Deviation (Upstream Station) ............................................................... 111

Figure 5.13 Mean Detection Time (minutes) for One-Lane Blockage in Link 34 with Bars
of One Standard Deviation (Downstream Station) .......................................................... 112

Figure 5.14 Mean Detection Time (minutes) for One-Lane Blockage in Link 34 with Bars
of One Standard Deviation (Both Upstream and Downstream Stations) ......................... 112

Figure 5.15 Detection Rate Changes for Full-Link Blockage in Link 34 over Interval
Changes ............................................................................................................................ 113

Figure 5.16 Mean Detection Time (minutes) for F ull-Link Blockage in Link 34 with Bars
of One Standard Deviation (Upstream Station) ............................................................... 113

Figure 5.17 Mean Detection Time (minutes) for Full-Link Blockage in Link 34 with Bars
of One Standard Deviation (Downstream Station) .......................................................... 114

Figure 5.18 Mean Detection Time (minutes) for Full-Link Blockage in Link 34 with Bars
of One Standard Deviation (Both Stations) ..................................................................... 114

Figure 5.19 Parameter (CD) Patterns in Different Geometric and Traffic Conditions ..... 121

xi

Chapter 1

INTRODUCTION

1.1 Statement of the Problem

The congestion on urban freeways and arterial streets is an important issue in urban
mobility. It is also an area with potential for successful implementation of ITS (Intelligent
Transportation Systems). A large component of ITS is directed toward congestion
detection and management.

The traditional way to resolve the congestion problem is through new road
construction and road-widening programs. In Oakland County, Michigan, for instance,
over one billion dollars was identified for road construction needs during the next decade.
Almost seventy percent of that investment was for improving mobility in congested areas
(Barbaresso 1993). However, this type of solution is very expensive and can not always
resolve the problem.

For many years, traffic engineers have studied techniques to reduce the impact of
incidents on urban freeways. They have also developed automatic incident detection
algorithms and responsive control systems. However, they were all developed for
freeways. The traffic flow pattern in urban arterial streets is different from that of

freeways.

2

The geometric and functional aspects of arterial streets are quite different from
freeways. From an incident detection perspective, the traffic flow pattern on arterial
streets makes it difficult to implement an incident detection program. While the traffic
flow on freeways is continuous and uninterrupted, the traffic flow on arterial streets is
often interrupted due to various external sources such as side street traffic, slow moving
vehicles, on-street parking, etc. Above all, turning vehicles prevent the direct comparison
of the upstream and the downstream condition, which is a commonly used incident
detection method on freeways. A signal system also prevents the construction of a short-
time data aggregating time interval because of stop-and-go patterns in a signal cycle.

A few recent studies have attempted to develop an incident detection algorithm for
urban arterial streets. These efforts are, however, still in the conceptual stage. Most of the
algorithms have not yet been applied to real field operations.

Some of these studies approach the problem similar to the way it is done in freeway
incident detection applications. They detect abrupt changes in traffic features to identify
an incident.

The other approach to incident detection on an arterial street is through using new
technologies. However, not all the existing arterial streets are equipped with probe
vehicles and video cameras required to obtain the necessary information.

There are aspects of freeway models and methodologies that may be applicable to
detecting incidents on arterial streets. In this study, existing freeway and arterial street
incident detection algorithms are reviewed to establish the usefulness of freeway based

algorithm aspects for arterial streets.

3

A real-time adaptive linear Kalman filtering approach was selected for this study
because it is simple, and operates in real-time to reflect dynamic arterial street traffic
flow. The use of flexible thresholds by using prediction intervals constructed by a priori
error covariance and the use of raw data from a single detector station are also advantages
of this approach.

Field collected arterial data and simulated data are used to validate the algorithm.
TRAF-NETSIM simulation is used to generate incident data sets due to the absence of

sufficient arterial incident data.

1.2 Characteristics of an Incident
In this section, characteristics of an incident are investigated. A definition of an incident
is postulated. Impacts on roadways are investigated and traffic variable changes resulting

from an incident are investigated since they are primary clues to the presence of an incident.

1.2.1 Definition of an Incident

An incident is a relative term since some may consider a specific event an incident, and
some may not. There is also no absolute quantitative definition of an incident in terms of
incident duration and roadway capacity reduction. Traditionally an incident is defined as
“an accident, breakdown, spilled load or other random event that reduces the capacity of
the roadway (Judycki, and Robinson 1991: 359).” Other definitions are “any occurrence
that affects roadway capacity, either by obstructing travel lanes or by causing gawkers
block (Giuliano 1989: 387)” and “a spill, breakdown, accident, or any other extraordinary

event that causes congestion and delay by restricting normal traflic flow (Urbanek, and

4

Rogers 1978: 1).” A red signal light and a slow truck in a no passing zone were also defined
as “simple incidents” (Wirasinghe 1978). As can be understood by these definitions, an
incident can be any event on a highway that negatively affects the traffic flow. However, it
is worthy to note what “event” should be considered as an incident. As Han (1995) argued,
an incident should be detectable and worth detecting. This is true especially in arterial street
conditions since there are many “incident—like” events. On-street parking vehicles and slow
buses could obstruct traffic flow but they are not incidents. A temporary stop to load or
unload a passenger may block traffic flow but it would not be long enough to affect traffic
flow significantly. A signalized intersection interrupts traffic flow regularly but is not an
incident. Therefore, it is advantageous to define an incident as any event on a highway that

lasts long enough to reduce roadway capacity over an extended time period.

1.2.2 Impacts of Incidents

The impact of an incident can be viewed in various perspectives: the traffic operation
perspective, the economic perspective, and the safety perspective.

In operations, the capacity reduction is more than the physical reduction in the available
lanes. For instance, in a three lane section, an accident in one lane reduces physical capacity
by 33 percent, but operational capacity is reduced by 47 to 48 percent or 50 percent. The
two lane-blocking accident in the three lane section reduces 79 percent of operational
capacity. In a two lane section, one lane blocking resulted in operational capacity reductions
of 60 percent to 79 percent while the physical reduction was 50 percent (Goolsby 1971).

In economic terms, a recent F HWA study shows that incidents account for 60 percent of

all freeway congestion in the 37 largest urban areas. Accordingly, user costs exceed five

5
billion dollars a year in delay and wasted fuel, based on 1984 data. The FHWA study also

shows that 70 percent of urban freeway congestion is accounted for by incidents, and will
cost users in excess of 35 billion dollars a year by the year 2005 in the absence of significant
improvements (Judycki, and Robinson 1991).

From the perspective of safety, pedestrian accidents and accidents involving striking a
parked vehicle resulted in 19 to 20 percent of all interstate system fatalities, almost 800
fatalities each year, during the years 1983-1985 (Judycki, and Robinson 1991). Other
studies indicated that 13 percent of accidents (106 out of 812 accidents) were secondary

accidents (Lari, Christianson and Porter 1982).

1.2.3 Traffic Variable Changes During a Freeway Incident

Traffic flow associated with an incident can be quantitatively expressed by traffic
features. These traffic features are used as control variables in automatic incident
detection algorithms and the abrupt change in these variables are used to activate these
algorithms. These features are usually traffic volume (flow), occupancy, density, speed,
and travel time. During a freeway incident, the traffic flow features can be categorized in

four regions as shown in Figure 1.1 (Hughes Aircraft and JHK & Associates 1992).

Flow region A is far upstream from the incident where traffic moves at normal speeds
with normal density.
Flow region B is the area located immediately behind the incident where vehicles are

queuing. In this region, low speed occurs and higher density is experienced.

6

Flow region C is the region immediately downstream of the incident, where traffic is
flowing at a metered rate due to the restricted capacity caused by the incident. The density
of region C is lower than the normal density and the speed may be slightly greater.
However, the speed of the downstream immediately behind the incident is not necessarily
higher since drivers slow down to look at the incident in some cases.

Flow region D is far downstream from the incident where traffic is not disturbed any
more by the incident. Like region A, traffic in region D is flowing at normal density and

speed.

Although not shown in Figure 1.1, occupancy is the traffic flow variable used in most
incident detection algorithms rather than density. Occupancy is defined as the percentage
of time a detector is occupied by a vehicle (or vehicles) during the reporting interval.
Since occupancy is raw data and most properly corresponds to the point measures of
speed and volume that are available, the use of this variable is beneficial. The use of
density requires the measurement over a section of a roadway in order to calculate the
number of vehicles per mile, and the converting of measured occupancy to density results
in an unnecessary degree of uncertainty. In addition, associated with speed, travel time
can be used as the indication of the onset of congestion as it is inversely related with the

speed.

 

UPSTREAM

DOWNSTREAM

 

 

I.’ NORUAL —C--— CONGKS‘DOI -.-

 

“ODENT
LOCATIW

.— uzmo —-.—-— mu». --0-

 

 

 

 

 

 

 

 

 

l anocxw

/////////

 

 

 

1.3 Objective and Scope of the Research
The objective of this research is to develop an incident detection algorithm for urban

arterial streets and test it with field data and simulated data. The scope of the research

includes:

0 Reviews of existing freeway and arterial street incident detection algorithms to
determine advantages and disadvantages for this research;

0 Development of an incident detection algorithm based on the reviews and arterial
street characteristics;

0 Tests of the algorithm with field collected incident-free and incident data to Show the
potential as an incident detection algorithm;

0 Tests of the algorithm with TRAF-NETSIM simulated data on an existing network to
determine the performance of the algorithm;

0 A search for best criteria in terms of evaluation metrics (the false alarm rate, the
detection rate, and the mean detection time); and

o Tests for parameter changes for different geometric and traffic condition.

1.4 Research Methodology

The development of an incident detection algorithm for arterial streets was based on
the freeway and arterial street incident literature reviews. The original Kalman filtering
algorithm was modified and programmed in FORTRAN 77.

The evaluation of the incident detection algorithm was conducted with field collected

data and TRAF-NETSIM simulated data.

9

Incidents were identified by using flexible prediction intervals (limits) of a priori
covariance estimates of the traffic control variables. If the filtered estimates were out of
intervals, then an incident is alarmed.

Field collected data included incident-free data and incident data. Field incident data
were collected by installing experimental detectors on arterial streets. Incident-free data
were used to validate the algorithm and test for flase alarms. A representative set of
incident data used to show how the algorithm determines an incident.

The TRAF-NETSIM model was selected for generating incident data. The simulation
program reflects the surface street traffic and drivers characteristics realistically.
Simulated data were generated based on the geometric and traffic conditions of an
existing arterial street, Washtenaw Avenue in Ann Arbor, Michigan.

Three metrics were used to evaluate the algorithm performance. They are the false
alarm rate, the detection rate, and the mean detection time. Incident-free data well used to
check the false alarm rate.

Two typical links were selected, representing large and small signal spacing, to
generate various types of incidents. The incident data were then tested for the detection
rate and the mean detection time.

By using different prediction intervals, an best incident detection criterion was found.
The inverse relationship between false alarm rate and detection rate are considered in
reaching an best point for these two metrics. Additionally, the mean detection time was

also considered.

Chapter 2

EXISTING INCIDENT DETECTION ALGORITHM REVIEW

Since the Emergency Traffic Patrol (the "Minutemem") was established in 1961 in
Chicago (the first efforts to deal with incident problems), various incident management
systems and automated incident detection systems have been proposed and implemented
(Dudek and Ullman 1992; Roper 1990). Popular methods of detecting incidents vary from
observations and reporting by motorists or the police to closed-circuit television and
cellular phone hotlines. These methods are based on direct observations. The alternative
is to use indirect methods. This type of detection is usually accomplished by using
sensors (e.g., loop detectors) placed along the roadway to obtain and process traffic flow
data. This section will describe the current freeway and arterial street incident detection

algorithms, and their performance characteristics.

2.1 Existing Freeway Incident Detection Algorithms

Current freeway incident detection algorithms have been placed in categories based on
their theoretical aspects and operating characteristics (Balke 1993; Chen, and Chang
1992; Hughes Aircraft, and JHK & Associates 1992; Stephanedes, Chassiakos, and

Michalopoulos 1992). They can be classified as follows.

ll

1. Comparative Algorithms (Pattern Recognition Algorithms)
0 California Algorithm
0 Stream Discontinuity Model Algorithm
0 Pattern Recognition (PATREG) Algorithm
0 All Purpose Incident Detection (APID) Algorithm
2. Statistical Approach Algorithms
0 Standard Normal Deviate (SND) Model
0 Bayesian Approach Algorithm
0 Exponential Smoothing Algorithm
0 Dynamic Model Algorithm (Sakasita et al. 1975)
o Box-Jenkins Technique-Based Algorithm
0 Filtering Technique-Based Algorithm
0 Kalman Predictor Algorithm
3. Dynamic Model-Based Algorithm (Willsky 1980)
4. HIOCC Algorithm (TRRL Algorithm)
5. McMaster Algorithm (Catastrophe Theory Algorithm)
6. Advanced Technology-Based Algorithms
7. Video Image Processing Technique-Based Algorithm
8. VRC Sensor Incident Detection Algorithm
- Headway Algorithm
0 Lane Switch Algorithm

0 Lane-Monitoring Algorithm

12

2.1.1 Comparative Algorithms (Pattern Recognition Algorithms)

The comparative algorithms or the pattern recognition algorithms are based on the
assumption that there is some recognizable continuity in the traffic pattern at two
consecutive sites under steady incident-free conditions. These algorithms detect an
inconsistency in a selective traffic flow variable between adjacent detectors to activate the
incident warning message.

The basic approach of the California algorithm to detect incident is by comparing the
traffic flow data at one detector station and the data from an adjacent downstream station.
The threshold values are pre-determined by a trial-and-error process. This algorithm is
executed for each adjacent pair of sensor stations at regular intervals (e. g., 20 seconds, 30
seconds, or 1 minute). The generalized form of the algorithm is a binary decision tree,
which includes a tree-like arrangement of binary decisions. Traffic features commonly
used for the California algorithms are temporal and spatial average occupancy difference
and relative average difference (Knobel, and Helfenbein 1976; Levin, and Kraus
1979a,b,c; Masters, Lam, and Wong 1991; Payne, Helfenbein, and Knobel 1976; Payne,
and Tigor 1978; Tignor, and Payne 1977). Ten more extended versions of the California
algorithm were developed to provide for more accurate incident detection under various
traffic conditions. Algorithm 7 is the California Algorithm with a check for persistence.
Algorithm 8 is the California Algorithm with a check for a compression wave and
persistence, especially for “stop and go” traffic (Payne, Helfenbein, and Knobel 1976).
Since the effectiveness of these types of algorithms is dependent on the thresholds
chosen, a program, called CALB, was developed, which uses a random number generator

that produces increments to be added to the current threshold vector to produce a new

13
threshold vector for evaluation (Knobel, and Helfenbein 1976; Leven, and Krause 1979a).

Also, an optimal threshold calibration tool, called DAISI (Detection Automatique
d’Incidents par Seuillage Iteratif - Automatic Incident Detection by Iterative
Thresholding), for automatic incident detection algorithms including the California
algorithm was developed. This tool uses usual performance criteria such as the false
alarm rate and the detection rate. The optimization of a threshold is achieved by
minimizing the false alarm rate, subject to a detection rate or by maximizing the detection
rate, subject to a false alarm rate (Cohen 1994; Cohen, and Ketselidou Unknown Date).

In the Stream Discontinuity Model Algorithm, it is assumed that the occurrence of an
incident creates a large difference in traffic measurements at an upstream and downstream
detector. Occupancies at the upstream and downstream detectors are obtained and two
parameters are determined, (1) the temporal difference of upstream and downstream
occupancy by time interval t, and (2) the temporal ratio of downstream occupancy and
upstream occupancy by time interval I. These two parameters are compared with pre-
determined thresholds. The time interval t is estimated from downstream information
such as detector spacing, the length of vehicle and detector, and number of vehicles
counted at the downstream detector (Sakasita et al. 1975).

As an algorithm to test the feasibility of estimating the average journey time between
measuring sites, the Pattern Recognition (PATREG) algorithm was also considered as an
incident detection algorithm (Collins 1977; Collins, Hopkins, and Martin 1979). The
incident is detected when a significant change in the speed (calculated at every 40
seconds by using a cross-correlation procedure) occurs and exceeds pre-determined lower

and upper thresholds for a pre-set number of consecutive seconds.

14
The All Purpose Incident Detection (APID) algorithm is based on the California

algorithm but is characterized by handling different traffic patterns (e.g., light/medium
traffic conditions need a particular algorithm suited for these conditions). A compression
wave test and persistent test are added to reduce false alarm rate in this algorithm as well
(Lam and Tritter 1994; Masters, Lam, and Wong 1991).

The advantage of the comparative algorithm is a simple process to detect an incident.
Complicated calibration is not required to interpret the data from detectors. Rather, raw
data (occupancy) is used to identify an abrupt change in traffic flow between adjacent
detectors. However, the simplicity does not explain dynamic traffic patterns between
adjacent detector stations so the probability of false alarms is high. This could be a fatal
flow on urban arterial streets because the traffic patterns in these roadways are more
complex between detector stations. The use of threshold values between adjacent
detectors is not flexible either. Since these values are obtained by historical observed data
on particular conditions, they need to be re-calibrated under different conditions. The

comparative algorithms is not an adequate strategy to be used on arterial streets.

2.1.2 Statistical Algorithms

Statistical algorithms include the Standard Normal Deviate (SND) Model, the
Bayesian Approach, the Exponential Smoothing Algorithm, the Dynamic Model
Algorithm (Sakasita et al. 1975), the Box-Jenkins Technique Algorithm, the Filtering
Technique Algorithm, and the Kalman Predictor Algorithm.

The Standard Normal Deviate (SND) Model algorithm was developed to circumvent

the weakness of other models, the requirement of each different frequency distribution for

15

the measured traffic variable for defining threshold values for each different traffic
operating conditions and detector stations (Dudek, Messer, and Nuckles 1974). This
algorithm considers the rate of change of the control traffic variable (e.g., occupancy and
kinetic energy that is computed from volume and speed for 3 and 5 min) rather than an
absolute threshold values. The standard normal deviate (SND) reflects the degree to
which the control variable has changed in relationship to the average trends measured
during previous intervals. That is, a large SND value would reflect a major change in
operating conditions on the freeway. This model was also used for a stoppage-wave-
detection algorithm previously studied which detects queues or stoppage waves in the
freeway to alert on-coming motorists (Dudek and Messer 1973).

The Bayesian Approach Algorithm assumes that the normal traffic flow follows its
historical trend (Levin and Krause 1978; Levin and Krause 1979a,d). This algorithm
obtains the frequency distribution functions of certain traffic feature (Z) during incident
(U1) and incident free (U2) situations for a certain geometric and traffic conditions. These
traffic features are measured at either the upstream or downstream detector or at both
located on one of the freeway lanes. If Z is greater than a threshold (21), a possible
incident state is signaled and if Z is less than 2., and incident-free state is signaled. The
probabilities of an incident, P(Ul), occurring based on the history of capacity-reducing
incidents on that freeway then is used to obtain the probability of incident state, P(1), and
the probability of an incident-free state, P(O). The Bayesian consideration is applied to
obtain the probability of having an incident given an incident signal (1) and the
probability of incident-free given a non-incident signal (0). The optimal threshold 21 is

obtained by maximizing the sum of P(incidentll) and P(no-incidenth). The signal can be

16

extended n successive time intervals to provide more reliable information. These
conditional probabilities could be calibrated for any particular freeway section and
specific traffic and environmental conditions, utilizing historical incident data.

The Exponential Smoothing Algorithm uses a short-term forecasting technique for
detecting irregularities of a traffic variable (Cook and Cleveland 1974; Masters, Lam, and
Wong 1991). The variables used are usually volume, occupancy and speed. An incident is
signaled if the tracking signal deviates significantly from a pre-defined threshold value
(zero). The tracking signal is defined as the algebraic sum to the present minute of all the
previous estimate errors divided by the current estimate of the standard deviation. The
estimate of the standard deviation is estimated by the mean absolute deviation, which is
obtained by single (or double) exponential smoothing method. The output of the
algorithm is two incident states: confirmed incident, and incident free.

In the Dynamic Model Algorithm, it is assumed that the arrival pattern of vehicles on a
freeway lane remains essentially unchanged under non-incident conditions. A
longitudinal pair of upstream and downstream detectors in a freeway lane was considered
for input and output traffic information and linearly related in travel time period (I)
between two detectors. Parameters (v,-) for this linear model can be summed to the steady-
state gain (g), which is the ratio of the mean values of traffic information at the upstream
and downstream detectors. Under non-incident conditions, g is assumed to come very
close to 1, and g becomes much smaller than 1 under incident conditions. Then, the
incident detection criterion is a small value of the summation of estimated v,~ (i.e., a) and

a large value of the temporal ratio of input and output in travel time period (i.e., fl).

Precisely, an incident is determined if an observed a , and ,6 point is in the critical region

17

in two dimensional space. This two dimensional region is determined empirically
(Sakasita et al. 1975).

Based on past observations of traffic variables (i.e., volume and occupancy), a forecast
can be made and compared with the next observation of the traffic variable. The
difference of the forecast and the observation signals a possible change in the traffic
stream behavior. The Box-Jenkins approach uses this idea to construct a prediction model
for traffic variable (Ahmed and Cook 1979; Ahmed and Cook 1980; Ahmed and Cook
1982; Ahmed 1983). There is an assumption in this theory. A non-seasonable time series
of freeway traffic observation (e.g., occupancy) taken at equally spaced time intervals is
assumed to be stationary or can be reducible to a stationary form. Then, the dynamics of
this observations (e.g., occupancy) is represented in the general class of linear models in
the form of autoregressive integrated moving-average models of order p, d, q (i.e.,
ARIMA(p,d,q)). The ARIMA models are obtained by a three-stage iterative procedure.
They are preliminary identification, estimation, and diagnostic check stages. In
preliminary identification stage, the values of p, d, and q are determined by inspecting the
autocorrelations and partial autocorrelations of the series or its differences, or both, and
by comparing them with those of some basic stochastic processes. In the estimation stage,
the autoregressive and moving average parameters are estimated by using non-linear least
squares techniques after the values of p, d, and q have been obtained. Finally, in the
diagnostic check stage, the goodness of the model fit is checked. If the model is
satisfactory, then the resulting residuals should be uncorrelated random deviations.

Once an ARIMA model is obtained, a short-term forecast can be made by an

Operational expression for updating the forecasts of the model that is expressed in terms

18

of forecast errors. An incident is detected if the observed value of the control variable lies
outside the confidence interval (e. g., 95 percent) for the corresponding point forecast.

In the Filtering Technique-Based Algorithm, a short-term averaging (low-pass filter) is
employed to reduce the adverse effects of short-term traffic fluctuations and high-
magnitude impulses in volume/occupancy measurement from a loop detector (Chassiakos
1992; Chassiakos and Stephanedes 1993; Stephanedes, and Chassiakos 1993a,b;
Stephanedes, Chassiakos, and Michalopoulos 1992). The performance is based on the
spatial occupancy difference between adjacent stations. Three smoothing techniques are
considered in the detection algorithm: the Moving Average (linear transformation), the
Median (non-linear transformation), and the Exponential Smoothing. The Moving
Average technique is first formulated and built into the algorithm as the spatial difference
of 30 second-occupancy outputs at time 1. Given the hypothesis of an incident occurrence
at time I, this algorithm uses 3-minute averages of the spatial occupancy difference after
time t between adjacent stations to provide an indication of the traffic condition (traffic
congestion). Then, the temporal change in the 5-minute average of the spatial occupancy
difference past time t seeks the cause of the traffic congestion (incident or bottleneck).
Therefore, the significant difference of 3-minute average of the spatial occupancy
difference after time t and 5-minute average of the spatial occupancy difference past time
t indicates a temporal change to distinguish incident related congestion from recurrent
congestion. The transferability across stations and the time of day is achieved by the
normalization obtained by the maximum of upstream and downstream station occupancy

averaged over the most recent 5-minute period before the incident.

19

The Predictor Algorithm utilizes well known Kalman Filtering to detect anomalies
which may arise from a loop detector system malfunction or traffic incidents. To reflect
freeway traffic dynamics, the transition matrix of state equations was developed. Also the
number of standard deviations over time from the prediction is used to determine
threshold values to be compared with traffic flow observations. An incident is alarmed if
the difference of the predicted and the observed parameters is significantly different from
the number of standard deviations (Dailey 1993).

This method suggests very good potential for general application to various traffic
conditions. Since the model is adaptive, it reflects freeway traffic flow over time. Also,
the dynamic thresholds is another good positive feature.

These models contributed to the state-of the art in incident detection on freeways.

Some of features of these algorithms are worth of comment.

SND Model: This algorithm takes advantage of the rate of change of the control traffic
variable instead of an absolute threshold value (Dudek, Messer, and Nuckles 1974). Also
this method is very simple and efficient. However, assigning equal weights to the
previous n sampling period for the mean in SND model will contaminate the most recent
information by obscuring unusual outlying observation (Montgomery, Johnson, and
Gradiner 1990). The determination of time span for the average value is also important

because the sensitivity of the SND model is dependent on that span.

Bayesian Approach Algorithm: The advantage of this algorithm comes from the

simplicity based on Bayesian theory. In addition, since an operator is given the likelihood

20

of an incident, he or she is not in the decision of the limited possibilities (incident or no-
incident). However, this algorithm is heavily dependent on historical incident and
incident-free data to obtain the control variable’s frequency distribution. It is not always
easy to obtain such a large data set to calibrate the probability of an incident. These data
sets are also heavily dependent on a particular location and time for the threshold.
Because of this attribute, this model is not operational in different location and time
condition with the once-determined threshold. Also the higher probability of correct
detection of an incident requires longer mean time to detect. For example, it took 1
minute for P(incident)=0.00305 and 4 minutes for P(incident)=0.81662 (Levin and

Krause 1978; Levin and Krause 1979; Levin and Krause 1979a, d).

Exponential Smoothing Algorithm: This exponential smoothing method is the procedure
that adjusts the smoothed statistic by an amount that is proportional to the most recent
forecast error. This reflects the most recent data sets unlike the simple moving average.
Theoretically, the exponential smoothing application needs to specify a value for the
smoothing constant. A widely used technique is to conduct trials of various values on a
set of historical data. Therefore, the smoothing constant for a particular data set could not
be directly used for other data. Large data are also needed to establish the standard
deviation of the tracking signal to determine an incident. The pre-defined threshold value
that is based on a particular data sets results in wild fluctuation in the number of false

alarms from day to day and from station to station (Masters, Lam, and Wong 1991).

21

Dynamic Model Algorithm: This algorithm includes a noise term in the linear model to
account for lane changes and speed changes between adjacent detectors. The upstream
and downstream information (occupancy and speed) is correlated, but the determination

of an incident is basically static. The terms a0 , and ,60 are obtained empirically and take

on different values depending on the type of incident and prevailing conditions.

Therefore, the more flexible the algorithm, the more historical incident data required.

Box-Jenkins T echnique-Based Algorithm: This algorithm is based on the assumption
that the ARIMA model fits freeway traffic flow well. Theoretically this time series
analysis requires a large historical data set to develop an acceptable ARIMA model
(Mongomery, Johnson, and Gardiner 1990). In the ARIMA(O,1,3) model obtained from
freeway data, the values of the moving average parameters vary from one detector station
to another and over time due to variations in surrounding conditions (Ahmed and Cook
1982). These parameters are not changed automatically and requires manual examinations
(Mongomery, Johnson, and Gardiner 1990). This fact will result in less flexibility in

operation for different conditions.

Filtering Technique Algorithm: The idea of smoothing short-term traffic
inhomogeneities is beneficial to get rid of corrupting impulsive noises that may impair
detection performance. The consideration of the transferability across the various data
sets is good as well. In addition, adoption of functions to identify incident-caused and
recurrent traffic congestion based on the effects of an incident is a positive attribute of the

algorithm. However, basically the DELOS algorithm uses the filtered spatial occupancy

22

difference between adjacent detector stations through time, which may not be valid
method in areas where the traffic pattern is varying between adjacent detector stations as
described in the comparative algorithms. Furthermore, the structure of the algorithm
requires time delay (3 minutes) as the smoothing time, which brings slower detection
response time in freeway detection algorithms. Although the moving average is simple
and effective, it possibly obscures information contents by applying the same weights to

averaged data.

2.1.3 Dynamic Model-Based Technique Algorithm

This technique for an incident detection is based on the use of a macroscopic dynamic
model describing the evolution of spatial-average traffic variables (velocity, flows, and
density) over sections of the freeway (Willsky 1976; Willsky et al. 1980). The authors
used the dynamic traffic flow model suggested by Payne for traffic flow and developed
incident detection systems using two different hypothesis testing techniques, the Multiple
Model (MM) and the Generalized Likelihood Ratio (GLR) algorithms. The dynamic
freeway traffic flow is based on the basic aspects of both the fluid flow and car-following
models of traffic dynamics that yields a spatial discretized set of coupled highly non-
linear equations. Specifically three types of events for each link of the freeway were
modeled. They are a capacity-reducing incident, a pulse of traffic that is lasting for a
specified duration, and sensor failure. Based on the model, incident detection algorithms,
MM (Multiple Model) and GLR (Generalized Likelihood Ratio) techniques, were

developed.

23

That the inter-relation between principal traffic is directly modeled by the algorithms
and the parameters are changed over time is a good aspect of the algorithm (Willsky,
1976). This algorithm, however, in among the more complex of the methods, and it has
large data requirements, in terms of types of variables (density, space mean speed) and
short time-space discretization (Stephanedes, Chassiakos, and Michalopoulos 1992). In
addition, as the authors of the algorithms indicated, the use of the nominal-linearized
Kalman filter is suspicious for a dynamic system. The literature indicated that such a use
naturally introduces bias into the model, and thus limits the performance of the MM

method and the GLR method (Willsky 1976; Willsky et al. 1980).

2.1.4 High Occupancy (HIOCC) Algorithm

The purpose of this algorithm is to detect an incident by looking for the presence of
stationary or slow moving vehicles over a detector. This can be achieved by examining
the one second instantaneous occupancy values (the occupancy outputs are scanned at 0.1
second intervals for the presence of absence of vehicles) for all detectors and looking for
a short sequence (2 or 3 second long) of 100 percent occupancy. Once this occurs, an
incident alarm is given. To continue the alarm during an incident, allowance is made for
the situation where vehicles may be stopped clear of the detector for a certain time period
(i.e., 8 seconds) (Collins 1977; Collins 1983).

The HIOCC algorithm is beneficial in that it does not require the complex calibration
because the incident was detected by the occurrence of 100 percent values of
instantaneous occupancy. It uses l-second consecutive occupancy data (scanned in 0.1

second interval) to detect stationary or slowing moving vehicles. This method is simple to

24

execute and uses the direct observations (occupancy) from the detectors. However, such
data (1 sec occupancy data) is not always available all the time with existing surveillance
system (Stephanedes, Chassiakos, and Michalopoulos 1992). Especially in urban
signalized arterial streets, one second data is meaningless since vehicles would stop over

detectors for at least the red phase duration of the signal at intersections.

2.1.5 McMaster Algorithm (Catastrophe Theory Algorithm)

This algorithm is based on the previous works on the freeway flow-density
relationships (Hall 1987; Hall, Allen, and Gunter 1986; Hall, Shi, and Atala 1993;
Persaud and Hall 1989). The application of catastrophe theory (especially the cusp
catastrophe) on the relationships between speed, flow and occupancy is introduced from
these works. The fact that the sudden drop in speed is obvious in traffic flow when
congestion occurs is the basis for catastrophe theory. This idea led to the McMaster
incident detection algorithm. A basic version of this algorithm uses the flow-occupancy
data that is separated into four traffic states (the uncongested state and several types of
congested states). The cause of congestion or a drop in speed makes the data change from
the uncongested state to one of the congested states. If either or both of these continued
for P consecutive periods, congestion is declared. In addition, the modified McMaster
algorithm provides the capability to discern if the congestion is incident-related or
recurrent-related (Gall, and Hall 1989; Persaud, Hall, and Hall 1990).

In this algorithm, the major advantages are the simplicity of design and a potential for
improved detection performance (Stephanedes, Chassiakos, and Michalopoulos 1992).

However, the use of particular data in defining the boundaries of congestion free and

25

congestion (incident-caused and recurrent-caused) have limitations in use for particular

traffic and surrounding conditions.

2.1.6 Advanced Technology-Based Algorithms

The neural network approach has been studied in combination with other incident
algorithms since 1991 (Abdulhai and Ritchie 1995; Chang 1992; Chang and Huamg
1993; Cheu, Ritche, and Recker 1991; Ritchie and Chen 1993). The basic concept of the
neural network is to simulate the knowledge reasoning of the human brain, organizing
massive amounts of information, and presenting this information in various forms.
Information transmission is achieved by cells called neurons through nerve connection.
Connection adjustment results in so called intelligent behavior. In freeway incident
detection aspects, neural networks can be trained by the data to recognize volume patterns
on freeways and to detect incidents. After being trained, the neural networks can be used
to assist controllers in freeway management systems.

The basic idea of fuzzy logic is that it provides formal reasoning schemes that are
approximate rather than exact as opposed to crisp logic. In the incident detection aspect,
the fuzzy logic approach allows imprecision in data received. Perfect information is very
expensive to obtain and not always available. In addition, the data needed for the
detection process is usually accompanied with noise, resulting in false alarm problems.
The fuzzy approach deals with imperfect data to get approximate results (Chang and
Huamg 1993; Tarko, Tsai, and Rouphail 1993).

There are some features that should be noted for these methods. The problem in

providing all the data for the neural network is that information should include all

26

operational characteristics such as learning data and possible solutions. If the level of
information is insufficient, the training results may not cover all cases (Chang and Huamg
1993). A particular data set for a self-learning mechanism to recognize traffic pattern uses
only one data set. For different traffic or geometric conditions, additional data must be
input to train the system. Although the PNN was recognized by the authors for
transferability and flexibility for different traffic conditions, the longer time in recall
without a parallel hardware implementation and the reliance on Euclidean distance
between the patterns, which is inaccurate in cases of correlated input variables and input
variables with very different variances were suggested as disadvantages (Abdulhai and
Ritchie 1995).

Although the Fuzzy approach takes advantage of interpreting uncertain data to
approximate results, it has computation problems. Since the fuzzy approach needs to
compute for all the nodes to draw conclusions, more computations and comparisons are

needed, which would result in detection time delay (Chang and Huamg 1993).

2.1.7 Video Image Processing Technique-Base Algorithms

A computer image processing technique, the IMPACTS, was also developed for
automatic incident detection (Hoose 1994; Kelly 1991). The strategy is to mimic the way
in which a human observer might describe the pattern of traffic when viewing a CCTV
monitor. A state of “cells” are used to identify a scene. A cell represents an area of the
roadway within a video image. They are arranged along traffic lanes and are sized so as to
convert an area of road approximately 1 lane wide by approximately 1 car length long.

Three states are defined for a cell: None, Moving, and Stop. These cells are then grouped

27

along each traffic lane to form “objects”. These objects correspond to groups of adjacent
cells that have a common “state”. Sampling time interval and image processing
operations are conducted in 5-second cycles. An incident “alarm level” is generated when
two or more stop cells are detected within three cycles and that the alarm level would be
turned off when the above condition is not detected.

Another video detector system, AUTOSCOPE was also used in incident detection on
freeways. However, this system uses raw data (e.g., speed or occupancy) from
AUTOSCOPE detectors instead of using image processing methods. Two incident
detection algorithms were developed. First, Speed Profile Incident Evaluation System
(SPIES) employs two speed-traps placed in each lane of traffic. Then, volume-smoothed
basis at roughly 15 second intervals and also 15-minute time intervals (for historical data)
are generated. Incidents are then detected by comparing the difference from the volume
smoothed speeds of adjacent detectors and historical data with an incident alarm
threshold. The second method is AUTOSCOPE Incident Detection Algorithm (AIDA),
combining the strength of McMaster and SPIES, but taking advantage of temporal
variations of the traffic parameters in addition to spatial ones. Rapid breakdowns in traffic
(e.g., sharp speed-drops or occupancy increases) is compared with speed thresholds for
determining the congestion level (Michalopoulos and Jacobson 1992; Michalopoulos et
al. 1993).

Fast incident detection is a favorable aspect of the technique. However,
misclassification of stationary (a stop cell when there is no stationary traffic in the field of
view), camera position problems and fixed threshold problems due to lighting conditions

must be overcome (Kelly 1991). Traffic detection itself is also a problem for video

28

detection. Leading headlight reflections at night on wet pavement and strong shadows
were found to cause a false detection of vehicles, which eventually results in false alarm

in incident detection (Michalopoulos and Jacobson 1992; Michalopoulos et al. 1993).

2.1.8 VRC Sensor-Based Incident Detection Algorithms

The VRC (Vehicle-to-Road Communication) or the AVI (Automatic Vehicle
Identification) consists of a transponder or “tag” on the vehicle and a reader along or over
the road and the communications link between the two. VRC readers can obtain at least
the individual vehicle identification number from each transponder-equipped vehicle that
passes. With this information, the system can obtain the car’s travel time, lane specific
and station-specific headways, the volume of tagged vehicles on a section, their lane
switching, and etc. Incident detection is implemented using this information. Three
algorithms were suggested using the data from new traffic sensor VRC (Parkany, and
Bernstein 1993; Parkany, and Bernstein 1995). They are the Headway Algorithm, the
Lane Switches Algorithm, and the Lane-Monitoring Algorithm.

In the Headway Algorithm, both temporal and spatial comparisons of travel times and
headways are used. A significant temporal difference in travel time from a pre-determined
threshold brings up a temporal difference of headways. A signal difference of headways
from a pre-determined thresholds then brings up spatial difference of headways. An
incident is alarmed if all three conditions were satisfied.

The Lane Switches Algorithm counts the number of vehicles that have switched
between readers. If the percentage of vehicles that have switched lanes exceeds a certain

threshold, an incident alarm is declared.

29

In the Lane-Monitoring Algorithm, the lane average number of vehicles for two time
intervals in lanes are obtained. If this average value of one of lanes is less than the pre-
determined low threshold, then another lane average is compared with pre-determined
high threshold. If any of the other volumes exceed the high threshold, then an incident is
declared.

Even though, a new technology is used, these three algorithms are classical methods as
they approach the problem similarly to the comparative algorithms. They are simple, easy
to understand, and show potential for a freeway incident detection. However, even though
the data sets that can be obtained from this type of sensors are valuable, they are not
available on all freeways. Since this system depends on vehicles which are equipped with
the “tag”, if there are not many equipped vehicles, then, there is less chance of success. In
addition, the threshold values should be more dynamic than static for the algorithms to be

applied under other conditions.

2.2 Existing Urban Arterial Street Incident Detection Algorithms

Recently, the issue of automatic incident detection in signalized arterial streets has
attracted traffic engineers. They include new approaches such as neural network and
computer image processing as well as conventional ways that have been studied on

freeways.

2.2.1 Automatic Incident Detection (AID) within Urban Traffic Control System
As a pilot study, Bell and Thancanamootoo (1988) developed an automatic incident

detection (AID) within an urban traffic control system. Basic exponential smoothing was

30

used for updating a traffic control variable, cyclic occupancy, aggregated in a signal cycle
time interval. Occupancy variance was also exponentially smoothed and used to
determine an incident if the newly obtained cyclic occupancy was out-of-bounds of a
certain confidence limits. If the upper bound is infringed, an incident is suspected, and the
lower bound is infringed at the next downstream detector, the incident is confirmed. The
smoothing factors for exponential smoothing cyclic occupancy were 0.8 and 0.2. Also 0.9
and 0.1 were used for exponential smoothing occupancy variance parameters. The
MONICA (Monitoring Incidents and Congestion Automatically) is the extended version
of this work by Bretherton and Bowen (1991). While Bell and Thancanamootoo used
information from only two links, which resulted in unacceptable false alarms, MONICA
(Monitoring Incidents and Congestion Automatically) used surrounding links as well.
This algorithm uses the simple exponential smoothing method. The use of signal cycle
length as data aggregation is also realistic for signal controlled streets. However, the
determination of parameters for the exponential smoothing functions varies site by site.
Also, the incident confirmation method by checking adjacent detector stations turned out
to be impractical because some incidents were not even detected at adjacent detector

stations.

2.2.2 A Dynamic Real-Time Incident Detection System for Urban Arterial System

A dynamic real-time incident detection system for urban arterials were proposed
recently (Chen, and Chang 1993; Chen 1994). This system consists of three major parts:
1) a dynamic traffic flow prediction model, 2) an incident identification, and 3) an

incident monitoring process. Associated with the incident detection scheme, the

31

occurrence of an incident is alarmed by the discrepancy between the projected and
detected traffic conditions. Time evolving traffic conditions are modeled with three traffic
variables (flow, occupancy, and speed). This model considers the time-spatial dynamics
of traffic information from upstream and downstream detectors. Exogenous effects in
urban arterial streets (lane-changes, spilled-back queues, buses, parking vehicles, illegal
pedestrians, green time of traffic signal, and uncontrolled access in the middle of a road
sections) are involved in the time-space models. The Kalman Filtering technique was
used to update model parameters at each time-step in real-time based to execute its self-
learning mechanism.

By using Kalman filtering technique, the strength of this algorithm appears in dynamic
parameter changes at each time step. The inclusion of various exogenous traffic flow
disturbance factors between upstream and downstream detector stations is also another
good consideration if the prediction model uses adjacent detector information. However,
the use of two detector stations within a link is not realistic. In most streets there is only
one detector station placed usually at the downstream end of a link. If the above dynamic
prediction model is applied to a one detector station street, more complicated traffic
interaction models within an intersection should be included. Also the method of data
aggregating time interval is not general if the network has varying cycle time and link
length. For example, a link that is 2000 feet long would need data aggregating time
interval much greater than 30 seconds (probably 80 seconds if travel time is assumed to
increase linearly worth the length of a link). Different data aggregating time intervals are
required for each different-length link, resulting in loss of generality over the network and

longer detection time if the time interval is long.

32

2.2.3 Neural Networks for Real-Time Data Fusion Method

This method uses information that is provided and preprocessed from three data
sources (i.e., fixed detectors, probe vehicles, and anecdotal sources) to determine an
incident after integrated in a data fusion process (Ivan et al. 1993; Ivan et al. 1995;
Bhandari et al. 1995; Ivan 1996). Fixed detector algorithm preprocesses occupancy and
volume data averaged over a fixed time interval (e.g., 5 or 7 minutes). Probe vehicles
report link travel times by radio, and anecdotal information provides particular events
affecting traffic flow. Fixed detector algorithm compares current and historical volume
and occupancy data from fixed detectors at the end of every period. A discriminant score
is then computed which determines incident state in the proximity of each detector. The
probe vehicle algorithm uses current and historical (non-incident) probe travel time data.
Average travel times are computed by aggregating individual probe reports at the end of
each period in the first stage. Travel time ratio and speed ratio are computed using the
observed travel time on the links and the corresponding historical travel times stored in a
database. Incident determination is done by computing the discriminant score. The
anecdotal information algorithm utilizes a qualitative description of incidents reported by
field observers, both trained and untrained, to detect incidents in real time. These separate
algorithm outputs are then combined for a neural network.

As discussed in the freeway section, neural networks require enough information
which should include all operational characteristics and possible solutions. Practically it
is hard to obtain sufficient valued data sets. Bad data sets such as malfunctioning of fixed

detectors also may mislead the model. Especially in arterial streets applications, the data

33

aggregation time interval should be reconsidered. Aggregation time intervals of 7 minutes
(Ivan et al. 1995) may be good to reduce fluctuations of traffic volume but long to detect

an incident.

2.2.4 Video Image Processing Technique-Based Algorithm

Hoose, Vicencio, and Zhang (1992), and Martinez and et al. (1944) suggested video
image processing as an incident detection system for urban roads. This method is similar
to that used in freeway incident section. Video images of traffic scenes obtained from a
video surveillance camera is used to determine an incident. The information by analyzing
video images is a periodic description of the pattern of the traffic along the length of road
and an intersection under surveillance. Information extraction is made from both urban
links and intersections. Information for urban links is obtained in a basis of the traffic
lane. Each lane has data set of spatial occupancy, number of objects (groups of adjacent
cells that have a common state), and speed and errors in objects. Intersection information
provides spatial occupancy, motion occupancy, static occupancy, number of groups
(moving or stationary), and associated information in groups (motion and errors.) An
incident is detected based on the data presented to it by the computer vision process,
combined with external data (e. g., data from existing loop detector). The threshold values
to be compared with the time length for vehicles to stop at stop bar for a red phase of a
signal was also suggested. Sellam, and Boulmakoul (1994) also presented an automatic
incident detection methodology at intersections. The method is based on several
technologies such as object-oriented design, real-time knowledge-based systems and a

video image processing. To differentiate incident-related stops of vehicles from non-

34

incident-related stOps due to traffic red-light or parking, they characterized pre-defined
(associated with area classes) punctual incidents and incidents of a functional type
(shortcoming movements, overload of a phase or cycle etc.).

The system is useful for fast detection of incidents (although the qualitative value was
not shown) (Martinez et al. 1994). However, high false alarms (24.23 percent on two field
tests) is a problem for practical use over a larger network. Further research problems
relate to variations in camera position and stability or extremes of weather and lighting

condition (Hoose, Vicencio, and Zhang 1992).

2.2.5 Intersection Accident Detection Algorithm

An accident detection method in an intersection is based on observing an intersection
under accident and accident-free conditions (Stephanedes, and Vasilakis 1995). It is
assumed, under accident-free condition, that traffic volume is evenly distributed across
all main moving lanes and the probability that a driver chooses one of N lanes is normally
distributed. Under accident conditions, severe lane change by drivers occurs and shows
significant differences in comparison to normal traffic parameters (traffic volume), prior
to and after the accident. A lane is defined as blocked if the detector transmits a 100
percent occupancy value at that location. These differences are compared with pre-
determined threshold values. Significant deviation from the historical average indicates
the occurrence of an accident.

This method uses abrupt change of traffic parameters such as volume and occupancy
to detect an accident in an intersection. Basically this is true to some extent for accident

cases. However, incidents at mid-block are not clearly supported by this method because

35

an incident at mid-block (if the link is long) would not affect traffic inter-actions easily at
an intersection. Drivers who notice a stalled car in one of lanes would simply change lane
for the block and switch again to then original lane, especially if they are turning at the
intersection. This method also should consider data sampling collection time-interval and

detector locations.

2.2.6 Smart Corridor Arterial Incident Detection System

As a part of the Santa Monica Freeway Smart Corridor Demonstration Project, an
automated incident surveillance system was developed (Roseman, and Skehan 1995). The
methodology is based on a method of time weighted data analysis. Volume and
occupancy data from each system detector are gathered from the real-time traffic control
system on a minute-by-minute basis. They are then transmitted to the Smart Corridor
system. The detector data is then combined into link data which is smoothed with data
from the previous four minutes to obtain a sliding five minute average link value. That is
then compared to a matrix which represents the normal, congested and maximum traffic
values expected on this particular link. From the comparison, data values within range are
given a persistence value. If traffic conditions warrants over time, the persistence value
will increase, and ultimately an incident will be identified. Links in the system are
assigned to a particular group based on their roadway conditions, which may change by
time of day. Within each group, there is a table of threshold values which the system uses
to compare current data to determine a persistence value. These threshold tables are

selected by time of day, and day of week based on historic traffic flow conditions. The

36

Smart Corridor system uses a total of 144 persistence matrices (16 groups, each of which
contains nine time-of-day threshold tables).

This methodology is simple and flexible to allow non-programmers to change the
database for performance adjustments. But this simplicity comes at the expense of
manually creating, fine—tuning, and maintaining the large link specific database. A
significant effort is required to develop database for a particular location such as
assigning detector and link parameters, assigning link group numbers, and developing the
threshold values for the persistence matrices (Roseman, and Skehan 1995). The pre-

determined threshold values are still static, which does not reflect dynamic traffic flows.

2.3 Summary of Existing Freeway and Urban Arterial Street Incident Detection
Algorithms

Existing incident detection algorithms for freeways and signalized arterial streets were
reviewed. For freeway incident detection, many studies have been conducted over 20
years. From the conventional California Algorithms to emerging new technology-based
algorithms such as neural network and video image processing methods, they have given
valuable contributions. It is hard to say which model or algorithm is better than others
since their direct comparisons were unavailable for all the different conditions. Each of
these algorithms or models has shown superior results for a certain condition. However,
by reviewing previous freeway studies, some commonly found issues should be
recognized. Chen (1994) also presented these issues in his work. Some of these issues are

inter-related.

37

Data Dependency: Usually models or algorithms that use fixed thresholds require
comprehensive historical data for a particular condition. The comparative algorithms,
some of the statistical approaches such as the Bayesian model, the Exponential
Smoothing Algorithms, the Dynamic Model Algorithm, and the Box-Jenkins Technique-
Based Algorithms need heavy historical data to obtain fixed thresholds or model
constructions. This is also true for the McMaster Algorithm, the Neural Network-Based

Algorithms.

Transferability: Transferability is inter-related with the data dependence problem.
Although many studies showed good performance on a particular condition, they lack
general application to other conditions. For example, the California Algorithm Family
uses pre-determined thresholds that are obtained by trial and error from a particular traffic
condition and geometry. These thresholds are not directly applicable to other conditions.
Some of the statistical approaches such as the Bayesian-Based Algorithm, the Dynamic-
Model Algorithm (Sakasita et al. 1975), and the Box-Jenkins Technique-Based Algorithm
are also limited by heavy data dependence, preventing from being easily transferable to
other conditions. This is true for the McMaster Algorithm since it uses the boundaries of
flow-occupancy data of a particular geometric and traffic condition to determine an
incident. The Neural Network-Based Algorithm needs a comprehensive descent data for
training itself. This algorithm is not applicable for other conditions, unless corresponding
data was provided. The VRC Sensor-Based algorithm still has this limitation since the

incident approach is dependent on pre-determined thresholds.

 

38

Adaptability (Real-Time Operation): This issue is also inter-related with the
transferability problem. Since many algorithms use pre-determined thresholds or models
with fixed parameters, they are also limited in transferring to other conditions. Some of
models such as the SND model, the Filtering models, and the Dynamic Model-Based
Algorithm (Willsky et al. 1980) are prominent in this context. They take advantage of
flexible thresholds and parameters to system changes. This adaptability is also an issue
with the computing efficiency for the real-time operation of models. For the model to be
implemented in real-time, these two factors should be satisfactory. The Kalman Predictor
Algorithm and the Dynamic Model-Based Algorithm is superior in adaptability due to the
Kalman filtering method. However, the complexity and long computing time of the
Dynamic Model-Based Algorithm would not satisfy the real-time operation requirements

(Chen 1994).

Availability: Newly emerging techniques such as video image processing and VRC
sensor-based methods demonstrated the potential in incident detection. However, not all
the freeways, streets, and cars are equipped with these systems. Penetration rate of these

equipment, especially VRC sensor and “tags”, is an important issue to be studied.

The above issues are also applicable to the existing arterial street incident detection
methodologies. Although newly developed methods show potential, these algorithms do
not meet all the requirements. The AID study is limited with fixed smoothing parameters

in transferability and adaptability. The Dynamic Real-Time Incident Detection System

39

meets above requirements, but failed in the availability due to existing detector
configurations in arterial streets. The neural network for data fusion study also is
restricted in the transferability and the adaptability due to heavy data dependence from
various sources. The video image processing method is still in the study stage and not
available on all the streets. The incident detection in intersections method should be
generalized to cover the mid-block incidents. The Smart Corridor study also does not
meet the adaptability and the transferability criteria due to heavy geometric and traffic

flow data dependence.

Chapter 3

INCIDENT DETECTION ALGORITHM DEVELOPMENT

3.1 Kalman Filtering Theoretical Background

The Kalman filter was derived as the solution to the Wiener problems (i.e., the
prediction of random signal, the separation of random signals from random noise, and the
detection of signals of known form in the presence of random noise). In Kalman filtering,
Wiener’s problem is approached from the point of view of conditional distribution and
expectations. It also uses the state-space model for the linear dynamic and random
process to control system. It is not always possible or desirable to measure every variable
to be controlled, and the Kalman filter provides a means for inferring the missing
information from noisy measurements. A brief theoretical background is contained in the

follwing sections (Kalman 1960; Grewal and Andrews 1993).

3.1.1 Optimal Estimate and Orthogonal Projection

In Kalman filtering, as in Wiener’s problem, the estimation (predicted, filtered, and
smoothed estimates) problem belongs to the realm of probability theory and statistics.
The conditional probability distribution function of random variables, signal, noise and

their linear function (random process), represents the basic idea of the estimation. For

40

41

example, any statistical estimate, X,(t,|t), X,(t,), or simply X, , of the random variable
x, (t) is some function of the conditional distribution,

Prrxxt.) s §1|y(to) = n<t.).---,y<t> = 77(1)] = at.) (3.1)
and therefore a (non-random) function of the random variables y(tO ),---,y(t). This

conditional distribution represents the probability of the occurrence of variables values

tf,(t) of the random variable x,(t,) given the set of measured values 77(t0 ),-~,77(t) of
the random variable y(t).

Since there is, in general, a difference between the estimate X, (t,|t) and (unknown)
true value x, (t, ) , a penalty or loss for an incorrect estimate is assigned. The loss should
be a positive and nondecreasing function of the estimation error 5 = x, (t, ) — X, (t, ). The

loss function is, therefore, defined by

L(0) = 0
L(6‘2)2 L(e,)20when 62 28, 20 (3.2)
L(£) = L(—£)

 

Some common loss functions are L(e) = aez, ae4, aej, a[1 —exp(—.€2 )1, etc., where a

is a positive constant.

A way to choose X, (t, ) may be derived, then, by minimizing risk (average loss)
E{L[x.(t) - X1001} = EIElLIX(n ) - X1 (tl)lly(to),-",y(t)}l (3-3)
The theorem (Kalman 1960: 37) shows that assuming that L satisfies (3.2) loss function
type and the conditional distribution function F (5) defined in equation (3.1) is
symmetric about the mean E and convex for 4‘ S E, the optimal estimate of x, (t,) which

minimizes the risk is the conditional expectation

42
xi<nlt> = Erx.<n>ly<to>,~-.y(r>r (3.4)

Especially, if L(e) = .92, the assumption that F (5) is symmetry is unnecessary for the

2 . .
, this theorem rs

 

 

optimal estimate. For vector valued random variables, when L(e) = ”a

still true.
This optimal estimate can be obtained from the orthogonality theory under a certain

condition. If the optimal estimate is restricted to be a linear firnction of the observed

random variables and L(5) = .92 , then

optimal estimate of x(t,) given y(tO ), ..., y(t)

XIUIII) (3.5)

orthogonal projection 5?(t,|t) of x(t,) on Y (t)

where f denotes the orthogonal projection of random variable x, and Y (t) a vector space

(linear manifold) that is formed by y(tO ), ..., y(t).

3.1.2 Linear Optimal Filters
A stochastic process (or random process) is the evolution of the state of the system if
the possible states of a non-deterministic system at any time can be represented by a
random variable. It is supposed that stochastic linear systems associated with the Kalman
filter can be represented and estimated by the types of plant (or state) and measurement
models. These models can be described as
System Dynamic Model (State Space Equation)
x, =<I>k_,xk_,+wk_,,wk_, EN(0,Qk_1) (3.6)
Measurement Model (Observation Equation)

2,, :Hkxk +vk, vk E N(0,Rk) (3.7)

43
where
x k is a n x 1 state vector
2 k = l x 1 measurement vector
(1) k = n x n time-varying dynamic coefficient matrix
H k =1 x n time-varying measurement sensitivity matrix
w k = n x 1 zero-mean uncorrelated Gaussian “state noise” process
v k =1 x 1 zero-mean uncorrelated Gaussian “measurement noise” process
Q k = n x n time varying matrix
R k = l x l time-varying matrix

These models satisfy the following conditions.

EON) = Eka) = 0

E(w,,wjf) = {32,10 ::j

EMVI) = {3" :2. (3.8)
E(wkaT.) = 0

and

E(ka,:) = 0

The state variable represents the variable to be controlled. They can not generally be
measured directly, but are inferred from what can be measured. The future state of a
dynamic system is uniquely determined by its current state and its future inputs. The
dynamic behavior of each state variable of the system must be a known function of the
instantaneous values of other state variables and the system inputs for the future state of a
system to be determined. The state-space model represents these functional dependencies
in terms of first-order differential equations (in continuous time) or difference equations
(in discrete time).

In a linear dynamic system we can estimate the state of a linear stochastic system in

three ways.

44

Predictors use observations strictly prior to the time that the state of the dynamic system
is to be estimated (10b5, < test)

Filters use observations up to and including the time that the state of the dynamic system
is to be estimated (tobs. 5 test)

Smoothers use observations beyond the time that the state of the dynamic system is to be
estimated (tobs, > test)

Predicted estimates will be used to predict traffic variables. The predicted estimates are

the dynamic changes of filtered estimates.

3.1.3 Kalman Filtering Algorithm

Kalman filtering is an algorithm used to estimate the state of a linear or non-linear
dynamic system by assuming the state-space model and the measurement model. The
typical procedure to estimate the state of a linear system is as follows.

1. Initialize the state and its error covariance, that is

~ ... (3.9)
E("0"0 ) = P0
where it”, = i, —x,,.
2. Compute Pk (—) using Pk_,(+),CI>k_,, and OH. That is,
Pk (—) = ¢k—lPk—l(+)q)Z—l + Qk—l (3-10)

where P, (—) is the a priori covariance (the error covariance matrix before the

measurement), which is

45

P. (—) = no. - i.<—»(x. - i.<—))T1

r (3.11)
=E(’i."i.‘ )
where ”i,- = x, —x,.(—).
3. Compute K, using P, (—) (computed in step 2), H k and R k, that is,
K, = P,(-)H{[H,Pk(-)H[ + 12,1" (3.12)

where K k is the Kalman gain matrix.
4. Compute P, (+) using K, (computed in step 3) and P, (—) (from step 2). That is,

P. (+) = [I — K.H.]P. (—) (3.13)
where Pk(+) is the a posteriori covariance matrix (the error covariance matrix after
update), which is

P. (+) -- E[(x. —i.(+))(x. — i. (+))Tl
~+~+r (3.14)

= Eka xk )
5. Compute the value of i,(+) , a filtered estimate, using the computed values of K,

(from step 4), and measurement 2,. That is,

ik(+) = ik(") + EkIZk ‘ Hkik(-)]

. _ . (3.15)
= (bk—1M4“) + Kklzk " Hk¢k—lxk—l(+)]
6. Compute x, (—) , predicted estimate, using it, (+) . That is,

7. Go to step 2 for recursive filtered estimate and predicted estimate. The parameters,

(1) ,H ,Q ,and R, can be updated if they are time-varying.

46

These equations are called Kalman filter recursive equations. Steps 3, 4, and 5 describe
the observational update (i.e., filtered estimate) after obtaining the corresponding
measurements. Steps 6, 7, and 2 describe the temporal update (i.e., predicted estimates)
before the next measurements are obtained. They can be seen in the flow chart in Figure

3.1.

 

Initial State and
E rror Covariances

 

 

 

 

 

Compute Predicted
ESIIITIallOI'l Error Covariance

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Compute Predicted Compute Kalman
Estim ate Gain Matrix
Compute Filtered Compute Filtered
ESIIITI ate Estimation Error Covariance

 

 

 

 

 

 

Figure 3.1: Recursive Kalman Filtering Steps

3.2 Incident Detection Algorithm Development using The Kalman Filtering

An incident detection algorithm can be developed based on the Kalman filtering
algorithm. Since the Kalman filter is an algoritlun, some modification of steps are made
to determine an incident. Although, Kalman filtering was a basis for one of the freeway
incident detection algorithms (Dailey 1993), there are some differences in the approach

for arterial streets.

47

3.2.1 Considerations fer Signalized Arterial Street Incident Detection
In the investigation in the literature review section, valuable lessons were learned.

Some of the freeway studies present an approach that has potential arterial street

applications. These studies were examined in terms of the historical data dependence, the

transferability, the adaptability, and the availability.

Traffic flow is more dynamic and complex on arterial streets than on freeways. This
complexity comes from various sources. Side street entrance and exits cause frequent lane
changes of main street cars, slow vehicles such as buses, parking vehicles, and even
bicycles and pedestrians are all external disturbing sources which interrupt the traffic
stream on arterial streets that are rarely seen on freeways. A traffic signal system at an
intersection is a major disturbance to main street traffic. There are, therefore, basic
parameters that to be examined to consider these factors.

0 Data Sampling Time Interval: In most freeway studies, it was possible to utilize a
short time interval such as 30 seconds for collecting data. The HIOCC Algorithm uses
0.1-second data readings to track traffic flow dynamics. This was possible and useful
owing to freeway traffic flow continuity. However, a fundamental problem exists in
urban arterial streets. A signal system at an intersection prevents utilizing such a short
sampling time interval. Since most detectors are located at intersections, the flow
dynamics would appear as step functions due to the green, yellow, and red intervals.
Theoretically, each interval could be used as a sampling time interval, but signal
phases differ intersection by intersection, and even within the same intersection over

time, thus this approach would not be generalizable. Other data aggregation time

48

intervals such as the seven minutes used in a previous study (Ivan 1995) minimizes
the effect of flow variations through the cycle. However, seven minutes in incident
detection can be too long, especially, for time series models, which usually compare
predicted values and on-coming measurements. Therefore, the best and least data
sampling time interval is the common cycle length over a controlled network since the
network is usually optimized with the cycle length. This sampling time interval
reduces data fluctuations due to different phases of a signal. Fluctuations that still
come from a stop-and-go situation can be smoothed by some method, such as the
moving average or the exponential smoothing method, etc.

- Detector-Station Type: As learned in the freeway comparative algorithms, the use of
spatial information from adjacent detector stations for determining an incident is less
effective where traffic flow is adversely affected by a geometric discontinuity between
two consecutive detector stations. In urban arterial streets where there are many
varying disturbing sources to traffic flow, point information from a single detector of
a single station approach is essential. The correlation between upstream and
downstream measurements in an arterial street may not be useful. For example, the
number of vehicles counted at an upstream detector station would not be the same or
close to the number of vehicles counted at downstream detector stations due to

turning traffic in an intersection, traffic from side streets, and on-street parking, etc.

3.2.2 Linear State-Space Model
Since Kalman filtering uses steps to estimate the predicted and filtered parameter

values, this procedure can be directly used to form an incident algorithm. Some

49

modifications for adaptive state-transition matrix ((13),) and variance matrix (Q U of the

state-space model were made to obtain varying parameters in that procedure.
The state-space equation represents the dynamic (evolution) of the state variables with

the additional noisy term. The state-transition matrix ((13),) is usually a known function

but it is necessary to determine the relationships between traffic control variables at time
k and k-I since there is no dynamic equation or relationships for the time evolution of the
traffic control variables. Dailey (1993) constructed this relationship for raw
measurements such as traffic flow, percent occupancy, and average speed. To obtain the

state transition matrix, (I) k , first multiply the state equation by x).-. on both sides and take
the expected values. Then,
_ T T —1
(bk—I — E(xkxk-1)[E(xk—lxk—1)] (3-17)
where E(wk_,x,f_,) = 0 since wk_, and x,_, are uncorrelated. That is, the state transition

matrix is represented by the multiplication of variance of x with lag l and variance of x.

Therefore, (l)k_, is the matrix that represents the relationship of data from time k-I to k.

This was approximated by Dailey (1993) as:
k=K r k=K r I
(Dr E (Zxkxk—OKZXk-ka—OI— (3-18)
k=l k=l

since denominators K and K-I for expected values are approximately the same as time

evolves. Also, to have the time-dependent noisy term, v,_, z N(0, Q k_,), the matrix Q“
was also estimated. To obtain this matrix, first move <1>k_,xk_, from the right to the left,

square both sides of the state equation and take its expected value. Then

50
Q.-. = E(x.xl>— E<x.xi..)d>i-.

(3.19)
" [E(ka[-1)¢I{—1 IT + (bk—1E(xk—lxk—1)(Dk-l

. r
srnce E(wk—lwk—l ) = Qk—l'

3.2.3 A Recursive Incident Detection Algorithm

As the regular Kalman filtering algorithm begins, a recursive process is started by
initializing state variables and their error covariances from pre-obtained data. The
temporal update to obtain predicted estimates then follows. The transition matrix (CD) is
estimated. This transition matrix estimate is used to estimate next-time predicted
estimates of control variables. The corresponding variance matrix (Q) is then obtained for
that time step. Predicted error covariance at this time step is then estimated. The Kalman
gain matrix and filtered error covariance for the oncoming new noisy measurements
follow. Then filtered estimates are obtained using the just obtained Kalman gain matrix
and new noisy measurements. This process shows the observation update after obtaining
new noisy measurements from detector stations. An incident state is indicated by
comparing the corresponding traffic control variable thresholds computed from predicted
error covariance and filtered estimates. False alarms due to inhomogeneous fluctuations
can be reduced by smoothing raw aggregated data.

In this algoritlun, the prediction model is implemented at each detector station
separately. This feature is beneficial since there is no need to consider various exogenous
factors between adjacent stations, resulting in no control variable functions associated
with these factors and adjacent detector stations. This algorithm is self-learning using data

from detectors which already reflect oncoming traffic dynamics.

51

3.2.4 Incident State Determination

Traffic control variables show different patterns when an incident occurs. These patterns
are similar to those shown in Figure 1.1 although this figure is from a freeway system.
The patterns are dependent on traffic flow, geometry, detector configuration, incident
conditions, and incident location. However, if detectors that are separately placed and
operated in each lane of a roadway provide separate information, not aggregated
information from adjacent detectors, different patterns for the same traffic control
variables can be seen. For example, a partial-blocking incident (e.g., one-lane blockage in
a multiple lane-road) that occurs immediately behind a detector station (i.e., downstream
detector station) well soon show lower occupancy, lower flow rate, and higher speed (or
normal speed) for the detector in that lane. A detector (or detectors) in adjacent lanes then
shows higher occupancy, higher flow, and lower speed (or normal speed) due to the lane-
switching behavior of vehicles. These are different patterns of the same control variables
for the same incident and the same detector location (except lane). This knowledge can be
used to develop an incident detection strategy for this type of detector configuration.
Instead of using a particular criteria-combination such as higher occupancy, lower flow,
and lower speed; or lower occupancy, lower flow, and higher speed for a particular
detector station, any of these control variables which is detected to be out-of-intervals is
an indication of an incident state. This eliminates the chance of missing incidents due to a
combined criteria of each control variable. Also, the time requested to combine these

control variables can be reduced. For example, a higher flow indication may not be

52

obtained at the same time a lower occupancy is obtained. Higher flows may occur several
time steps later.

Determining an incident is implemented by constructing the upper and the lower
intervals obtained from the predicted error covariance and comparing them with filtered
estimates of traffic control variables. These intervals are not fixed, but change over time
since updated predicted variances and estimates are made as soon as new measurements
are obtained. These intervals can be written as

Upper Limit at time k: i, (—) + no,
' 3.20
Lower Limit at time k: i, (—) — nag, ( )

where 02, is JP, (—) ,and n is the number of 032, (standard deviations).

The procedure for determining an incident can be seen in Figure 3.2.

53

 

Initialiqtion
(in ’ P0)

I

 

 

 

 

t=t+l

 

 

D,

 

Transition Matirix
and
Variance Update
( ¢,Q )

 

 

 

 

  
 

Are Filtered
Estimates Out
of Limits?

   
   

 

Filtered Estimates

Update (i(+))

 

 

 

 

Filtered Error

 

 

(P(+))

 

Figure 3.2 Incident Detection Algorithm

Yes

Tt=t+l

Incident
Notice

 

 

 

 

 

Covariance Update ‘—_.

 

Kalman Gain

>1

 

Predicted
Estimates Update

(217))

 

 

 

Predicted Error
Covariance Update

(P(- ))

 

 

 

Upper and Lower
Limits Update

(it( )in\lP1(-) ))

 

New Smoothed

 

Matrix _Update a:
( K )

 

 

Data(z)

____/// '///

 

 

 

 

 

/// ///—

using the Kalman Filtering Recursive Procedure

Chapter 4
INCIDENT DETECTION ALGORITHM TESTS

ON ARTERIAL STREET FIELD DATA

The Kalman filtering technique is not new in the transportation field. Gazis, and
Knapp (1971) applied the technique to the estimation of traffic densities at the Lincoln
Tunnel. Some freeway incident detection models adopted this model for real-time
operation.

It is possible to use Kalman filtering if the situation can be expressed as a state-space
model. Actually, Kalman filtering is beneficial because it can run variable models,
variable parameters, and variable variances simultaneously. Also, this filtering method
considers the correlation between data series, so that the predicted and filtered estimates
are more realistic. For instance, the traffic flow and speed are correlated. The data show
that the flow increases while the corresponding speed drops in peak hours and the
contrary phenomena occurs in off-peak hours. Also, the flow in one lane affects the flow
in adjacent lanes (e.g., lane changes from the congested to the non-congested). On
freeways, the upstream traffic condition affects the downstream condition, but this is not
always true on arterial streets. Kalman filtering can be used to estimate these interactive

traffic variable changes adaptively.

54

55
4.1 Application to Incident-Free Field Data

The incident-free field data were obtained from a two-lane approach along Hagadom
Road where it meets Grand River Avenue in East Lansing, Michigan in July 1995.
Detectors (HI-STAR from NU-METRICS, Model NC-90A) were placed in each lane
about 500 feet upstream from the stop bar of the intersection. The detector senses a
vehicle, the speed of the vehicle, and provides the length of that particular vehicle. Since
the detector provides microscopic information on traffic, it was necessary to aggregate the
data in a certain time interval. The best and least time interval to use as a unit time for
time series data in this case was the signal cycle length. In this test, however, the signal
system is an actuated type, therefore, it was necessary to select an appropriate, but
constant analysis period. The time interval of 3 minutes was used for the data set.

The available data is the traffic flow (counts) in a given time unit, and the time mean
speed. Data were obtained from 12:00 RM. to 8:00 PM. The data shows that in off-peak
hours (12:00 - 4:00 PM.) a relatively consistent pattern of traffic flow and speed evolves
over time in each lane. However, in peak hours (4:00 - 6:00 PM), speed drops over time
while traffic flow increases until the off-peak period (after 6:00 PM.) This phenomenon

is expected in normal days.

4.1.1 Linear State-Space Model for the Data
To implement the Kalman filtering algorithm, it is necessary to construct the state-
space model and measurement model. The state vector in the state space model is

represented as

x, = [x{, xf. x,{, x‘]T (4.1)
where superscripts f and s represent traffic flow and speed. The subscript represents the
lane numbers where detectors are placed. These are (unknown) true flow and speed

values that we are interested in estimating. Then, the linear state-space model (equation

3.7) for these measurements can be constructed by the equations 3.17 or 3.18 and 3.19.

4.1.2 Linear Measurement Model for the Data
In the stochastic process system in this study, the observation at time k is assumed to
be linearly related to the state of the system at the corresponding time as given in the

measurement equation. The measurement vector for this data set is represented
_ f s f s T
zk _ Z] 3 len'zl 9 Z] (42)

where the subscript and the superscript represent the same as in the state vector. These
measurements are noisy observations from detectors.

It is also necessary to obtain the relationship between detector observations and true
observations for the measurement equation. For this particular detector, two tests were
conducted.

To obtain the measurement sensitivity for traffic flow, detectors were placed on
Harrison Road in East Lansing, Michigan, in February 1996. Traffic flow then was video-
taped for A.M., RM. and off-peak periods. The usable data set was from 7:00 AM. on
February 6, 1996 to 8:00 PM. the next day, resulting in a total of 116 records when

aggregated in 3 minutes. This data set was then smoothed by moving average of order of

57

3 to eliminate impulsive noises. Detector measurements and the video-tape measurements
were compared.

The measurement sensitivity test for speed was also performed. Since it was difficult
to obtain true speed of random vehicles, a field driving test was conducted. The test was
conducted in a large parking lot. A detector was placed in the middle location of the
longest driving path. The driving path was long enough to increase the speed to more than
40 mph. This speed was high enough because most of arterial streets are posted around 30
to 35 mph, at most 45 mph. A total of 45 tests were conducted. Then, detector speed
measurements and recorded speedometer measurements were compared. Since there are
errors in reading the speedometer, detector measurements were assumed to be the equal
to the recorded measurements if they were in the range of recorded measurements i 3
mph.

The linear data fitting for flow and speed for this particular detector are as follows.

. 1
xf = ———yf — 0.441608
0.987767
1 (4.3)

x, = ——-y, — 2.0364170
0.907173

where

if = true traffic flow estimates
y f = observed traffic flow

is = true speed estimates

y, = observed speed

58

These are plotted in Figure 4.1. The slopes and the intercepts explain the noisy factors for

detector performance. Then, the measurement equation can be shown as follows.

 

 

 

r21/ "0.987767
2f 0

22f = 0
-23 _ - 0

4.1.3 Initialization

0
0.907173
0
0

0 ‘ 5e{ ”0.441608“

0 56,‘ 2.364170

0.987767 0 .2; + 0.441608
0.907173_ _25, _2.364170_

0
0

0

— —

 

 

 

 

 

(4.4)

To implement Kalman filtering, it is necessary to have initial 11,, and P0 values. To

obtain iro , the averages of the first 30 minute data were used for traffic flow and speed.

The corresponding variance of the flow and speed are obtained and used as an initial error

covariance matrix. The 30 minute-data contains 10 data records. Therefore,

x,{0
352,0
.520.
and
_Var(x{0) 0
,. 0 Var(xf0)
P0 = ‘
0 O
_ O 0

 

 

 

P10

 

 

0 0
0 0
Var(x{0) 0

 

O Var(x‘2‘,0 )_

(4.5)

(4.6)

59

4.1.4 Results of Incident-Free Field Data Application

A prediction of obtained data was conducted by the linear Kalman filtering algorithm.
The typical filtered and predicted values of traffic flow and speed for each lane are shown
in Figure 4.2. Good tracking can be seen in the plots. Filtered estimates are also in the 95
percent upper and lower intervals for the time period from 4:00 PM. to 8:00 PM.
Corresponding scatter plots for residuals of respective traffic flow and average speed are
in Figure 4.3. These scatter plots show good randomness around zero. The resulting mean
squared error (MSE) values, obtained from the discrepancy of predicted and filtered

estimates, for flow and speed in lane 1 and lane 2 are in Table 4.1.

Table 4.1 MSEs for Flow and Speed in Lane 1 and Lane 2

 

 

 

 

 

r Flow Speed I
| Lane 1 6.278 3.968 |
F Lane 2 5.887 5.166 ]

 

Figure 4.4 shows the changing pattern of parameter (I) over time. These are
parameters of flow in lane 1 (CD (1,1)), average speed in lane 1 (CD (2,2)), flow in lane 2
(CD (3,3)), and average speed in lane 2 (CD (4,4)). TheseCD values concentrate after initial
fluctuation. After about three fourth of time period (about 6 PM.) these values begin to
stabilize. Figure 4.5 shows the changing pattern of variance (Q) of the Gaussian noise

term in the state-space equation. Values of Q decrease as new data is obtained over time.

 

 

60

A priori error covariance changes over time is also shown in Figure 4.6. As time goes by,

these error covariances concentrate.

 

61

Data Fitting for True Flow and Observed Flow

 

 

 

 

 

 

 

 

 

 

35 . r l I r I m l ' r T
3
_C_) -i
U.
.1
B
> A
I—
(D J
U)
.0 _
O
.1
fi
1
0 5 10 15 20 25 30 35
True Flow
(8)
Data Fitting for True Speed and Observed Speed
45 I l I I 1 I I I I I Y
I
40 -— d
'O
(D 1.
(D
Q
a) 35 - -
.0
9 t
E
a) 30 + ..
(I)
.0 i ,,
O
25 — _
20 - -
15 A l 1 1 I 1 . A A l 4
15 20 25 30 35 4o 45
True Speed
0))

Figure 4.1 Data Fitting for True and Observed Flow and Speed

62

6O

 

Flow in Lane 1

 

 

,0 _ ----- Upper and Lower Limits ~ -. . " ~ . ~
— Filtered Estimates ‘J

-------- Predicted Estimates
0 I l 4 l I 1 I I I l I J I l I 1

10 20 30 4o 50 so 70 80
Time (4:00 - 8:00 PM.)

 

 

 

 

 

 

(a) Flow in Lane 1

 

601-

Average Speed in Lane 1

 

 

 

 

Time (4:00 - 8:00 PM.)

(b) Average Speed in Lane 1

Figure 4.2 Predicted and Filtered Estimates of Flow and Speed in Lane 1 and 2
and 95 Prediction Interval

Flow in Lane 2

Average Speed in Lane 2

63

50

 

 

 

 

 

Time (4:00 - 8:00 P.M.)

(c) Flow in Lane 2

 

 

 

 

 

10 20 30 4o 50 60 70 80
Time (4:00 - 8:00 P.M.)

((1) Average Speed in Lane 2

Figure 4.2 (cont’d)

Residuals for Flow in Lane 1

Residuals for Average Speed in Lane 1

64

 

 

 

 

 

 

 

 

 

I
5- I ' ' I I
-. ' I . I ' I ' ' II
I I . I I - I
-. . .- I a. I I.-
0?"-¢ ..‘ i...‘ .I 'l .- f...| II-
I If. Ir.- . ‘- ~ I '— d
4 I. I - . . - 5 I I i
I . i -I I I
-5J' I I ‘ ‘ I. I
I
‘10 T I l I l f 1 I t
20 4O 60 80 100 120 140
Time (12:00 - 8:00 P.M.)
(a) Residuals of Flow in Lane I
10
S-i '
I . l
:- I - I.- ..‘I - ' I "
I ' '- II.}‘-. I '-.lI - ‘.fi‘ .5.
0? ~.-- .."It - - I. ..-
q. l"- I ll. I..IJ'_..I
J - I f .l' -
I I - i
- .
-5..
. I
10T r T I f I r I I I
20 40 60 80 100 120 140

Time (12:00 - 8:00 P.M.)

(b) Residuals of Average Speed in Lane 1

Figure 4.3 Scatter Plots of Residuals for Traffic Flow

and Average Speed in Lane 1 and Lane 2

Residuals for Flow in Lane 2

Residuals for Average Speed in Lane 2

65

 

 

 

 

 

 

 

 

 

10 }
5- . .
I I .. I
‘ I . - I. ' I. - . ' .I
I! . I . I I I II ‘ I. -
0d . f'-‘ .I h. I.. . .. -
-. -- .I 5 f.' A I" ' . -I- .I'I
I. II. II. I.' l I. ...,
'I I - . I .. .
-5- I I I
I
'10 I I I I I I I I I I I I I
0 20 40 60 80 100 120 140
Time (12200 - 8:00 P.M.)
(c) Residuals of F low in Lane 2
10
l
I '-
5‘ I I
'- - - l: - I I
' . I i ' ; ‘- I I I
i I
o . 1.3-f I..Ifi :' ‘- f‘Ifl-.I.I '
I
i
l. I - I I I---. '. -'
.-
-5.l .
I
'10 I I I I I I I—I I I I I I
20 40 60 80 100 120 140

Time (12:00 - 8:00 P.M.)

(d) Residuals of Average Speed in Lane 2

Figure 4.3 (cont’d)

66

as; .85 mews—C C56 v... «Sufi

22¢ 85 - comm: we;

 

 

 

 

 

3: our cor om om ov om
. q _ . . . _ q . . . . d .
so
. in...)
(I)! “IRAs-clan
3‘37}? 1.
:1; _Ialil ..
8.8 fell
3.8 ...E .
2.: fall L

 

 

 

 

 

o.o

Nd

Yo

0.0

md

09

Né

v...

we

mé

9N

lHd

Error Covariance

67

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

300
250 .
— Q (1.1)
200 l, ------- Q (2,2)
\ O (3.3)
150 _\ ————— own
A.
100 l.“
L
so _
0 I
20 4o 60 so 100 120 140
Time (12:00 - 8:00 P.M.)
Figure 4.5 Q(i,i) Changes over Time
350
300 —
—— P (1.1)
250 — ---- P (2.2)
l‘ ......... P (3.3)
200 -'~.‘ - -------- P (4,4)
r
150 -
100 _\
50
o

 

Time (12:00 - 8:00 P.M.)

Figure 4.6 A Priori Error Covariance Changes over Time

68
4.2 Application to the Incident Data Set

A test was conducted to apply the incident detection algorithm on an arterial street.
Incident data were included as part of the Harrison Road data set. An incident was found
by analyzing the data set. In lane 1 (outer lane) a large headway (662 seconds) was found
between consecutive passing vehicles. The cumulative headway of the last vehicle before
that headway was 20476 seconds from 7:00 A.M. on the first day of data collection (i.e.,
12:41.4 P.M.) The cumulative headway of the first vehicle after that headway was 21138
seconds (i.e., 12:52.2 P.M.) It was assumed that an incident occurred in that lane. In
analyzing traffic in lane 2 (inner lane), an abrupt increase of traffic flow was also found
for that time period. Therefore, a lane blockage before the detector in lane 1 was inferred
from this traffic flow pattern.

The incident detection algorithm was applied to this data set and i 2 standard
deviations of predicted error covariance were used as the upper and lower intervals. An
incident is indicated if a filtered estimate is out of intervals. Figure 4.7 shows the traffic
flow change in lane 1 and lane 2. These figures show the time when the incident occurred
and how that was detected. In Figure 4.7a, the incident occurred at time step between
12:42:00 and 12:45:00 (arrow A). Using the incident detection algorithm, this incident
was detected at time 12:45:00 with a filtered flow estimate of 9.446 (arrow B) and the
lower interval value of 10.724 (arrow C) for the predicted flow estimate. The incident
was terminated in the time period between 12:51:00 and 12:54:00 (arrow D). The speed
change in lane 1 also showed the same pattern with a filtered speed estimate of 17.273

(arrow A) and lower interval of 19.579 (arrow B) for the predicted speed estimate (Figure

4.7b).

69
This incident also affected the traffic in lane 2 and was detected (Figure 4.7c). The

sudden flow increase at time 12:42:00 was detected at 12:48:00. The filtered flow
estimate was 39.817 (arrow A) with the upper interval of 33.677 (arrow B). However, the
speed in lane 2 was not affected enough to be out of intervals (Figure 4.7d). Therefore, it

took less than 3 minutes to detect the incident in lane 1 and less than 6 minutes in lane 2.

70

«:5 Earn .8 9:639:— 53889 .522: 5.9 95»:—

~ 284 E 32m A3

32¢ comm - comm: we:
ov mm on mm cm or or m

 

      

Soumamr
cm 3823 236::

85533 8662a
:53 526..

mmumEzmm U905...

=5: 53:

 

u q q q u .

 

om

 

Laue-l u! MO|-_-|

71

CV

8283 S. 9...»:—

_ 284 5 Beam 030313

62.n— ooum - comm: mEE.

mm on

mm ON mr or

 

 

u a -

 

 

 

mm

0v

1 sue-l u! peeds eBBJeAv

72

CV

5253 S. 2:»...—
N 83 s 30$ Go

22¢ comm - comm: we:
on on mm om we or m

 

 

 

- I - I d I - I - I l- d J I o

l or
or
0N
mm
on
on

10v

 

lmv

tom

 

 

mm

2 eue1 u! M0|:|

73

8.233 S. 2:3...

m 095 E 3on owmeo>< A3

65...". com .. CONN: mEC.

 

0v mm on mN ON m: or m
- q - a _ u u q - a u u u q - .
.... J
\J’ ... II.
\ , .
s * ..
~ .. l
.. .. \ a,» ......
.. .. . .. ,* . .1.
s . .. a
\ .. . v/ I.
. x ,\ x/x .
l/ t * s .. \ * l Ix
¥**\& VF‘ *l*\ I***... I I\* *\* *lxl**\* *****
III 1

 

 

 

or

or

0N

mm

3 sum u! peeds afieJeAv

Chapter 5

TRAF-NETSIM SIMULATION TEST

5.1 Simulation Test for the Algorithm

The incident detection algorithm was tested with simulated data sets. The TRAF-
NETSIM, Version 5 was selected for simulating incident-free and incident data sets. The
NETSIM (NETwork SIMulation) is one module of the TRAP family, and is a
microscopic simulation model of urban traffic (FHWA 1995). This microscopic model is
based on the behavior of individual vehicles, and provides many features. One of these
features is the ability to generate incidents in the network. Partial blockages and full
blockages in a link can be generated for various durations. In this section, incident-free
data were generated and used to test the false alarm rates. Then, incident-data were

generated to test incident detection rates and the mean detection time statistics.

5.2 Description of the Arterial Street Network

The network for the simulation was selected to represent a real-world situation. The
urban arterial street for the simulation test was Washtenaw Avenue located in Ann Arbor,
Michigan. This site was used in a previous bus-preemption signal study (Al-Sahili, and
Taylor 1996). This arterial street contained 12 intersections and stretched from the west to

the east. The lengths of links between intersections vary from 400 feet to 4000 feet.

74

75

Available network data included intersection geometry, and traffic volumes from 7:00
A.M. to 8:00 A.M.(i.e., A.M. peak hour volume). The traffic flow from the east to the
west was high in the A.M. peak hour period so this direction was selected for the test. The
traffic signal timing was optimized by using TRANSYT-7F, Release 7. The TRANSYT-
7F minimizes an objective function called the performance index (P1) to optimize the
signal timing (Wallace, Courage and Hadi 1991). The PI is a linear combination of delay
and stops, fuel consumption and excessive maximum length of queue. The best common
cycle length was 120 seconds over the network. Each direction of the street had bus
stations. A total of 34 bus stations were located along the street for both directions. The
type of bus stop location was protected, which was pull out and no-parking bus stop. The
average time the bus spends stopped at each bus station to load and unload passengers
was assumed to be 30 seconds. The mean headway between buses was also assumed to be

900 seconds. The network is represented as a link and node diagram in Figure 5.1.

5.3 Measures of Effectiveness
In evaluating various incident detection methods, the most widely used metrics are the

detection rate, the false alarm rate, and the mean detection time.

0 Detection Rate (DR) - The ratio of incidents detected to all incidents that occur

during a specified time period.

76

Number of Incidents Detected x

DR % =
( ) All Incidents

 

100 (5.1)

0 False Alarm Rate (FAR) - The ratio of false alarms to all time intervals tested during
a specified time period. There are two ways to express the false alarm rate: on-line
definition and off-line definition. The on-line definition is the percentage of false
incident messages out of the total incident messages generated by the algorithm. The
off-line definition is the percentage of incident messages generated by the algorithm
using representative incident-free data (Levin and Krause 1978). In this study, the
false alarm rate (%) per detector station is calibrated as the ratio of the number of
false alarms per hour per station to 30 records per hour because the data is aggregated

in 2 minute slices (common best cycle length). In mathematical form, it is

False Alarm Rate (%) = Number of False Alarms / Hour / Station x 100 (5.2)

30 Records / Hours

 

0 Mean Detection Time (MDT) - The average amount of time to detect an incident.
This is defined as the difference between the detection time by the algorithm and the
earliest possible detection time determined by examining both the upstream and
downstream station data of the incident location. The process of estimating the

earliest possible detection time requires subjective judgment.

77

These metrics are related with each other. For example, a lower false alarm rate is
achieved at the cost of a lower detection rate. There are no definite standards for selecting
the best combination of these metrics. The performance comparisons of various historical
detection methods is also hard to make on an absolute scale, because the definition of
incidents has varied study by study. For instance, some studies include stalled vehicles on
the shoulder as incidents, although this vehicle may not impact traffic flow (Hughes

Aircraft and JHK & Associates 1992).

5.4 Incident-Free Data Set Description

Incident-free data were generated for a false alarm check. A false alarm check was
conducted for both directions of all 13 links of Washtenaw Avenue. One-hour
simulations representing a weekday peak hour were conducted for generating normal
A.M. peak hour condition. The first day was used for the initializations of filtered and
corresponding error covariance estimates. The next two days were used to stabilize
parameters and predicted error covariances for constructing upper and lower intervals.
Then, ten days of data were generated to test the incident free condition.

This data is obtained by placing surveillance detectors at the downstream end of each
links in the network. A total of 24 detector stations were placed for each direction. They
are presence detectors and placed in through moving lanes because our interest is the
detection of blockages for the through traffic. The detector provides traffic measurements
of vehicle counts, average speeds, and percent occupancy. These measurements are

collected every 120 seconds (best common cycle length in the network). Therefore, 30

78

records were created each hour (i.e., a day). For the ten-day incident free data set, a total
of 300 records were created at each detector station. Therefore, a total of 7,200 records of
data were generated for the 24 detector stations in the network. A record contained traffic
measurements of each through lane. For example, six traffic measurements were in a
record for a two-through lane link (i.e., vehicle counts, average speed, and percent

occupancy in each through lane, respectively).

5.5 Incident Data Set Description

As in the incident-free data generation, one-hour simulation representing a day was
conducted for generating incidents in the A.M. peak hour. As in the incident-free data set,
the first day was used for initialization, and the second and the third days for stabilization
were conducted with incident-free data. The fourth day incident data was generated. The
fifth day was incident-free and represented a recovery to normal traffic flow after an
incident in the fourth day. Therefore, the predicted error covariance for the upper interval
and the lower interval in the third day was used for the incident-check on the fourth day

data set.

5.5.1 Link Selection for Incident Generation

There are a total of 13 links on Washtenaw Avenue in the network. Two links were
selected for incident detection tests. This selection was based on traffic flow and
geometric conditions. The link 34 was selected because it represented high traffic and a

long link length. A bus station was located 700 feet away from a stop bar of the

79

downstream intersection. Link 45 was selected because it represented high traffic and a
short link length. A bus stop was located 100 feet away from a stop bar of the
downstream intersection. Intersections associated with these links are node 3, node 4, and
node 5. Level of Service (LOS) of these intersections are in Table 5.1. These LOSS are
obtained from the Highway Capacity Software (McTrans Center 1995). Their geometry

and traffic conditions are shown in Figure 5.1-5.3.

Table 5.1 LOS in Node 3, 4, and 5

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Direction Lane Node 3 Node 4 Node 5
Movement

Left D A A

EB Through C D A A A A
Right C A* A*
Left D C B

WB Through C C B B B B
Right 0" B 8*
Left F N/A E

NB Through D E E E
Right C E* D*
Left D N/A D

SB Through D D E D D
Right D* D* D*

 

 

* Shared Movement with Through Movement

 

80

xuflwo:

a 30> AquEm cob: buwosucfiz

teams—33m
vmow

6 mm;

baa. 806.. .8». @305
boom .83 .8E .nm .3. boom 6 33
Son
8. 6 9
cow

035>< 350.533 .0091. a
.owo.

.99.. ...e e 6
...o .e ..e ..

Sow a
flow a
. .oom
burg—flu .
a 5 :52 221.3; caxbzm e e 0

93m :00 ”Egon

52% a: e a: :3 >< 33:83
£325 .m

2

Figure 5.1 Simulation Network

81

 

 

 

 

 

 

 

 

l|l

 

J301
_———->664 231®68 “6242:;—
§§l73 f0

392 224

Figure 5.2 Node 3 Geometric and Hourly Traffic Volume

82

 

 

231 66 ll
—/39 \
jg“ ® l664é——

2 94%

ll 96

Figure 5.3 Node 4 Geometric and Hourly Traffic Volume

83

12 10 23
4_<897 ® 1628:—
28
68
99 13 20

Figure 5.4 Node 5 Geometric and Hourly Traffic Volume

84

5.5.2 Incident Generation in TRAF-NETSIM

In TRAP-NETSIM, Version 5, an incident can be generated either in a link or an
intersection. An incident in this simulation program represents long-term blockages due
to illegal parking or vehicle breakdown. A limitation on an incidents generation in this
simulation program is that they represent only one-lane blockages instead of full-link
blockages (e.g., two lane blockages for two lane roadways). Another limitation is the
randomness of the incident location along a subject link. For example, if a user specifies
an incident on a certain link, an incident can occur at either end of the downstream link or
the middle of the link. Since these incident types and incident location are important
factors for an incident detection algorithm test some techniques were applied to generate

realistic incidents.

o Full-Link Blockages Generation - A full-link blockage was created by using a
signalized dummy node in a link. For normal traffic flow, the signal was given all
green for the interested direction. When an incident occurs, the signal was given 5
seconds of green and 115 seconds of red phases. This 5-second green time represented
the time period for metered vehicles to pass the incident location in a 120-second
cycle length. The 115 seconds represent an incident event that blocks oncoming

vehicles.

0 Incident Location Specification - A one-lane blockage in a multi-lane roadway can

be made using the microintersection feature in NETSIM. A dummy node for a

85

microintersection in the interested link was made. Since this was a microintersection,
it was possible to specify a blockage in the intersection that was a particular incident
location by specifying the distance between adjacent nodes. For example, a blockage
in lane number 1 in an approach link (interested link) and lane number 0 on a cross
street represents the incident in lane 1 of the approach link. By moving this dummy
node (microintersection) along the interested link, an incident can be placed in

various locations.

5.5.3 Various Incident Generation
Mid-block incidents can be represented by the type, and duration, as well as the

location on a link.

Incident Type: For two-lane streets, there are two types of incidents. An incident
blocking one lane, lane 1 (outer lane) or a lane 2 (inner lane), is a partial blockage which
reduces the capacity of the roadway. These incidents allow drivers to switch their driving

lane to the non-incident lane. The other incident type is a full-link blockage.

Incident Duration: The duration of incidents is variable. That is, there is no time
duration common to all incidents. In this study, the duration of an incident was studied in
10 minute increments. For one-lane blocks, incidents were assumed to last 10, 20, 30, 40,
and 50 minutes. For full-link blocks, incidents were assumed to last 10, 20, and 30

minutes.

86

Incident Location: The location is an important factor for incident detection techniques.
It is easy for an algorithm to detect an incident occurring near a detector since the impact
of the incident will be sensed quickly. The closer an incident occurs to a detector, the
faster the incident is detected. For link 34, three locations were selected for incident. They
were an upstream location (100 feet away from an upstream intersection), a downstream
location (100 feet away from a detector station), and a middle location (1050 feet away
from both the upstream intersection and the downstream intersection). For link 45, two
locations were determined since the link was short. They were the upstream (100 feet
away from the upstream intersection) and the middle location (256 feet away from both
the upstream and the downstream intersections). These locations were applied to the one-
lane blockages and the full-link blockages.

As a result, a total of 65 incidents were generated for the incident test. For link 34, a
total of 30 and 9 incidents were simulated for the one-lane blockage and full-link
blockage, respectively. For link 45, a total of 20 and 6 incidents were simulated for one-
lane blockage and full-link blockage, respectively. The characteristics of these incidents

are shown in Table 5.2 and Table 5.3.

Table 5.2 Incidents in Link 34 Descriptions

87

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Incident Type Lane Blocked Incident Location Duration
One-Lane Blockage 1 Middle 10 minutes
One-Lane Blockage 1 Middle 20 minutes
One-Lane Blockage 1 Middle 30 minutes
One-Lane Blockage 1 Middle 40 minutes
One-Lane Blockage 1 Middle 50 minutes
One-Lane Blockage 2 Middle 10 minutes
One-Lane Blockage 2 Middle 20 minutes
One-Lane Blockage 2 Middle 30 minutes
One-Lane Blockage 2 Middle 40 minutes
One-Lane Blockage 2 Middle 50 minutes
One-Lane Blockge 1 Downstream 10 minutes
One-Lane Blockage 1 Downstream 20 minutes
One-Lane Blockage 1 Downstream 30 minutes
One-Lane Blockage 1 Downstream 40 minutes
One-Lane Blockage l Downstream 50 minutes
One-Lane Blockage 2 Downstream 10 minutes
One-Lane Blockage 2 Downstream 20 minutes
One-Lane Blockage 2 Downstream 30 minutes
One-Lane Blockage 2 Downstream 40 minutes
One-Lane Blockage 2 Downstream 50 minutes
One-Lane Blockage l Upstream 10 minutes
One-Lane Blockage 1 Upstream 20 minutes
One-Lane Blockage 1 Upstream 30 minutes
One-Lane Blockage l Upstream 40 minutes
One-Lane Blockage 1 Upstream 50 minutes
One-Lane Blockage 2 Upstream 10 minutes
One-Lane Blockage 2 Upstream 20 minutes
One-Lane Blockage 2 Upstream 30 minutes
One-Lane Blockage 2 Upstream 40 minutes
One-Lane Blockage 2 Upstream 50 minutes

 

 

 

88

 

 

 

 

 

 

 

 

 

 

 

Table 5.2 (cont’d)

Incident Type Lane Blocked Incident Location Duration
Full-Link Blockage 1,2 Middle 10 minutes
F ull-Link Blockage 1,2 Middle 20 minutes
F ull-Link Blockage 1,2 Middle 30 minutes
Full-Link Blockage 1,2 Downstream 10 minutes
F ull-Link Blockage 1,2 Downstream 20 minutes
Full-Link Blockage 1,2 Downstream 30 minutes
F ull-Link Blockage 1,2 Upstream 10 minutes
Full-Link Blockage 1,2 Upstream 20 minutes
Full-Link Blockage 1,2 Upstream 30 minutes

 

 

 

 

 

Table 5.3 Incidents in Link 45 Descriptions

89

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Incident Type Lane Blocked Incident Location Duration
One-Lane Blockage 1 Middle 10 minutes
One-Lane Blockage 1 Middle 20 minutes
One-Lane Blockage 1 Middle 30 minutes
One-Lane Blockage 1 Middle 40 minutes
One-Lane Blockage 1 Middle 50 minutes
One-Lane Blockage 2 Middle 10 minutes
One-Lane Blockage 2 Middle 20 minutes
One-Lane Blockage 2 Middle 30 minutes
One-Lane Blockage 2 Middle 40 minutes
One-Lane Blockage 2 Middle 50 minutes
One-Lane Blockage 1 Upstream 10 minutes
One-Lane Blockage l Upstream 20 minutes
One-Lane Blockage 1 Upstream 30 minutes
One-Lane Blockage 1 Upstream 40 minutes
One-Lane Blockage 1 Upstream 50 minutes
One-Lane Blockage 2 Upstream 10 minutes
One-Lane Blockage 2 Upstream 20 minutes
One-Lane Blockage 2 Upstream 30 minutes
One-Lane Blockage 2 Upstream 40 minutes
One-Lane Blockage 2 Upstream 50 minutes
Full-Link Blockage 1,2 Middle 10 minutes
F ull-Link Blockage 1,2 Middle 20 minutes
F ull-Link Blockage 1,2 Middle 30 minutes
F ull-Link Blockage 1,2 Upstream 10 minutes
Full-Link Blockage 1,2 Upstream 20 minutes
Full-Link Blockage 1,2 Upstream 30 minutes

 

 

90

5.6 Results of the Incident-Free Test

The incident free condition was tested to check the false alarm rate. An incident
indication was made by comparing the filtered estimates and the prediction intervals of
the predicted estimate in each time increment. If any filtered estimate is out of intervals,

then an incident alarm is recorded. For the upper and the lower intervals, 95 percent (e. g.,
i2,/P(—)) prediction intervals were used. Raw traffic counts, percent occupancy, and the

average speed in each lane were smoothed by the moving average of 3 time intervals.
Results showed that a total of 36 false alarms were indicated in the network. These
false alarms were accumulated from 24 detector stations over 10 hours. This is 0.15 false
alarms per hour per station. If the system was operated for 2 hours a day, then there could
be 0.3 false alarms per day per station. For the network, a total of 7.2 false alarms would
occur in a day. Since 0.15 false alarms were made in an hour (30 records of data an hour)
per station, the false alarm rate is 0.5 percent. For the direction of high traffic (e.g., from
east to west), a total of 4 false alarms were indicated. This is 0.033 false alarms per hour
per station. If the system were operated for 2 hours a day, then there would be 0.067 false
alarms per day per station. For the high traffic direction, this is 0.800 false alarms per day
per station. The false alarm rate is 0.111 per hour per station. They are shown in Table

5.4.

91

Table 5.4 Results of Incident Free Test for 95 Percent Prediction Intervals

 

 

 

 

 

 

 

 

 

 

Statistics Network High Traffic Direction
(Both Directions) (from east to west)

Total # ofFA/10 Hr/

Total # of Stations 36 4

FA/Hr/ Station 0.150 0.033

F A/2Hr/Station 0.300 0.067
FA/2Hr/Total # of Station 7.200 0.800

F AR/Hr/Station (%) 0.500 0.111

 

By analyzing the result, a distribution of false alarms with the associated traffic control

variables are shown in Table 5.5.

92

Table 5.5 Distribution of False Alarms over Detector Stations

Station No. No. of FA Traffic Control Variables*
1 1 4

2

2

1
1
3
2
1
1
1

20
2
4
Total 36 52 34 2 5

*The number in a parenthesis indicates a traffic control variable. The number before
a parenthesis indicates the number of that traffic control variables in the parenthesis.

 

Traffic Control Variable Description

1: traffic count in lane 1

2: percent occupancy in lane 1
3: average speed in lane 1

4: traffic count in lane 2

5: percent occupancy in lane 2
6: average speed in lane 2

A total of 11 detector stations showed false alarms. Eight of these stations were in the
low traffic flow direction (i.e., from west to east). These are even-numbered stations.
Three of these stations were in the direction of high traffic flow (i.e., from east to west).
These are odd-numbered stations. This result suggests that the incident detection

algorithm is more effective in the high traffic direction. This is illustrated by comparing

93

the stations with the most false alarm and the station with no false alarm. Station 20 and

station 5 one-hour traffic statistics are shown in Table 5.6, and 5.7, respectively.

Table 5.6 One-Hour Traffic Statistics at Detector Station 20

 

 

 

 

 

 

 

 

 

 

 

 

 

Traffic Control Variables
1 2 3 4 5 6
Average 7.433 14.660 25.033 7.033 15.570 25.457
Std. Dev.* 0.971 4.220 1.948 1.426 6.385 2.804
*Standard Deviations

By looking at the statistics for station 20, the average time interval between
consecutive vehicles can be calculated. Assuming a 2 second discharge interval, the
effective green time for the signal is 100 seconds (i.e., 98 seconds for green, 4 seconds for
yellow, and 2 second for the lost time). The average number of vehicles per lane is 7.233

for that phase. Therefore,

129— = 13.826 seconds is the average time interval (5.3)

7.233

A measure of the variations in the percent occupancy in lane 1 and lane 2 can be seen by

taking the ratio of the standard deviation to the average, respectively.

4.220
14.66

 

x 100 = 28.79% in lane 1 (5.4)

94

6.385

x100 = 41.04% in lane 2 (5.5)
15.570

 

As seen in Table 5.5, the most sensitive false alarm variable is the percent occupancy
in lane 2. It is probable that the high variation in percent occupancy in lane 2 caused
many false alarms. This high variation in percent occupancy might be attributable to low
traffic with a high average time interval between consecutive vehicles. This hypothesis is
confirmed with data from the no-false detector station.

Detector station 5 produced no false alarms in the test. Typical one-hour traffic

statistics is shown in Table 5.7.

Table 5.7 One-Hour Traffic Statistics at Detector Station 5

 

 

 

 

 

 

 

 

 

Traffic Control Variables
l 2 3 4 5 6
Average 26.133 24.783 33.960 27.167 24.507 33.750
Std. Dev. 2.389 4.639 1.561 1.683 5.697 1.143

 

 

 

 

The effective green time for the associated phase of the intersection (i.e., node 4) is 97
seconds (i.e., 95 seconds for green, 4 seconds for yellow, and 2 second for the lost time).

The average per lane volume is 26.650 vehicles for that phase. Therefore,

97
26.650

 

= 3.640 seconds is the average time interval (5.6)

95

Variation in the percent occupancy in lane 1 and lane 2 as shown by the ratio of

standard deviation to the average are:

4.639

 

 

x 100 =18.718% in lane 1 (5.7)
24.783
5697 x 100 = 23.246% in lane 2 (5.8)
24.507

By comparing these results to the detector station with the greatest number of false
alarm confirmed that the high variation in percent occupancy due to low traffic be the

cause of the high false alarm rate.

5.7 Results of the Incident Test
Results of the incident tests are shown in four sub-sections. First, a partial-blockage in
link 34 is shown, followed by a full-blockage in link 34. Next, a partial blockage and a

full blockage in link 45 are discussed.

5.7.1 Results of One-Lane Blockage in Link 34

An incident detection evaluation was conducted at both the upstream and the
downstream stations. The upstream station (station 3) and the downstream station (station
5) showed the same number of incident detections. A total of 21 out of 30 incidents were
detected in both stations. This is a 70 percent detection rate at either detector station.

However, the mean detection time in these stations was different. The mean detection

96

time is 8.286 minutes with a standard deviation of 3.364 for the upstream station while
the mean detection time is 4.476 minutes with a standard deviation of 1.250. This results
implies that detection efficiency is higher in the downstream detector station.

There are incidents that were detected by the upstream detector station that were not
detected by the downstream station, and vice versa. Considering both stations together, all
the incidents were detected. The mean detection time for the combined data is 4.8

minutes with the standard deviation of 1.243.

5.7.2 Results of Full-Link Blockage in Link 34

The incident detection algorithm for full-blockages in link 34 was evaluated in the
same way. As expected, the impact of a full-blockage is very large. At both the upstream
and the downstream detector stations, all 9 incidents were detected regardless of the
incident location, respectively. The mean detection time for the upstream station is 6.667
minutes with the standard deviation of 1.414. The mean detection time for the
downstream station is 4 minutes with a standard deviation of 1.000. By considering both
stations together, the mean detection time is 3.778 minutes with a standard deviation of
0.667. While the incidents were eventually detected by the upstream detector station, all

of them were more rapidly detected by the downstream detector station.

97

5.7.3 Results of One-Lane Blockage in Link 45

Since this link is short (i.e., 530 feet), the impact of an incident is easily detected on
either the upstream or the downstream station. All one-lane blockages occurring at the
upstream or the middle location were detected by the upstream and downstream detector
stations. The mean detection time for the upstream detector station is 2.9 minutes with a
standard deviation of 1.373. For the downstream detector station, the mean detection time
is 4 minutes with the standard deviation of 0.918. On this link, the overall performance of
the upstream detector station was better than the downstream station because no incidents
were assumed at the downstream location. By considering both stations, all the incidents

were detected as well. The mean detection time is 2.7 minutes with a standard deviation

ofl.174.

5.7.4 Results of Full-Link Blockage in Link 45

As expected, all the incidents were detected by either detector station. The mean
detector time for the upstream detector station is 2.333 minutes with a standard deviation
of 0.817. For the downstream detector station, the mean detection time is 4 minutes with

0.000 standard deviation.

5.8 A Sensitivity Test for Various Prediction Intervals
There is a trade-off among the detection rate, false alarm rate, and the mean detection
time. For example, wider upper and lower intervals would be less sensitive in detecting

an abnormality of the control variable, resulting in fewer false alarms, but also a lower

98

detection rate. Narrower intervals would be more sensitive providing more false alarms
but better detection rate. If a confirmation check is used to reduce the number of false
alarms, then the incident rate and the false alarm rate would be better, but they would
require a longer mean detection time.

The previous results were from a test using 95 percent prediction intervals (i.e., 2 a“

of the predicted control variable estimates for the upper and the lower intervals). Wider

intervals using 2.1 0'“ , 2.5 0i]. , and for 3.0 oi, were tested with the same incident-

free and incident data sets to investigate how the false alarm rate, detection rate, and the

mean detection time change.

5.8.1 False Alarm Changes over Different Intervals

False alarms were tested with 3 different intervals to see the overall change. These

intervals were 2.0, 2.5, and 3.0 a“ . The results are shown in Table 5.8.

Table 5.8 False Alarm Statistics Changes for 2.0, 2.5, and 3.0 aik

 

Interval (No. of ai k ) Changes

 

 

 

 

 

 

 

 

 

 

2.0 2.5 3.0
FA/2Hr/Network 7.2 1.80 0.60
FA/2Hr/Station 0.3 0.08 0.03
F AR/Hr/Station (%) 0.5 0.08 0.04
FA/2Hr/High Traffic Direction 0.8 0.20 0.00

 

99

The number of false alarms decreases as the upper intervals and lower intervals

increase. The ratio of the changes in the number of false alarms from 2.0 to 3.0 can is
significant and one to twelve. At 3.0 0's]. , there are no false alarms detected for the high

traffic direction.

5.8.2 Detection Rate and Mean Detection Time Changes over Different Intervals
The detection rates and mean detection time of one-lane blockages and full-link
blockages in link 34 and link 45 are shown in Table 5.9, 5.10, and 5.11, 5.12,

respectively, for the same interval changes.

100

Table 5.9 Detection Rates of One-Lane Blockage over 2.0, 2.5, and 3.0 0'“ Changes

 

 

 

 

 

 

 

 

 

(percent)
Detector Location Interval Changes

2.0 2.5 3.0
Upstream 70 53.3 46.7
Link 34 Downstream 70 56.7 6.7
Both 100 83.3 46.7
Upstream 100 100 100

Link 45 Downstream 100 70 5
Both 100 100 100

 

 

 

 

 

Table 5.10 Detection Rates of Full-Link Blockage over 2.0, 2.5, and 3.0 aik Changes

 

 

 

 

 

 

 

 

 

(percent)
Detector Location Interval (No. of 0') Changes

2.0 2.5 3.0
Upstream 100 100 88.9
Link 34 Downstream 100 100 33.3
Both 100 100 100
Upstream 100 100 100

Link 45 Downstream 100 100 0
Both 100 100 100

 

 

 

 

 

 

 

101

Table 5.11 Mean Detection Time of One-Lane Blockage over 2.0, 2.5, and 3.0 0'“
Changes (minutes)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Detector Location Interval Changes

2.0 2.5 3.0

Upstream Mean 8.29 9.13 10.71

Std. Dev. 3.37 3.26 4.61

Link 34 Down- Mean 4.48 5.29 5.00
stream Std. Dev. 1.25 1.21 1.41

Both Mean 4.8 6.00 9.87

Std. Dev. 1.24 2.38 4.87

Upstream Mean 2.9 9.13 10.71

Std. Dev. 1.37 3.26 4.61

Link 45 Down- Mean 4.00 5.29 5.00
stream Std. Dev. 0.92 1.21 1.41

Both Mean 2.70 6.00 9.87

Std. Dev. 1.17 2.38 4.87

 

Table 5.12 Mean Detection Time of F ull-Link Blockage over 2.0, 2.5, and 3.0 on

Changes (minutes)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Detector Location Interval Changes

2.0 2.5 3.0

Upstream Mean 6.67 7.78 7.75

Std. Dev. 1.41 2.11 2.25

Link 34 Down- Mean 4.00 4.67 6.00
stream Std. Dev. 1.00 l .00 0.00

Both Mean 3.78 4.44 7.11

Std. Dev. 0.67 0.88 2.03

Upstream Mean 2.33 2.33 2.67

Std. Dev. 0.82 0.82 1.03
Link 45 Down- Mean 4.00 5.00 N/A*
stream Std. Dev. 0.00 1.67 N/A*

Both" Mean 2.33 2.33 2.67

Std. Dev. 0.82 0.82 1.03

 

*No detection is made.

"This statistics are the same as that of the upstream station because the upstream station

is faster than the downstream station.

 

 

102

For the incident tests, the detection rates and mean detection times were as expected.

The detection rate at a 2.0 0“ decreases about 50 percent at a 3.0 0's]. on both detector

stations for a one—lane blockages in link 34,. The detection time increases about two times

from the 2.0 a“ to the 3.0 an . The link 45, all the blockages were detected at both
stations even at the 3.0 0'“ since this link is short. The mean detection time in this link
increases about 4 times at the 3.0 0i]. . Full-link blockages are much more critical and

they were all detected on both stations earlier than one-lane blockages.

5.8.3 A Detailed Sensitivity Test for Link 34
Since the statistical changes in the false alarm test and one—lane blockage test in link

34 are relatively change, more detailed interval tests were made using 2.1 to 2.9 an . For

the network itself, the number of false alarms decreases as the intervals are increased. For

the high traffic direction, no false alarms were detected after reaching a 2.7 0i]. , as

shown in Table 5.13. The results are also shown in Figures 5.5-5.10.

Detection rates for the one-lane blockage in link 34 shows a consistent decrease at
both detector stations. These results are shown in Table 5.14, and Table 5.15,
respectively. At the upstream detector station the detection rate slowly decreases while

the detection rate at the downstream detector station drops sharply afier the 2.5 a,“ as in

Figure 5.11.

103

Generally the mean detection time at either station increases as intervals go higher.
But as seen in Figure 5.12, and Figure 5.13, respectively, the mean detection time

fluctuates after the 2.5 a“ . This is due to the reduction in the number of detections. That

is, when an incident that was detected at one time step is not detected at the next time
step, the number of recorded detections is reduced. The mean detection time computed

with the lower number of detections may also be reduced. For example, at the 2.50“ ,

the detection rate for the downstream detector station is 56.7 percent (17 detections out of

30 incidents) while the detection rate at the 2.6 0'“ is 46.7 percent (14 detections out of
30 incidents). The simple mean detection time at the 2.5 a“ is 5.25 minutes while it is
5.14 minutes at the 2.6 aik because one of detections (i.e. 8 minutes of detection time)

was missed, resulting in a lower mean detection time. Overall, the detection time, when

considering both stations together, consistent increases as the a“ increase as seen in

Figure 5.14.

The detection rate and the mean detection time of the full-link blockages in link 34 are
shown in Table 5.16, and Table 5.17, respectively. Since the impact of these blockages
are large, all the blockages are eventually detected at all intervals at both stations. The

detection rate at the downstream detector station shows drops suddenly after 2.5 0'“
while the upstream station shows a nearly constant rate until 2.8 032;. . These results are

shown in Figure 5.15.
The mean detection time at the upstream station shows a rather consistent pattern after

2.4 aik while this value consistently increases at the downstream station. However, the

104

mean detection time is still shorter at the downstream station. By considering both
stations together, the mean detection time consistently increases. They are shown in
Figures 5.16-5.18. The standard deviation of the mean detection time at each generally

increases as the 0i]. increases.

105

 

8:90

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

000.0 000.0 000.0 000.0 0m0.0 0N0.0 0N0.0 000.0 000.0 _ w _.0 _ 0 M0 \gA$v.m<m
Seam
000.0 000.0 000.0 000.0 20.0 :00 :00 000.0 0m0.0 500.0 $0.0 EEG/E :38on
8:80
000.0 000.0 000.0 000.0 000.0 000.0 000.0 A. 00.0 mm0.0 000.0 000.0 302.0 059.0.
$030th
0.0 0.0 0.0 0.0 N0 N0 N0 v.0 0.0 000.0 000.0 BAN); :05
~8onqu
0 0 0 0 0 _ 0 m 0 v v E30 <<m
8:80
$0.0 30.0 050.0 500.0 :_.0 03 .0 50—0 00N.0 NON.0 0 _ 0.0 0.0 \§A$vm<m
.8080
mm0.0 000.0 N000 0m0.0 500.0 050.0 02.0 mNfio m: .0 N20 0.0 \ENZm
.8080
m _0.0 n 0 0.0 50.0 30.0 000.0 000.0 000.0 000.0 000.0 000.0 0m _ .0 ES); 0.52502
619,302
0.0 0.0 _ v; 0; 0. _ «am 0.0 NV 04. NS ESQ/0,0
08332
0 w m h 0 0 N0 2 3 mm 00 E20<<m
0.0 0.N 0.N EN 9N m.N 1N 0.N Nd —.~ 0.N
3::

8052—0 338:—

 

 

 

m_a>._8:— 0:20.005 .35 3052-0 8.32 3.3,.— mfim «Ssh.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

80:25 .5...an

 

50..“q 00.0V 00.0 0N0 00.0. 00.0 0v.N 0v.m 500 00. 0 0N; .>oQ .90

50.0 50.0 :0 00.0 00.0 0.0 05.0 55.0 00.0 00.9V 0.0 582 50m
:2 0 _A 00.0 #0.— 00; 0N; 0N._ 3: 0; 0m; 0m.“ <63 .90 5390
0.0 00.0 50.0 00.0 3.0 0N0 2.0 0 00.0 0.0 006 :82 -8500—
~0.v 0...» 0.0 50.0 00.0 0N0 0N0 :0 00.0 N00 00.N Sufi .90 «58%

5.00 05.00 _5.0 0Y0 0.0 0_.0 00.0 N00 N00 05.5 05.5 582 -03
0.0 0.0 0.0 5.N 0.N 0.N v.0 0.N 0.0 fiN 0.N

 

106

$859.5 2.9.8:: 25.8.5: .82. 0 2:5 E 00.2er onus—6:0 .80 00:25 «8:. .833qu :32 0—0 035,—.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5.00q 0.00 0.00 0.05 5.05 0.00 0.00 5.00 5.00 000 000 9cm

5.0 0.0. 0.00 5.00 5.0V 5.00 0.00 0.00 0.00 5.00 05 8858500

5.00 5.09V 5.0V 5.0V 0.00 0.00 0.00 5.00 0.00 0.00 05 88503
0.0 0.0 0.0 5.0 0.N 0.0 v.0 0.0 0.0 fin 0.0

gouge;

8052—0 338:— :98on

 

c.8983 2.3.5.:— anoba .52. V0 «9: 5 003—8:— onus—6:9 .8.— 30530 35m .533qu 3.0 035—.

 

 

107

 

 

 

 

 

 

 

00.0 00.0 00.: 0:0 00.: 00.0 00.0 50.0 00.: 50.0 50.0 .>0Q .90

:.5 50.0 00.0 00 .0 00.:V 30.0 30.0 and 00.00 05.0 05.0 :8: 500:
00.0 00.0 00.0 00. _ 00.0 00. : 00. : 00.0 mm: 00. ~ 00. : .80: .90 E890
00.0 0.0 50.0 50.9V 50 .v 50...V 50.:u $0.0 mmé 00...V 00.:V :82 -8509
00.0 00.0 :0 :0 :0 :.N :0 00.: 00.: 00.: I: .>0Q .90 E890
05.5 05.5 05.5 05.5 05.5 05.5 05.5 :.5 00.0 00.0 50.0 :82 -03
0.0 0." 0.0 5.N 0.N 0.0 Ya 0.0 ~.N fin 0.N

 

 

 

 

 

 

 

 

 

 

 

 

 

00052.0 .038:—

 

 

Am00=:_:.v 0.520.:— 0:0..0.«.=D 00>: 00 0.5-— 9 005—005 vii-=5: .8.— 00:a__0 0E:- :900000D :8: 50.0 030,—.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

00: 00 _ 00 : 00 _ 00 _ 00 _ 00 _ 00 : 00 _ 00 _ 00 _ 900:

00.00 00.00 50.00 50.00 05.55 00: 00: 00: 00: 00: 00: E89832:

00.00 00.00 00: 00. 00_ 00_ 00: 00: 00: 00: 00: 88:33
0.0 0.0 0.0 5.N 0.0 0.0 1N 0.0 0.0 0.0 0.0

8080.:

80:20 .3000:— 090003:

 

 

3:09.03 0.5.000:— .:0..0.§: 00>: v0 0...: 5 003—005 09549.: .30 00:30 003— :030002— 0—0 030,—.

 

 

108

 

 

 

 

 

 

g 4- [ —x—— FA/Hr/Network ]

E

<

a 3*

(0

LI.

33 2-1 \

3 .

E \

3 1- *\*

Z \x\¥

X\*\

o.,.,.,.l .,sr.

' l ' l ' l
2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0
Limit (sigma) Changes

Figure 5.5 Number of False Alarms for the Network over 0'“ Changes

0.20

 

 

—x-— FA/Hr/Station for Network

 

0.15

0.10:
\

005 ..i \l\

0.00 7* I T I Y I V I U I I I I I V l V I V
2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0

Limit (sigma) Changes

Number of False Alarms

I\

\*\*\
*\
K

 

 

 

Figure 5.6 Number of False Alarms per Station for the Network over 0’“ Changes

109

 

 

 

 

 

0.5-H
F .
L ——-— FAR/Hr/Station for Network 1
0.4 ~
:0? .
d)
‘5 0.3 r-
n:
E
(U
2 0.2 —
d.)
2
63
LL 0.1 - \.\'\I
i- \K‘
00 4 L A l 1 l n l A 1 i l ‘ l + l L l

 

2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0
Limit (sigma) Changes

Figure 5.7 False Alarm Rate per Station for the Network over on Changes

 

 

 

 

 

 

 

0.5
A ' u a
l L + FNHr/th Traffic Direction
0) 0.4 —
E
E
<
a, 0.3 -
E
(U
Ll.
“5 0.2 —
S-
o
.0
E
0.0 I I I l I l f fi' I ‘ l I I ‘ I

 

U I I
2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0
Limit (sigma) Changes

Figure 5.8 Number of False Alarms for High Traffic Direction over on Changes

llO

 

 

 

 

 

0 05
m —x— FA/Hr/Station for High Traffic Direction I
E O 04 -
5
<
E 0.03 -4
CU
LL -i
s.—
S o 02 a
a)
.D 4
E
z 0.01 -
000 Y I Y I V T V I i ‘7 I V I I l

 

I l r '
2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0

Limit (sigma) Changes

Figure 5.9 Number of False Alarms per Station for High Traffic Direction
over 0'“ Changes

 

 

 

 

 

 

 

 

0‘“ [ —I— FAR/Hr/Station for High Traffic Direction]
A 0.12-
o\° u
2 0.10-4
m
m 008
E ' .
m
:2 0.06~
m 4
% 0.04-i
LL 4 .
0.02-i
J
ooo'i'T‘TTfifr'I'lfl'l

 

20 2.1 2.2 23 2.4 2.5 2.6 2.7 2.8 2.9 3,0
Limit (sigma) Changes

Figure 5.10 False Alarm Rate per Station for High Traffic Direction
over 0'“ Changes

Ill

 

100-. .\

9° ‘ .\I
4 \-
80 'i \.

.4
70
N: \-\'
so - N *\ _—

50-

 

 

4o-

   
 

3O
20 s ——I— Both Detector Stations
10 ‘ —-— Upstream Detector Stations

1 + Downstream Detector Stations

0 U I Y I V 1 Y I V I U l I I I l T l V
2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0

Limit (sigma) Changes

 

Detection Rate (%) for Link 34

 

 

 

 

 

Figure 5.11 Detection Rate Changes for One-Lane Blockage
in Link 34 over Interval Changes

 

 

 

 

 

 

 

 

 

 

9‘3 ‘6 - __ _
a ‘ 1
.E 14 -
E « _ __
V 1,. w—
0 12 -i “*7 __ T
E ‘ —I
i— 1° - L,
d /‘*./
.9 a- _ /"
‘6 ‘ L d_
2 5 - L ._. _‘L __ _JL .sL. _L
o . _
0 4 - . .
c —I— Mean Detection Time for
3 2- One-Lane Blockage in Link 34
2 « (Upstream Station)

 

 

 

 

 

 

0 ‘ I ' I ' I ' I I r I ' I ' I ' I ‘ I

. . . .
1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.3 2.9 3.0 3.1
Limit (sigma) Changes

Figure 5.12 Mean Detection Time (minutes) for One-Lane Blockage in Link 34
with Bars of One Standard Deviation (Upstream Station)

112

 

5-

‘ /\-—_
51 /

.i ______

.1

 

 

Mean Detection Time (minutes)

 

 

 

 

 

3—
2‘ —I—— Mean Detection Time for
1; One-Lane Blockage in Link 34
1 (Downstream Station)
OTI'I'I i ljfijl'I‘I‘

 

. , . , 1 .
1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.5 2.7 2.8 2.9 3.0 3.1
Limit (sigma) Changes

Figure 5.13 Mean Detection Time (minutes) for One-Lane Blockage in Link 34
with Bars of One Standard Deviation (Downstream Station)

 

 

 

 

 

 

 

 

 

 

 

 

 

16 >-
E 14 ' —I— Mean Detection Time for T _L
3 _ One-Lane Blockage in Link 34
E, 12 _ (Both Station)
0
.§ 10 _ /‘
l- b T
S e -
‘6
2’. 6 - /u/
8 . / / __
c 4 -
m _ _
0
2 2 k
0 4 l 1 l 1 J A I A i 1 I 1 l A l L #4 l ‘ l ‘
1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0 3.1

Limit (sigma) Changes

Figure 5.14 Mean Detection Time (minutes) for One-Lane Blockage in Link 34
with Bars of One Standard Deviation (Both Upstream and Downstream Stations)

113

100‘! F+ * I + a \I
W
90-4 ._—_
80 -1
.4
70 -i

 

 

 

 

 

V

m

x

.E

_J

.c‘z

s 6°:

2: so -

g 40 A

5 30: -——I— Both Detector Stations

'5 . —o— Upstream Detector Stations
.3 2°: + Downstream Detector Stations
0 10 -

 

 

 

o I I I I I I 1 I fi' 7 I I I I I I I U
2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0

Limit (sigma) Changes

Figure 5.15 Detection Rate Changes for Full-Link Blockage
in Link 34 over Interval Changes

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

A 10 -_ 3- -_ __ __
w i
2 90
g i
'- 8“ II ii- Ill
5 /i‘ T T
a: 7‘ / —- /
E
l- 6‘ —~ —— —— —— —— __ __
C .1
.9 5
13' 4-
2 Ii
3 3 - —I— Mean Detection Time for
c 2 - Full-Link Blockage in Link 34
m . .
o 1_ (Upstream Station)
2
0 ' r ' I ' fi' I 'f ' I V a I T r

 

r I U l I’ V
1.9 2.0 2,1 2.2 2.3 2.4 2.5 2.6 2.7 2.3 2.9 3.0 3.1
Limit (sigma) Changes

Figure 5.16 Mean Detection Time (minutes) for Full-Link Blockage in Link 34
with Bars of One Standard Deviation (Upstream Station)

114

 

 

 

 

 

3? .
23 9 q —I— Mean Detection Time for
_E 8 4 Full-Link Blockage in Link 34
g 1 (Downstream Station)
0 71
E
; G-j *1/-
5 5i i 1 H—H—-
1'3 4~ —— / /
2 d
8 3‘
C 2 ..
(U
é’ 1-
o l 1 I ' 1 # T ' I ' I

 

 

 

V I I V l I I I I I T I
1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0 3.1
Lim it (sigma) Changes

Figure 5.17 Mean Detection Time (minutes) for Full-Link Blockage in Link 34
with Bars of One Standard Deviation (Downstream Station)

 

 

 

 

 

 

 

 

 

 

’UT _—
9 - n-

‘3 i —-— Mean Detection Time for

.E 8 ‘ Full-Link Blockage in Link 34 —— T

E 71 (Both Stations) /‘

a, . /#

.E 6 -

..- d

r: 5 - _—

.9 - / -— ~— ~-

3 w H/H 1 1 #—

33 3 —

o .

5 2 1

m .

2 1 ‘
0 I l v 1 T 1 I ' I ' '

 

 

I I I l I ‘ I l V l I U
1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0 3.1
Limit (sigma) Changes

Figure 5.18 Mean Detection Time (minutes) for Full-Link Blockage in Link 34
with Bars of One Standard Deviation (Both Stations)

115

5.8.4 Determination of Recommended Intervals
As seen in Tables 5.13-5.17 and Figures 5.5-5.18, the false alarm rate has an inverse

relationship with the detection rate as an increases. Generally the mean detection time
also increases as oik increases. These changes in the false alarm rate, the detection rate,

and the mean detection time can be expressed by the change in the slope of the
relationship. These slope changes are shown in Table 5.18-5.20. The detection rate and
the mean detection time were investigated only for link 34 and 45. Therefore, the
relationship between false alarms, detection rate and corresponding mean detection time
does not necessarily precisely represent the relationship for all links in a network.

The recommended a“ values were found by analyzing the results for both the

network as a whole, and for the high traffic direction. These values are found by
analyzing the false alarm rate and the mean detection time for intervals where the
detection rate is 100 percent for both one-lane blockages and full-link blockages.

For the network as a whole, a 100 percent detection rate was found at 2.0 and 2.1 a“

for the one-lane blockage and full-link blockage in link 34, respectively. The
corresponding number of false alarms are 7.2 and 4.6, respectively, if the system were
operated for 2 hours a day. The mean detection time for the one-lane blockages is 4.80
and 4.93 minutes at these respective intervals. The mean detection time for a full-link
blockage is identical (3.78 minutes) at these respective intervals. As can be seen in Table

5.18-5.20, the slope change for false alarms is the largest, and the slope change for mean

116

detection time is the smallest at 2.1 a“ for the one-lane blockage. The slope change of

the mean detection time for full-link blockages is not changed.
For the high traffic direction (from east to west), the detection rate is 100 percent for

values up to 2.1 a“ . The corresponding false alarms in this direction showed no slope
changes from 2.0 to 2.1 aik . The corresponding slope of the mean detection time for a
one-lane blockage is 0.13 (from 4.80 minutes to 4.93 minutes) at 2.1 032k . The full-link

blockage again show no change. From these results, it is recommended that a value of 2.0

0“ defines the best interval, because there are no changes in false alarms and the mean

detection time and the gain in the mean detection time is greatest in this interval for 2.0
0'“ to 2.1 0'“.

These results are summarized in Table 5.21. It should be noted that if the cycle length
were less than 120 seconds, then the mean detection time would be lower. For example,
the 4.93 minutes to detect a one-lane blockage in link 34 would be reduced to 2.47

minutes if the cycle length were 60 seconds.

117

Table 5.18 Change in False Alarms over Different Intervals

 

 

 

 

 

 

 

Interval Changes
Unit
2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0
FA/21-lr/ 7.2 4.6 4.2 3.0 2.4 1.8 1.6 1.4 1 0.8 0.6
Network Network
SC (—)* 0.0 2.6 0.4 1.2 0.6 0.6 0.2 0.2 0.4 0.2 0.2
High FA/2Hr/ 0.8 0.8 0.6 0.4 0.2 0.2 0.2 0 0 0 0
Traffic Direction
Volume SC (-)* 0.0 0.0 0.2 0.2 0.2 0.0 0.0 0.2 0.0 0.0 0.0
Direction

 

 

 

 

 

 

 

 

 

 

 

 

 

*Slope Change: (-) indicates a negative slope change and (+) a positive slope change

Table 5.19 Change in DR and MDT for One-Lane Blockages over Different

Intervals (Both Stations)

 

Interval Changes

 

 

 

 

 

 

Unit
2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0
DR 100 100 96.7 86.7 83.3 83.3 76.7 70.0 60.0 50.0 46.7
SC(-) 0.0 0.0 3.3 10.0 3.4 0.0 6.6 6.7 10.0 10.0 3.3
MDT 4.8 4.93 5.31 5.77 5.76 6.0 6.35 6.85 8.11 9.47 9.87
SC(+) 0.0 0.13 0.38 0.46 -0.01 0.24 0.35 0.5 1.26 1.36 0.40

 

 

 

 

 

 

 

 

 

 

 

 

 

 

118

Table 5.20 Change in DR and MDT for Full-Link Blockages over Different Intervals

 

 

 

 

 

 

 

 

 

 

 

 

 

(Both Stations)
Interval Changes

Unit

2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0
DR 100 100 100 100 100 100 100 100 100 100 100
SC(-) 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
MDT 3.78 3.78 4.00 4.22 4.44 4.44 4.89 5.56 6.22 6.67 7.11
SC(+) 0.00 0.00 0.22 0.22 0.22 0.00 0.45 0.67 0.66 0.45 0.44

 

 

 

 

 

 

 

Table 5.21 Results of Incident Detection Performance at Recommended Intervals

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Network High Traffic
(2.1 a“) Direction (2.0 a“)
F A/ 10 Hr/Network 23 4
False Alarm FA/Hr/Station 0.096 0.033
FAR/Hr/Station (%) 0.319 0.111
One-Lane DR (%) 100 100
Link 34 Blockage MDT (minute) 4.93 (136*) 4.8 (124*)
Full-Link DR (%) 100 100
Blockage MDT (minth) 3.78 (0.67*) 3.78 (0.67*)
*Standard Deviation

 

119

5.9 Parameter Estimation for Different Conditions

In this section, the state-space model for different geometric and traffic conditions are
varied. The values of (1) for these different conditions are plotted to see how they change
over time and under different conditions.

Test locations were selected based on geometry and traffic volumes. F our links in the
network were selected for the test. They were link 34, link 45, link 89, and link 910. Link
34 represents high traffic and long distance, while link 45 is high traffic and short
distance. Link 89 is low traffic and long distance, and link 910 is low traffic and short

distance. The characteristics of these links are shown in Table 5.22.

Table 5.22 Characteristics of Selected Links

 

 

 

 

 

Selected Links Link Length (feet) Entering Traffic Volume for the Link
(vph)
34 2100 1870
45 530 1696
89 2900 890
910 925 940

 

 

 

 

 

Simulations were run for 20 days, with one hour of simulation representing one day.
The first day was used for initialization, with the remaining 19 days were used to obtain
parameter stabilization.

Plots of these parameters for various conditions are shown in Figure 5.19. Phi (1,1)

represents the error covariance of the flow, Phi (2,2) the error covariance of the percent

120

occupancy, Phi (3,3) the error covariance of the average speed in lane 1 while Phi (4,4),
Phi (5,5), and Phi (6,6) represent the error covariance of the flow, percent occupancy, and
average speed in lane 2, respectively. In these figures, the parameters fluctuate for the
first part of the time period, and then become stable. These fluctuations are different
under different conditions. For example, CD (1,1) at link 34 begins to concentrate at about
one-fourth of the time period to around 0.27 while (I)(l,l) at link 89 begins to
concentrate about one—third of the time period to around 0.1. Other (I) ’3 show different

patterns under different conditions as well.

2355.80 955,—. can 2.50830 393$: 5 nEotam E: .5883!— 36 95mg

3 a: a smudge Ea 3

A930 9. .2 ..2.< comm .. 00”: $87..
omv own omm cor

b I P i- . o.o

com
L

b

 

 

121

 

$11-

 

I "l‘l‘\|’\\l.“\"
'

 

...0
N6
m6
v.0
m.o
m.o

5.0

I w.o

:1; _In_

 

 

 

5.8__._n_|...|$.m:1n_
33:5 ......... $.32}. ...... :.:_In_lll

 

 

 

I m6

 

 

 

 

o._.

178 >|U!'| JO} S.IHd

122

8.253 36 as»;

a. as a $9.20 Ea 3V

E30 2 .2 .s_.< comm - 8”: we:
own omv own emu oov

 

 

 

b i- . Hi 0.0

 

0.0

i m.o

lmd

 

0.?

 

917 >1U1'l J015.1Hd

123

Gags an 95E..—

a ca: 3 £56 Ea 3

A930 9. .8 .S..< comm - cont mEC.

 

 

com 00? com CON on:

n n p n p - P — p b pa 0 o
.. ..... , is

. 1 1r I'll-51...}... 4.. ... or M. «.0
.) if .

i.» \st N .

_ N o
u
_

. m.o

 

I m6

 

 

o...

 

68 nun Jo; S.|Hd

124

8.25: an 2:5

o; E. a $55 Ea A3

Ema 3 .8 .52 85 - 8”: we:

oov com
b

oom
.

cor

 

 

 

 

 

If

.14.} ..J a
I ’J. a";’— . .4

    
 

 

 

 

016 >1U11 101 s.lHd

125

5.10 Summary of TRAF-NETSIM Simulation Results

Simulations using the TRAP-NETSIM program were made to generate incident-free
and incident data in order to evaluate the incident detection algorithm. The network
simulated was Washtenaw Avenue located in Ann Arbor, Michigan. The traffic variables
obtained from the simulation for the algorithm were traffic counts, percent occupancy,

and average speed in each lane.

5.10.1 A Test at 95 Prediction Intervals (2.0 0'“ )

Two tests were conducted: an incident-free test, and an incident test.

Incident-free data sets were generated for a false alarm check. One-hour of simulation
representing a weekday peak hour were conducted for generating the normal A.M. peak
hour condition. A total of ten-days of incident-free data were generated. The raw data
from detectors were smoothed using a moving average with 3 time intervals. The records
were then aggregated in 2 minute time slice for analysis. Incidents were indicated by
checking the upper and the lower intervals constructed with 2 standard deviations of the
predicted error covariance.

Results of this test show a total of 7.2 false alarms for the network if the system were
operated 2 hours per weekday. This converts to 0.15 false alarms per hour per station. For
the high volume traffic direction (from east to west), a total of 0.8 false alarms occurred
when the system was simulated for 2 hours a weekday. This is 0.033 false alarms per hour

per station.

126

Incident data sets were also generated to test the performance of the algorithm. Link
34 and 45 were selected to represent different geometric characteristics under high
volume traffic conditions (i.e., long distance and short distance, respectively). A total of
65 incidents were generated. For link 34, a total of 39 incidents were simulated at
different locations. For link 45, a total of 26 incidents were simulated.

By considering the upstream and the downstream detector stations together, the
detection rate of one-lane blockages was 100 percent for both link 34 and link 45. The
mean detection time for each link is 4.8 minutes and 2.7 minutes, respectively. The
detection rate of full-link blockages showed 100 percent for each link, with a mean

detection time of 3.78 and 2.33 minutes, respectively.

5.10.2 A Sensitivity Test for Various Prediction intervals

Various prediction intervals from 2.0 to 3.0 a“ were applied to test a trade-off

between false alarms and detection rate on link 34. The corresponding mean detection
time was also investigated. As expected the false alarms decreased as the intervals
became wider and the detection rate was decreased. The mean detection time shows a
consistent increase as the interval became wider.

For the network, if the system were operated for 2 hours a day, there would be
reasonable number of false alarms (4.6), perfect detection rate, and reasonable mean

detection time (4.93, and 3.78 minutes, respectively) using a 2.10“ . Wider prediction

intervals result in the algorithm failing to detect incidents and long mean detection time.

127

For the high volume traffic direction, there is less than one false alarm per day using a

2.0 a“ . Detection rate at this interval was 100 percent with a mean detection time of 4.8

minutes and 3.78 minutes for one-lane blockages and full-link blockages. Therefore,

2.0 a“ prediction intervals are appropriate for the high volume traffic direction.

5.10.3 Parameter Estimation for Different Conditions

The parameter (CD) in the state-space model was investigated under different
geometric and traffic volume conditions. Links 34, link 45, link 89, and link 910 were
selected to represent high traffic and long distance, high traflic and short distance, low
traffic and long distance, and low traffic and short distance, respectively. Plots of (D on
these different links show different patterns. These parameters concentrate on and change
as new measurements from different geometry and temporal traffic flow are obtained.

Conventional incident detection algorithms, such as the California algorithm, the
Bayesian approach, the Exponential Smoothing algorithm, and the Box-Jenkins
Technique-based algorithm are not advantageous in this aspect because they utilize fixed
parameters. However, an adaptive model-based algorithm as used in this study is
advantageous by utilizing flexible parameters to represent different spatial and temporal

conditions, reducing the disadvantage of the fixed-parameter model.

Chapter 6

CONCLUSIONS AND DISCUSSION

6.1 Conclusions
In this section, the findings of this research are summarized. Also, the comparison

with previous incident detection algorithms for arterial streets is discussed.

6.1.1 Summary of the Research
This section briefly summarizes the research results.

1. In the literature reviewed, the theoretical background and limitation of various
freeway and arterial street algorithms were examined. No algorithm could be defined
as the best for all applications since no direct comparisons were available for all the
different conditions. Commonly found issues were: data dependency, transferability,
adaptability, and availability.

2. An incident detection algorithm was developed based on the Kalman filtering
algorithm. Some modifications were made to the algorithm to assist in defining an
incident. The common cycle length over a controlled network was determined to be
the best data sampling time interval. This aggregation reduces data fluctuations due to

different phases of a signal. Also, it was determined that point information from a

129

single detector at an intersection approach was essential since there are many forces

that disturb the traffic flow between successive detectors on an urban arterial street.

. A state-space model and a measurement model were constructed. Traffic flow

dynamics on an urban arterial street was represented by a transition matrix. The time-
dependent noisy term in the state-space model was also determined. An incident was
identified by using the upper and lower intervals obtained from the predicted error
covariance, compared with filtered estimates of the traffic control variables.
. The incident detection algorithm was applied to field incident-free and incident data
sets. The purpose of the test was to fit the obtained data into the state-space model
and estimate states with the noisy flow and speed data. The field data were collected
by using experimental detectors on Hagadom Road in East Lansing, Michigan. Raw
data (traffic flow and speed) were then aggregated in 3 minute time slices and
smoothed by a moving average over 3 time intervals to eliminate impulsive noises.
Field data which included an incident were obtained from Harrison Road in East
Lansing, Michigan. Raw data were aggregated and smoothed in the same way as the
Hagadom Road data. Results showed good tracking of predicted and filtered
estimates in the incident-free data, as there were no filtered estimates beyond the 95
percent prediction intervals. The corresponding scatter plots for residuals of traffic
flow and average speed showed good randomness around zero.

Results from the incident-data showed the potential for detecting abnormality in
arterial street traffic flow. The plots of these variables versus time (Figure 4.7) show
how traffic flow and speed change when there is an incident. These figures showed

good potential in detecting incidents.

130
5. The incident detection algorithm was also tested with simulated data. The TRAF-

NETSIM, Version 5 was used to simulate incident-free and incident data sets.
Washtenaw Avenue located in Ann Arbor, Michigan was selected as the network to
represent a real-world situation. The arterial street was optimized using a 120 second
signal cycle length. This signal cycle length was also used as the data sampling time
interval. Raw data (traffic flow, percent occupancy, and average speed) for the A.M.
peak hour was obtained and smoothed by the moving average of three orders.

Ten hours (one hour represents one-day) of incident-free data was simulated,
resulting in 300 records at each detector station. A total of 7,200 records of data were
generated for the 24 detector stations in the network. To generate incident data, two
links were selected to represent various geometry (a long and short link length) with
high traffic volume. Various incidents were generated. They included partial
blockages and full-link blockages at various locations with various duration. As a
result, a total of 65 incidents were generated. Both the incident-free and incident data

were tested using the 95 percent prediction interval (i.e., 2 an ).

Results for the incident-free data showed that a total of 36 false alarms over 10
hours were indicated in the network. This is 0.15 false alarms per hour per station.
The corresponding false alarm rate is 0.5 percent per hour per station for the network.
For the direction of the high volume traffic (e.g., from east to west), a total of 4 false
alarms were indicated over 10 hours. This is 0.033 false alarms per hour per station.
The corresponding false alarm rate is 0.111 percent per hour per station. As expected,
the incident detection algorithm showed better performance in the high traffic flow

direction than light traffic flow direction.

131
Results for the incident data showed a detection rate of 100 percent at both the

upstream and the downstream stations for one-lane blockages and full-link blockages
in the two links tested, link 34 and link 45. The mean detection time for a one-lane
blockage was 4.8 minutes with a standard deviation of 1.24 for both stations on link
34. On link 45, the mean detection time was 2.7 minutes with a standard deviation of
1.17. For full-blockages, the mean detection time was 3.78 minutes with a standard
deviation of 0.67 for link 34 and 2.33 minutes with a standard deviation of 0.82 for
link 45. The analysis showed that the downstream detector station had a lower mean
detection time.

A sensitivity test for prediction intervals was conducted to search for a
recommended criterion. For the network, a reasonable false alarm rate (0.319 percent)
with a perfect detection rate (100 percent for one-lane blockages and full-link
blockages) and a reasonable mean detection time (4.93 minutes with a standard
deviation of 1.24 for one-lane blockages and 3.78 minutes with a standard deviation

of 0.67 for full-link blockages, respectively) were found when using a 2.1 0's]. . For
the high traffic volume direction, the algorithm showed good performance at 2.0 0's]. .

That is, 0.111 percent false alarm rate, 100 percent detection rate for one-lane
blockages and full-blockages and 4.8 minutes for the mean detection time with a
standard deviation of 1.24 for one-lane blockages and 3.78 minutes for the mean

detection time with a standard deviation of 0.67 for full-link blockages, respectively.

132

6.1.2 Comparison with Previous Studies

This section discusses the performance results of existing incident detection
algorithms on arterial streets. Their test conditions are also mentioned. Then, the
performance of these algorithms are compared with the algorithm developed in this
research. It should be noted that the comparison of the algorithms do not represent
absolute superiority and inferiority, because the condition under which these algorithms
have been tested is different.

The incident detection system developed by Chen (1994) used a 750 foot long three
lane link to test the algorithm. This link contained two detector stations located at both
the upstream and downstream ends. The cycle length was 60 seconds. TRAP-NETSIM
was used to generate incidents. A total 72 observations were generated for normal traffic
conditions, and 60 cases were simulated for incidents with one-lane blockages and two-
lane blockages on the link. These cases were evenly divided for peak and off-peak periods
(18 cases for one-lane closures and 12 cases for two-lane closures). Incidents were
located in the center of the link. The three control variables were used: downstream
speed, flow and occupancy. The time unit for data aggregation was 30 seconds. The
results of the incident detection system were a 96.7 percent detection rate, a 12.8 percent
false alarm rate, and a 240.0 seconds for the mean detection time when traffic flow is less
than 900 vehicles per hour per lane. For traffic flow greater than or equal to 900 vehicle
per hour per lane, the algorithm resulted in a 100 percent detection rate, a 13.1 percent
false alarm rate, and 60 seconds of mean detection time.

The detection rate and the mean detection time in Chen’s study are good in the

congested condition. Compared to the results in this research, the detection rate is the

133
same and the mean detection time is better (60 seconds). For the algorithm developed in

this research, a simple mean of the mean detection time would be 3.4 minutes for all the

incidents in link 34 and link 45 using a 2.0 a“ for A.M. peak hours. However, the data

aggregation methodology is different, and the detector configuration in Chen’s study is
not practical due to the requirement for multiple detectors in a link. The false alarm rate
in Chen’s study is much higher (13.1 percent) than that of this algorithm (0.111 percent
per hour per station for the high traffic volume direction). If the incident detection
algorithm is to be used to trigger a response, a false alarm rate of over 13 percent would
be unacceptable.

Martinez et al. (1994) conducted field trials of an Urban AID system based on
computer vision techniques. Field tests included a field trial that contained one complex
junction, and field trials that covered four adjacent large junctions and the most important
adjacent links. Eight cameras were installed at these locations. A total of 951 incidents
were logged in the period from October 93 to March 94. Their duration was less than one
minutes for 86 percent of the cases. These incidents were stopped vehicles in one lane
disturbing traffic in the same lane, or stopped vehicles in two lanes at the same position,
or a stopped vehicle in one lane disturbing the adjacent lane in an entry zone near the
junction. Overall detection rate was 92.74 percent with the false alarm rate 24.23 percent.
No mean detection time was discussed.

The Urban AID system shows potential for detecting incidents on signalized streets by
using different sensor systems. However, the high false alarm rate compared with the
algorithm in this research suggested improvement is required in this technology. Also, the

mean detection time should be discussed.

134

Ivan et al. (1995) used the data fusion approach in detecting an incident on urban
arterial streets. The data fusion approach obtains data, such as traffic volume, occupancy,
and speed from fixed detectors, and observed travel times collected from probe vehicles
traveling the street network. These observed data are processed in uniquely developed
algorithms for each link to determine the likelihood of an incident. Then the Algorithm
Output Fusion Network combines these algorithm outputs using neural network logic. By
using INTRAS, a micrOSCOpic freeway corridor traffic simulation model, more than 100
simulations were run to generate the training data for a variety of incident and
corresponding non-incident conditions. The simulation network was a 5-km section of
major arterial streets that contained 8 intersections and 39 loop detectors. A data
aggregation interval of 7 minutes (all signals had cycle length of 140 seconds). Incidents
were simulated on six different links, at three or four locations on each link, for duration
of from 5 to 10 aggregation intervals. These are major incidents of 35 to 70 minute
duration. Ten percent of the training data files were selected randomly and set aside for
testing the algorithm. The Algorithm Output Fusion Network (the Two Input Network)
detected all of the incidents in the testing data with no false alarms. The algorithm
misclassified 0.11 percent of the non-incident observations in the training data set.

An additional study by Ivan (1996) with network alternatives was conducted. The
alternatives included an Output Memory Network and a Full Network which added input
from previous time periods to the current input. These alternatives showed better
performance. The Output Memory Network showed a detection rate of 86.11 percent in
the training data and 100 percent in the test data with a 0 percent false alarm rate in both

data sets. The Full Network showed a 92.59 percent detection rate in the training data and

135

100 percent in the test data, and a 0 percent false alarm rate in both data sets, respectively.
The mean detection time was not discussed in both studies.

This approach showed good detection and a better false alarm rate (0 percent) in the
test data compared to the algorithm in this research. However, the performance should
also be evaluated by the mean detection time. The data aggregation time interval of 5 or 7
minutes from fixed detectors is too long to allow an incident detection system to be
useful. Also, this algorithm is data intensive. That is, this test used 90 percent of total data
as training data to produce the 100 percent detection rate in the other 10 percent of the
test data. In comparison, the algorithm in this research used only three hours of data to

initialize and stabilize the parameters, resulting in less data dependency.

6.1.3 Conclusions of the Research

The incident detection algorithm in this research shows good potential for use on
signalized arterial streets. With less data dependency, this real-time algorithm will be
simple and efficient to apply to existing arterial conditions using raw data from existing
sensors such as loop detectors. This algorithm does not require special equipment and can
be applied to existing arterial conditions, although it is still in the experimental stage. The
comparison with other incident detection systems on arterial streets shows good detection

rate, good false alarm rate, and reasonable mean detection time for this algorithm.

6.2 Discussion on Further Research
In this section some factors that would improve this incident detection algorithm are

discussed. Also, further arterial street incident detection research is discussed.

136

Limitation of a Simulation

Although the incident detection algorithm in this research shows potential, it is based
on simulation. A simulation may be a good alternative when the real field data is not
available. However, there are always limitations in the ability of simulation to represent a
real world situation. For example, driver’s response to a stalled vehicle in a lane would be
varying individually. Some drivers may change their route to avoid the incident by taking
alternative roadways or some may still drive in that roadway by changing lanes. The rate
of diverting drivers may affect the performance of an incident detection system.

One of the variables in the simulation is the “seed number” to generate different traffic
patterns. For the same incident conditions, a different seed number will generate a
different traffic pattern for that incident scenario. This may result in various incident

detection results.

An Incident Detection Treatment after an Incident

In this study, the incident detection algorithm was developed mainly to identify the
beginning of an incident for efficient incident management. The pattern of traffic control
variables during an incident state are unstable and fluctuate due to the stop and go status.
Using these measurements to estimate the flow after termination of the incident (i.e.,
normal traffic flow) may increase the chance of a false alarm shortly after the incident
termination. Therefore, it is probably advantageous to use measurements from a time
period prior to the incident to eliminate false alarm chances after the incident termination.

This would be implemented by storing the prior measurements and corresponding

137

parameter values in memory, and using them after the incident termination as the new

process begins.

Detector Location, the Number of Detectors, and New Detectors

It is well known that the location of detectors is very important to an incident detection
algorithm, especially on long links. Existing detector location is usually at the
downstream end of a link, resulting in full occupancy during the red phase of a signal.
Alternative detector locations are at the upstream end of a link as used in Europe, or at the
middle of a link.

The best location for any link could be found by simulating various incidents against
various detector locations. That is, the relationship among mean detection time, detection
rate, and false alarm rate could be found from the simulated data. The best locations
would be dependent on various traffic conditions such as A.M. peak hour, PM. peak
hour, and off-peak hour volumes.

The more detectors available the better the performance of an incident detection
system. Some of incident detection systems on arterial streets, such as Chen’s (1994) and
Ivan’s (1995), obtain traffic measurements from multiple detectors and various sources.
Using information from multiple sensors enhances an incident detection system. If the
cost is justified by the benefits, it is worth considering this approach.

Martinez et al. (1994), and by Parkany, and Bernstein (1993) have demonstrated that new
technology-based sensors may improve an incident detection system. A video-image
process or a VRC (or “tags”) technology can provides a detailed information on the traffic

state or an individual vehicle that can not be obtained from conventional detectors. This

138

type of information suggests different approaches to an incident detection system. To use

a new-technology based sensor, the cost and benefit should be justified as well.

Efficient On-Line Implementation

The algorithm in this research was tested with field collected and simulated data. In
implementing an on-line detection system for an urban network, it will be necessary to
construct a computing structure to run the algorithm in multitasking because the
algorithm is run for all-detector stations at the same time. A suitable computing system

well require a fast and efficient computer.

Emergency Hot-Lines

Nothing is more accurate and faster than a driver who witnesses an incident occur
describing the incident. Many drivers now use a cellular phone or a car phone. Reports
from drivers may soon be quicker than an incident detection system. Further research

should include the possible use of this alternative.

APPENDICES

APPENDIX A

APPENDIX A

DESCRIPTION OF TRAFFIC VOLUME FOR EACH NODE IN THE NETWORK

Table A Descriptions of Traffic Volume for Each Node in the Network

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

East Bound West Bound North Bound South Bound
Node

No. T R L T R L T R L T R L
2 576 33* 152 920 37* 22 36 203* 85 143 235 126
3 664 31 1 301 1 128 62* 50 224 84 392 90 350* 60
4 1031 22* 39 1 664 1 12 94 1 1 96 N/A 10 32* N/A
5 897 28* 10* 1628 10* 68 18 20* 99* 10 12* 23
6 750 40* 1 57 1 165 1 74* 226 303 89* 1 81 253 1 26* 79
7 975 706 35 1 284 1 0* 44 62 1 07* 32 28 24* 47*
8 31 1 1 0 N/A 890 N/A 594* N/A 617 N/A N/A N/A N/A
9 382 33* N/A 863 N/A 27 N/A 35* 77* N/A N/A N/A
10 391 12* 17* 912 19* 10* 21 10* 15* 14 28* 11*
11 340 30* 11* 926 19* 10* 70 57* 71* 108* 29* 12
12 304 20* 10* 842 10* 174* 16 51* 18* 34 35* 13*

 

T: Through Traffic

R: Right Turn Traffic

L: Left Turn Traffic
*1 Shared Lane Traffic with Through Traffic

139

 

APPENDIX B

140
APPENDIX B

DESCRIPTION OF GEOMETRIC CONFIGURATIONS FOR NETWORK LINKS

Table B Description of Geometric Configurations for Network Links

 

 

 

 

 

 

 

 

 

 

 

 

 

 

East to West West to East
Link Length No. of Length of Length of Length of Length of
(ft) Through Left Turn Right Turn Left Turn Right Turn
Lanes* Bay (fi) Bay (ft) Bay (ft) Bay (it)

1 ,2 3000 2 800 N/A N/A N/A
2,3 4000 2 800 N/A 800 N/A
3,4 2100 2 200 200 200 200
4,5 530 2 200 200 200 N/A
5,6 1450 2 600 N/A N/A N/A
6,7 3960 3 400 N/A 600 N/A
7,8 400 3 200 N/A 150 150
8,9 2900 2 N/A N/A N/A 400
9,10 926 2 N/A N/A N/A N/A
1 0,1 1 2776 2 N/A N/A N/A N/A
1 1 ,12 1200 2 N/A N/A N/A N/A

 

 

 

 

 

 

 

 

*The number of turning bay lanes is one.

APPENDIX C

141

APPENDIX C

DESCRIPTION OF SURFACE STREET TURNING MOVEMENTS

Table C Description of Surface Street Turning Movements

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Upstream Node Downstream % of Traffic % of Traffic % of Traffic
Node Turning Left Going Though Turning Right
1 2 3 92 5
2 1 0 100 0
32 3 4 91 5
32 2 20 76 4
43 4 5 89 6
43 3 26 59 15
54 5 4 96 0
54 4 4 94 2
65 6 13 67 10
65 5 0 97 3
76 7 3 97 0
76 6 17 79 4
87 8 40 60 O
87 7 1 95 4
98 9 3 97 0
98 8 0 97 3
509 10 1 97 2
509 9 0 92 8
510 11 1 97 2
510 10 4 93 3
521 12 17 82 1
521 11 3 89 8

 

 

 

 

 

 

142
Table C (cont’d)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Upstream Node Downstream % of Traffic % of Traffic % of Traffic
Node Turning Lefi Going Though Turning Right

12 13 0 100 0

13 12 3 91 6

52 2 25 28 47
22 2 26 11 63
53 3 12 18 70
23 3 70 18 12
14 4 0 10 90
24 4 0 10 90
15 5 62 14 24
25 5 72 13 15
16 6 17 55 28
26 6 31 53 16
17 7 48 28 24
27 7 16 31 53
28 8 0 0 100
29 9 69 0 31

110 10 21 26 53
30 10 33 46 21

111 11 8 72 20
31 11 36 36 28
112 12 16 41 46
62 12 21 19 60

 

 

 

 

 

 

APPENDIX D

APPENDIX D

143

DESCRIPTION OF PRETIMED SIGNAL INTERVAL FOR EACH NODE

Table D Description of Pretimed Signal Interval for Each Node

 

 

 

 

 

 

 

 

 

 

 

 

 

Node Offset Interval Interval Interval Interval Interval Interval Interval Interval

1 2 3 4 5 6 7 8

2 119 17 4 44 4 24 4 19 4

3 49 21 4 41 4 26 4 16 4

4 110 97 4 1 13 4 1

5 108 99 4 1 11 5

6 60 20 4 50 4 16 4 18 4

7 4 91 4 1 19 5

8 7 92 4 20 4

9 36 16 4 2 94 4

10 53 12 4 2 98 4

11 108 19 4 93 4

12 14 13 4 99 4

 

 

 

 

 

 

 

 

 

 

 

LIST OF REFERENCES

LIST OF REFERENCES

Abdulhai, Baher, and S. G. Ritchie. 1995 Performance of Artificial Neural Networks for
Incident Detection in ITS. Transportation Congress, Volume 1, Civil Engineers-Key to

the World’s Infrastructure, Proceedings of the Conference, American Society of Civil
Engineers, New York. pp. 227-238.

Ahmed, Samir A. 1983. Stochastic Processes in Freeway Traffic, 1. Robust Prediction
Models, Traffic Engineering & Control, June/July, pp. 306-310, 314.

Ahmed, Mohamed S., and Allen R. Cook. 1979. Analysis of Freeway Traffic Time-Series
Data by Using Box-Jenkins Techniques. Transportation Research Record 722, TRB,
National Research Council, Washington, D. C., pp. 1-9.

Ahmed, Samir. A., and Allen R. Cook. 1980. Time Series Models for Freeway Incident.
Transportation Engineering Journal of ASCE , ASCE, November, pp. 731-745.

Ahmed, Samir A., and Allen R. Cook. 1982. Discrete Dynamic Models for Freeway
Incident Detection Systems. Transportation Planning and Technology, Vol. 7. Pp. 231-
242.

Al-Sahili, Khaled A., and William C. Taylor. 1996. Bus Preemption Signal (BPS) - An
Application of an Advanced Public Transportation System (APTS), Michigan State
Universifl, East Lansing.

Balke, Kevin N. 1993. An Evaluation of Existing Incident Detection Algorithms.
Research Report 1232-20. Texas Transportation Institute. College Station, TX.
November.

Barbaresso, James C. 1993. Overview of the FAST-TRAC IV HS Program. Pacific Rim
TransTech Conference, Proceedings Vol. 1, American Society of Civil Engineers, New
York, July 25-28, pp. 134-137.

Bell, M. G. H., Dr., and B. Thancanamootoo, Dr. 1988. Automatic Incident Detection

within Urban Traffic Control Systems. Roads and Traffic 2000, Vol. 4/2, Proceedings,
Berlin 6-9, page 35-39.

144

145

Bhandari, Nikhil, Frank S. Koppelman, Joseph L. Schofer, Vaneet Sethi, and John N.
Ivan. 1995. Arterial Incident Detection Integrating Data from Multiple Sources.
Transportation Research Record 1510, Washington, D. C., pp 60-69.

Bretherton, R. D., and G. T. Bowen. 1991. Incident Detection and Traffic Monitoring in
Urban Areas. Drive Conference, Brussels, Belgium, Advanced T elematics in Road
Transport, Vol. I., pp 740-751.

Chang, Edmond Chin-Ping. 1992. A Neural Network Approach to Freeway Incident
Detection. Vehicle Navigation & Information Systems, Oslo. Norway. September 2-4.
pp. 641-647.

Chang, Edmond Chin—Ping and Kunhuang Huamg. 1993. Incident Detection Using
Advanced Technologies. Preprint 930943, The 72nd Annual Meeting Transportation
Research Board, Washington, D. C. January.

Chassiakos, Athanasios P. 1992. Spatial-Temporal Filtering and Correlation Techniques
for Detection of Incidents and Traffic Disturbances. Doctoral Dissertation, December,
University of Minnesota,

Chassiakos, Athanasios P., and Yorgos J. Stephanedes. 1993. Smoothing Algorithms for
Incident Detection. Transportation Research Record 1394, TRB, National Research
Council, Washington, D. C., pp. 8-16.

Chen, Chao-Hua. 1994. A Dynamic Real-Time Incident Detection System for Urban
Arterials. Doctoral Dissertation, University of Maryland. College Park. Maryland.

Chen, Chao-Hua, and Gang-Leng Chang. 1992. A Review of Recent Freeway Incident
Detection Algorithms. Report No. DTFH61 -92-R-00122, Department of Civil
Engineering, University of Maryland,June 5, College Park, MD 20742.

. 1993. A Dynamic Real-Time Incident Detection System for Urban Arterials-
System Architecture and Preliminary Results. Proceedings of Pacific Rim Trans Tech
Conference, Vol. 1, American Society of Civil Engineers, July 25-28, Seattle,
Washington, pp. 98-104.

Cheu, R. L., S. G. Ritchie, and W. W. Recker. 1991. Investigation of a Neural Network
Model for Freeway Incident Detection. The Second International Conference on the
Application of Artificial Intelligence Techniques to Civil and Structural Engineering,
Oxford. England, September.

Cohen, S. 1994. Comparative Assessment of Conventional and New Incident Detection
Algorithms. Road Traffic Monitoring and Control. Conference, Publication No. 391,
April 26-28, pp. 156-159.

146

Cohen, S, and Z. Ketselidou. Unknown Date. DAISI: A Calibration Tool for Automatic
Incident Detection Algorithms. Application on the Paris Ringway. Unknown Source.

Collins, J. F. 1977. Incident Detection Experiments on the Boulevard Peripherique in
Paris. TRRL Supplementary Report 362.

Collins, J. F. 1983. Automatic Incident -Experience with TRRL Algorithm-HIOCC.
TRRL Supplementary Report 775.

Collins, J. F., C. M. Hopkins, and J. A. Martin. 1979. Automatic Incident Detection-
TRRL Algorithm HIOCC and PATREG. TRRL Supplementqrv Report 526.

Cook, Allen R., and Donald E. Cleveland. 1974. Detection of Freeway Capacity-
Reducing Incidents by Traffic-Stream Measurements. Highway Research Record 495,
Highway Research Board, National Research Council, Washington, D. C., pp. 1-1 1.

Dailey, Daniel J. 1993. Improved Error Detection for Inductive Loop Sensors. Report
Number WA -RD 300.1, Washington State Transportation Center. May.

Dudek, Conrad L., and Carroll J. Messer. 1973. Detecting Stoppage Waves for Freeway
Control. Highway Research Record 469, Highway Research Board, National Research
Council, Washington, D. C., pp. 1-15.

Dudek, Conrad L., Carroll J. Messer, and Nelson B. Nuckles. 1974. Incident Detection on
Urban Freeways. Highway Research Record 495, Highway Research Board, National
Research Council, Washington, D. C., pp. 12-24.

Dudek, Conrad L., and Gerald L. Ullman. 1992. Freeway Corridor Management.

National Cooperative High way Research Program Synthesis of High way Practice 1 77,
March, Washington, D. C.

Federal Highway Administration. 1995. T RAF User Reference Guide, Version 5. March.

Gall, Ana 1., and Fred L. Hall. 1989. Distinguishing between Incident Congestion and
Recurrent Congestion: A Proposed Logic. Transportation Research Record 1232, TRB,
National Research Council, Washington, D. C., pp. 1-8.

Giuliano, Geneview. Incident Characteristics, Frequency, and Duration on a High
Volume Urban Freeway. Transportation Research, Part A, Vol. 23A, No. 5, 1989, pp.
387-396.

Goolsby, M. E. 1971. Influence of Incidents on Freeway Quality of Service. Highway
Research Record 349, pp. 41-46.

I47

Grewal, Mohinder S., and Angus P. Andrews. 1993. Kalman Filtering Theory and
Practice. Prentice Hall, New Jersey

Hall, Fred L. 1987. An Interpretation of Speed-Flow-Concentration Relationships using
Catastrophe Theory. Transportation Research, Part A, Vol. 21A, No. 3, pp. 191-201.

Hall, F. L., Brian L. Allen, and Margot A. Gunter. 1986. Empirical Analysis of Freeway
F low-Density Relationships. Transportation Research, Part A, Vol. 20A, No. 3, pp. 197-
210.

Hall, F. L., Yong Shi, and Gerrge Atala. 1993. On-Line Testing of the McMaster Incident
Detection Algorithm under Recurrent Congestion. Transportation Research Record
1394, TRB, National Research Council, Washington, D. C., pp. 1-7.

Hoose, Neil. 1994. Incident Detection and Traffic Surveillance using the IMPACTS
Video Analysis System. Traffic Technology International ’94. pp. 94-99.

Hoose, Neil, M.-A. Vicencio and X. Zhang. 1992. Incident Detection in Urban Roads
Using Computer Image Processing. T raffle Engineering and Control, April, pp. 236-
244.

Hughes Aircraft and JHK & Associates, Direct Project, Incident Detection Issues. 1992
Report for Michigan Department of Transportation, MDOT—92-04l3, August,
Washington, D. C.

Ivan, John N. 1996. Neural Network Representations for Arterial Street Incident
Detection Data Fusion. Preprint 960096, Transportation Research Board 75th Annual
Meeting, January 7-1 1, Washington, D. C.

Ivan, John N., Joseph. L. Schofer, Chandra R. Bhat, Pen-Chi Liu, Frank S. Koppelman,
and Alvin Rodriguez. 1993. Arterial Street Incident Detection Using Multiple Data
Sources: Plans for ADVANCE. Pacific Rim T ransT ech Conference, July 25-28, Seattle,
Washington, USA, Proceedings Vol. 1, Advanced Technologies, pp. 429-435.

Ivan, John N., Joseph L. Schofer, Frank S. Koppelman, and Lina L. E. Massone. 1995.
Real-Time Data Fusion for Arterial Street Incident Detection Using Neural Networks.
Transportation Research Record 149 7, Washington, D. C., pp 27-35.

Judycki, D. C., and J. Robinson. 1991. Freeway Incident Management. Technical Papers
from ITE's 1990, 1989, and 1988 Conference, contained in Massachusetts Incident
Management Conference Proceeding, DOT-T-92-05, June 6. pp. 359-368.

Kalman, R. E. 1960. A New Approach to Linear Filtering and Prediction Problems.
Journal of Basic Engineering, March, pp. 35-45.

148

Kelly, David. 1991. Results of a Field Trial of the IMPACTS Image Processing System

for Traffic Monitoring. Vehicle Navigation & Information Systems Conference
Proceedings, P-253, Part 1. Society of Automotive Engineers, October, pp. 151- 167.

Knobel, H. C., and E. D. Helfenbein. 1976. Development and Testing of Incident
Detection Algorithms. Final Report, Volume 4: Program Documentation for Calibration,
Evaluation, and Implementation, FHWA-RD-76-22, Technology Service Corporation,
Santa Monica, California, April.

Lam, Joseph K., and D. Bowen Tritter. 1994. Innovative Technologies in the COMPASS
Advanced Traffic Management System. Traffic Technology International '94, pp. 44-
49.

Lan', A., Christianson, D., and S. Porter. 1982. I-35W Incident Management and Impact of
Incidents on Freeway Operations. Minnesota Department of Tran_._s_portation, January.

Levin, Moshe, and Gerianne M. Krause. 1978. Incident Detection: A Bayesian Approach.
Transportation Research Record 682, TRB, National Research Council, Washington, D.
C., pp. 52-58.

. 1979a. Evaluation of Incident Detection Algorithms. Report No.
FHWA/IL/PR-085, Illinois Department of Transportation, Bureaus of Materials and
Physical Research, December, Springfield, Illinois.

. 1979b, Incident-Detection Algorithms, Part 1. Off-Line Evaluation.
Transportation Research Record 722, TRB, National Research Council, Washington, D.
C., pp. 49-58.

1979c. Incident-Detection Algorithms, Part 2. On-Line Evaluation.
Transportation Research Record 722, TRB, National Research Council, Washington, D.
C., pp. 58-64.

. 197%. A Probabilistic Approach to Incident Detection on Urban Freeways.
Traffic Engineering and Control, March, pp. 107-109.

Martinez, J. J ., Guillen S., Sellam S., Boulmakoul A., Boilot F., Diaz E., Ferris R., and
Cavero V. 1994. Evaluation of a Computer Vision Based Automatic Incident Detection
System in an Urban Context. World Congress on Application of Transport Telematics
and Intelligent Vehicle-High way Systems, Toward an Intelligent Transportation
System, Vol. 2.

Masters, Philip H., Joseph K. Lam, and Kam Wong. 1991. Incident Detection Algorithms
for COMPASS - An Advanced Traffic Management System. Vehicle Navigation and

Information Systems Conference Proceedings, Part 1, Society of Automotive Engineers,
Inc., October, Warrendale, PA, pp. 295-310.

149

McTrans Center. 1995. Highway Capacity Software, Release 2.1b, University of Florida,
Gainesville, FL.

Michalopoulos, Panos G. and Richard D. Jacobson. 1992. Field Implementation of the
Minnesota Video Detection System. The 3rd International Conference on Vehicle
Navigation & Information Systems, Oslo, Norway. September 2-4. pp. 105-111.

Michalopoulos, Panos G., Richard D. Jacobson, Craig A. Anderson, and Thomas B.
DeBruycker. 1993. Amtomatic Incident Detection through Video Image Processing.
Traffic Engineering & Control, February. pp. 66-75.

Montgomery, Douglas C., Lynwood A. Johnson, and John S. Gardiner. 1990. Froecasting
& Time Series Analysis, Second Edition, McGraw-Hill, Inc. New York

Parkany, A. Emily, and David Bernstein. 1993. Using VRC Data for Incident Detection.
Pacific Rim. T ransT ech Conference, Proceedings, Volume 1. Advanced Technologies.
ASCE Third International Conference on Applications of Advanced Technologies in
Transportation Engineering. Seattle. Washington, USA. July 25-28. pp. 63-68.

. 1995. Design of Incident Detection Algorithms Using Vehicle-to-Roadside
Communication Sensors. Transportation Research Record 1494, TRB, National
Research Council, Washington, D. C., pp. 67-74.

Payne, H. J., E. D. Helfenbein, and H. C. Knobel. 1976. Development and Testing of
Incident Algorithms, Research Methodology and Detailed Results, Final Report, F HWA-
RD-76-20, Volume 2. April.

Payne, H. J., and S. C. Tigor. 1978. Freeway Incident-Detection Algorithms Based on
Decision Trees with States. Transportation Research Record 682, TRB, National
Research Council, Washington, D. C. pp. 30-37.

Persaud, Bhagwant N., and Fred L. Hall. 1989. Catastrophe Theory and Patterns in 30-

Second Freeway Traffic Data. Implications for Incident Detection. Transportation
Research, Part A: General, Vol. 23, No. 2, March, pp. 103-113

Persaud, Bhagwant N., F. L. Hall, and Lisa M. Hall. 1990. Congestion Identification
Aspects of the McMaster Incident Detection Algorithm. Transportation Research
Record 1287, TRB, National Research Council, Washington, D. C., pp. 167-175.

Ritchie, Stephen G., and Ruey L. Chen. 1993. Simulation of Freeway Incident Detection
using Artificial Neural Networks. Transportation Research, Part C, Volume 1., No. 3.
pp. 203-217.

150

ROper, David H. Freeway Incident Management. National Cooperative Highway
Research Program Synthesis of Highway Practice 156, National Research Council,
December 1990, Washington, D. C.

Roseman, David, and Sean Skehan 1995, Automated Arterial Incident Detection Santa
Monica Freeway Smart Corridor. Institute of Transportation Engineers 65th Annual
Meeting, Compendium of Technical Papers, page 27-31.

Sakasita, Masarni, De Leuw, Cather and Company, and Adolf D. May. 1975.
Development and Evaluation of Incident-Detection Algorithms for Electronic-Detector

Systems on Freeways. Transportation Research Record 533, TRB, National Research
Council, Washington, D. C. pp. 48-63.

Sellam S. and A. Boulmakoul. 1994, Intelligent Intersection: Artificial Intelligence and
Computer Vision Techniques for Automatic Incident Detection. Artificial Intelligence
Applications to Traffic Engineering, pp. 189-200.

Stephanedes, Yorgos J., and Athanasios P. Chassiakos. 1993a. Application of Filtering
Techniques for Incident Detection. Journal of Transportation Engineering, Vol. 119,
No. 1, ASCE, Jan/Feb, pp. 13-26.

. 1993b. Transportation Research, Part C, Vol. 1, No. 3, pp. 219-233.

Stephanedes, Yorgos J ., Athanasios P. Chassiakos, and Panos G. Michalopoulos. 1992.
Comparative Performance Evaluation of Incident Detection Algorithms. Transportation
Research Record 1360, TRB, National Research Council, Washington, D. C., pp. 50-57.

Stephanedes, Yorgos J ., and George Vasilakis. 1995, Accident Detection at Intersections
for ITS. Transportation Congress, Vol. 1, Civil Engineers - Key to the World’s
Infrastructure, Proceedings of the 1995 Conference, San Diego, California, Oct 22-26,
American Society of Civil Engineers, page 239-244.

Tarko, Andraej, Jeffrey Tsai, and Nagui M. Rouphail. 1993. Detection of Traffic
Congestion using Fuzzy Operator Logic in ADVANCE-like Systems. Proceedings of the
Intelligent Vehicles ’93 Symposium, IEEE Industrial Electronics Society, July 14-16,
Tokyo, Japan, pp. 189-194.

Tignor, Samuel C., and H. J. Payne. 1977. Improved Freeway Incident Detection
Algorithms. Public Roads, Vol. 41, No. 1, June, pp. 32-40.

Urbanek, G. L., and Rogers. Alternative Surveillance Concepts and Methods for Freeway
Incident Management - Volume 1: Executive Summary. Report No. FHWA RD-77-58,
March 1978

151

Wallace, Courage, and Hadi. 1991. TRANSYT-71", Users Guide, Volume 4 in a Series:
Methodology for Optimizing Signal Timing, Federal Highway Administration,
December.

Willsky, A. S. 1976. A Survey of Design Methods for Failure Detection in Dynamic
Systems. Automatica, Vol. 12, pp. 601-611.

Willsky, A. S., E. Y. Chow, S. B. Gershwin, C. S. Greene, P. K. Houpt, and A. L.
Kurkjian. 1980. Dynamic Model-Based Techniques for the Detection of Incidents on
Freeways. IEEE Transaction on Automatic Control, Vol. AC-25, No.3, June, pp. 347-
359.

Wirasinghe, S. C. 1978. Determination of Traffic Delays from Shock-Wave Analysis.
Transportation Research, Vol. 12, pp. 343-348.