"Object detection is considered one of the most critical components of any Computer Vision (CV) system. For many real world CV systems such as object tracking, recognition, and alignment, a good localization of the targeted object is necessary as an initialization step. In this thesis, we explore object detection when used in real life scenarios, and study the effect of challenging illumination scenarios impacting the general performance. More specifically, we study two main challenges: underexposed dark images captured during the night time, and overexposed bright images with sunlight projected on the object. We initially study two detection applications used in real-life scenarios. The first application used a Correlation Filter (CF) detector for detecting small objects. CF's were robust in handling object detection with minimum appearance variations, but suffers in handling illumination challenges causing many false alarms. To overcome this issue, we propose to use a post-refinement method to eliminate any false detections caused by specular light. The second detection application used Convolutional Neural Networks (CNN). We learned a total of five detection CNN's, constructing the inputs of a Multiplexer-based method which controls the flow of the CV system, and is driven to select the appropriate CNN based on physical estimated measurements. The five CNN's include a global frame object detector, three local object detectors used for tracking, and finally an object contour detector used to enhance the 2D detection and infer 3D localization of the target. With sufficient training data, all five CNN's were proven to generalize well for a wide range of illumination variations introduced from weather changes, along with many other visual challenges. However, the challenging underexposed light images collected during night time led to system failure. In this thesis, we propose two CNN models to handle both illumination challenges: (1) a temporal delighting Guided-CNN scheme for recovering overexposed video frames caused by sunlight, which is based on the analysis of directional light. (2) An iterative CNN-based technique to synthesize good lighting images suffering from underexposed low intensity lighting. We demonstrate that our approach allows the recovery of plausible illumination conditions and enables improved object detection capability. An extensive evaluation on several CV systems was carried out, including pedestrian detection, and trailer coupler detection."--Pages ii-iii.
Read
