Monocular camera-based object detection plays a critical role in widespread applications including robotics, security, self-driving cars, augmented reality and many more. Increased relevancy is often given to the detection and tracking of safety-critical objects like pedestrians, cyclists, and cars which are often in motion and in close association to people. Compared to other generic objects such as animals, tools, food — safety-critical objects in urban scenes tend to have unique challenges. Firstly, such objects usually have a wide range of detection scales such that they may appear anywhere from 5-50+ meters from the camera. Safety-critical objects also tend to have a high variety of textures and shapes, exemplified by the clothing of people and variability of vehicle models. Moreover, the high-density of objects in urban scenes leads to increased levels of self-occlusion compared to general objects in the wild. Off-the-shelf object detectors do not always work effectively due to these traits, and hence special attention is needed for accurate detection. Moreover, even successful detection of safety-critical is not inherently practical for applications designed to function in the real 3D world, without integration of expensive depth sensors. To remedy this, in this thesis we aim to improve the performance of 2D object detection and extend boxes into 3D, while using only monocular camera-based sensors. We first explore how pedestrian detection can be augmented using an efficient simultaneous detection and segmentation technique, while notably requiring no additional data or annotations. We then propose a multi-phased autoregressive network which progressively improves pedestrian detection precision for difficult samples, while critically maintaining an efficient runtime. We additionally propose a single-stage region proposal networks for 3D object detection in urban scenes, which is both more efficient and up to 3x more accurate than comparable state-of-the-art methods. We stabilize our 3D object detector using a highly tailored 3D Kalman filter, which both improves localization accuracy and provides useful byproducts such as ego-motion and per-object velocity. Lastly, we utilize differentiable rendering to discover the underlying 3D structure of objects beyond the cuboids used in detection, and without relying on expensive sensors or 3D supervision. For each method, we provide comprehensive experiments to demonstrate effectiveness, impact and runtime efficiency.
Read
