Vision-Based Intersection Monitoring: Behavior Analysis & Safety Issues

The main objective of my dissertation is to provide a vision-based system to automatically understands traffic patterns and analyze intersections. The system leverages the existing traffic cameras to provide safety and behavior analysis of intersection participants including behavior and safety. The first step is to provide a robust detection and tracking system for vehicles and pedestrians of intersection videos. The appearance and motion based detectors are evaluated on test videos and public available datasets are prepared and evaluated. The contextual fusion method is proposed for detecting pedestrians and motion-based technique is proposed for vehicles based on evaluation results. The detections are feed to the tracking system which uses the mutual cooperation of bipartite graph and enhance optical flow. The enhanced optical flow tracker handles the partial occlusion problem, and it cooperates with the detection module to provide long-term tracks of vehicles and pedestrians. The system evaluation shows 13% and 43% improvement in tracking of vehicles and pedestrians respectively when both participants are addressed by the proposed framework. Finally, trajectories are assessed to provide a comprehensive analysis of safety and behavior of intersection participants including vehicles and pedestrians. Different important applications are addressed such as turning movement count, pedestrians crossing count, turning speed, waiting time, queue length, and surrogate safety measurements. The contribution of the proposed methods are shown through the comparison with ground truths for each mentioned application, and finally heat-maps show benefits of using the proposed system through the visual depiction of intersection usage.