Towards rapid safety assessment of signalized intersections: an in-depth comparison of computer vision algorithms

The development of rapid safety assessment tools for signalized intersections is a pivotal focus in contemporary research. Recent advancements in technology and artificial intelligence (AI) have driven the development of proactive safety assessment tools that highlight hazardous instances and provide accurate safety insights. These tools utilize various detection methods, such as Closed-Circuit Television (CCTV) cameras, Light Detection and Ranging (LiDAR), Unmanned Aerial Vehicles (UAVs), infrared (IR), and loop detectors. For nationwide deployment, the most cost-effective methods use existing surveillance cameras. This study investigates different computer vision detection algorithms, focusing on training datasets, video analysis duration, detection methods, types of detected objects, and strengths and weaknesses of each algorithm. It examines convolutional neural network (CNN) detection algorithms based on bounding boxes, 3-D bounding boxes, and key points. These algorithms were applied to video footage from two case study intersections: the Town Square intersection in Jackson Hole, Wyoming, with three cameras in a rural setting, and the Four Corner Camera intersection in Cold Water, Michigan, with one fixed camera in an urban environment. Key findings suggest that 3-D bounding box detection algorithms are more effective at lower elevations for estimating occluded parts and extracting accurate trajectories. Higher-elevation cameras benefit more from bounding box algorithms for faster processing, while key point detection algorithms excel at intersections with multiple cameras, providing accurate depictions and localization of road users. These results offer valuable recommendations for developing accurate, cost-effective, and time-efficient safety assessment tools.