Use of Computer Vision to Identify Pedestrians' Nonconforming Behavior at Urban Intersections

Abstract

The detection and understanding of nonconforming behavior (violations) can be useful in forming safety diagnoses and developing safety countermeasures. Traffic violations occur when road users, including pedestrians, seek increased mobility and disregard traffic laws and regulations. Such behavior can cause additional collision risks. This paper's objective is to demonstrate the automated identification of pedestrian crossing violations with computer vision techniques. Two types of violations are considered. The first is spatial violations: pedestrians cross an intersection in nondesignated crossing regions. The second is temporal violations: pedestrians cross an intersection during an improper signal phase. The methodology primarily relies on the identification of road users' trajectories and separating pedestrians with nonconforming behavior from those with conforming behavior. The methodology is demonstrated on two urban intersections, one in downtown Vancouver, Canada, the other in Kuwait City, Kuwait. The results show satisfactory accuracy in the detection of spatial and temporal violations, with an approximately 90% correct violation detection rate having been achieved in both case studies.