Using automated video processing to identify pedestrian-vehicle conflicts

Abstract

Pedestrian-vehicle crashes are rare occurrences. Using solely crash data to improve safety on corridors and intersections, therefore, has its limitations. The Georgia Tech Research Institute (GTRI) has developed a tracking system for Georgia Department of Transportation (GDOT) which identifies pedestrians, vehicles, and cyclists from a video. This tracker system outputs the frame by frame coordinates of each of these objects. Using the tracker outputs and existing developed tools, the work of this thesis developed a conflict analysis tool. Written in Python, this tool analyzes each individual pedestrian to see whether a vehicle conflicts with a pedestrian. Once a conflict is identified, the intersection point between a vehicle and pedestrian is found, and the post-encroachment time is calculated (PET). Post encroachment time is the time differential between a vehicle and pedestrian at the intersection point. The conflict analysis tool outputs visualization plots to identify where conflicts occur, as well as general statistics about type of conflict. Results have shown that the conflict tool is 89% accurate in identifying conflicts. Additionally, this software works within the existing tools which GTRI developed and possesses the ability to jump to each identified conflict in the video. Using conflict analysis allows GDOT to identify unsafe locations and implement proactive measures to improve safety.