Evaluation of Machine Vision-Based Automated Detection of Freeway Incidents

Automatic incident detection (AID) can significantly decrease the time to detect incidents and better identify incident locations to allow faster response. The deployment of traffic surveillance cameras has motivated transportation agencies to investigate computer vision technologies for AID using live video streams from these cameras. This paper evaluated three commercially available computer vision-based AID systems. The evaluated AID systems were able to use video streams from six existing Pan-Tilt-Zoom (PTZ) surveillance cameras for incident detection when these cameras were on predefined presets. The three systems with the selected settings performed differently. For example, the system with the highest number of true stopped vehicle incident detections also had a relatively high number of false alerts. The AID system accuracy also varied by study site, likely due to traffic patterns, roadway geometry, and camera angles. Two major factors contributed to false alerts: stop-and-go traffic during non-incident conditions and other highway objects, including light reflection and glare, that were falsely detected as incidents. The three systems also detected additional incidents not reported in existing incident management databases, but these were mainly shoulder incidents and work zone vehicles.