A UAV-Based Automatic Traffic Incident Detection System for Low Volume Roads
Liye Zhang, Zhong‐Ren Peng, Daniel Sun, Xiaofeng Liu · Transportation Research Board 92nd Annual MeetingTransportation Research Board · 2013
This paper presents an automatic traffic incident detection system for low volume roads based on Unmanned Aerial Vehicle (UAV). Using computer vision techniques, the traffic flow parameters and vehicle trajectories were extracted from the real-time video. Then, various traffic incidents can be detected automatically by using the proposed algorithms, which can detect collisions that happened already or dangerous events that would result in collisions. A unique feature of the system is the Geographic Information System (GIS) based property while with Google Earth embedded, which enables the users to see real time large scale terrain image around the existing UAV location. This system supports video retrieving according to traffic semantics, such as stalling vehicles and vehicle running in wrong direction, which enables users to confirm incidents according to traffic video. To evaluate the performance of proposed system, two field experiments have been done. One was conducted in Gainesville, Florida U.S. in May, 2011. Another was carried out in Feb, 2012, in Xinjiang, China. In the first experiment, the video with 24567 frames contains 2 slow moving incidents, 3 wrong direction incidents, and 55 vehicle stalling incidents. The slow moving, wrong direction incidents detection rate is 100% and the vehicle stalling incident detection rate is 96.36%. For the second experiment, three test cars were driven to simulate incidents, all the vehicles including static ones were detected successfully when the UAV cruised along the highway.