System Evaluation and Development of ST-MRF Incident Detection System based on Statistical Analyses of Traffic Behavior

摘要

Surveillance for human activity has been of interest for some time now, especially in the field of ITS. With advances in computer vision, traffic monitoring has been making progress towards improved road and traffic safety. But problems still exist in trying to make computer vision recognize certain automobile and traffic incidents. And in cases where serious occlusion or more complex congestion situations occurs, developing a high accuracy real time incident detection system could be the key. Only with system evaluation and development of a high accuracy detection system can unnecessary work load of humans at traffic monitoring centers be reduced, emergency responses become faster, and collision avoidance becomes more possible. For these purposes, we performed system evaluation and further developed a high accuracy incident detection system with semantic hierarchy of operations that can precisely understand context of traffic events, similar to how an operator visually understands certain traffic scenes. This paper describes algorithms developed to hard to detect incidents such as single congestion alert for long congestion periods, beginning of congestion, and non-interfering slow cars with logical reasoning focusing on relative behavior among vehicles, and classification with continuous variables via hyperplane.