Anomaly detection on traffic videos based on trajectory simplification

Detecting anomalies in the Traffic Control Systems (TCS) could be very useful for the accident analysis, fault detection and other traffic-related topics. In this article we propose a general framework for the trajectory-based anomaly detection, which is fast and reliable. Experimental results show that the system could be used on a vast variety of camera types and configurations. We have used a semi-supervised anomaly detection in the framework which learns from the trajectories of “normal” movements and detects the trajectories that does not fit on the trained model. The trajectories are simplified using a line simplification algorithm to improve the performance while increasing robustness on the noisy inputs.