Unsupervised Traffic Anomaly Detection Using Trajectories

Abstract

Traffic anomaly detection of unsupervised videos has attracted great interests in computer vision field, and this task is very challenging since the scarcity of data and scene diversities. In this work, we present a robust framework for solving unsupervised traffic anomaly detection based on vehicle trajectories. The possible anomalies are detected and tracked from background image sequence of videos. The start time of the abnormal events is located by the decision module based on tracks. In order to better solve the problems of false detections and missed detections caused by the detector, we design a multi-object track (MOT) algorithm suitable for this task. We also present an adaptive unsupervised road mask generation method to filter out false anomalies outside the road area. Our method participated in the evaluation of 2019 AI CITY CHALLENGE Track3 and achieved good result.