Optical Flow Based Anomaly Detection in Traffic Scenes

Anomaly is any abnormality in a scene that largely differs from the regular pattern. In Indian traffic scenarios we will find this anomalous behavior of vehicles in a larger scale. In this paper we consider three specific kinds of anomalies that occur on the road, viz. vehicles crossing road, pedestrian crossing roads, and change in motion patterns of vehicles on road. Proposed method can assist the driver as a part of ADAS. Traditional camera based methods detect abnormal motion using orientation only of the optical flow. Our method fuses both magnitude and orientation information. Performance obtained on benchmark ITS dataset is presented and analyzed. An average accuracy of 75.33% was obtained with this method.