Video anomaly detection based on wake motion descriptors and perspective grids

Abstract

This paper proposes a video anomaly detection method based on wake motion descriptors. The method analyses the motion characteristics of the video data, on a video volume-by-video volume basis, by computing the wake left behind by moving objects in the scene. It then probabilistically identifies those never previously seen motion patterns in order to detect anomalies. The method also considers the perspective of the scene to compensate for the relative change in an object's size introduced by the camera's view angle. To this end, a perspective grid is proposed to define the size of video volumes for anomaly detection. Evaluation results against several state-of- the-art methods show that the proposed method attains high detection accuracies and competitive computational time.