Improvement in detection of presence in forbidden locations in video anomaly using optical flow map

Abstract

Anomaly detection has been in researchers’ scope of study for a long time. The wide variety of anomaly detection use cases ranges from quality control in production lines to providing security in public places. One of the most attractive topics in anomaly detection is in video surveillance systems. In this paper, we propose a method that works based on frame prediction and optical flow to improve anomaly detection in videos. The use of optical flows in normal frames helps the system to better detect the entrance of people or objects to forbidden areas by its information about the amount of movement in different regions of the frames. Based on the optical flow of normal videos and that of current video, the threshold for anomaly decision is adaptively adjusted. This could ultimately lead to a better overall performance of the anomaly detection system compared to the recent similar works. The presented method is general and can be simply incorporated to other video anomaly detection systems to improve the detection accuracy.

Keywords

• anomaly detection
• video surveillance
• optical flow
• deep neural network