Unsupervised Video Anomaly Detection Based on Sparse Reconstruction

Abstract

[Objective] Video surveillance technology is more and more used in all kinds of scenes, and a large number of surveillance videos are produced every day. In the video, most of the data are normal events, however people focus on only a few abnormal events, how to find individual abnormal events in the massive video is the focus of current research. [Methods] Firstly, the dense optical flow of video is obtained, and the information of optical flow is transformed into the histogram feature of optical flow. Secondly, the space-time cube of video is constructed by using the space-time correlation of video. Finally, sparse representation method is used to model the whole process. [Result] The part in the video with too much reconstruction error is considered as abnormal event. Experiments on the UCSD dataset show that proposed method can effectively detect the abnormal events in the video. [Conclusion] In this paper, a video anomaly event feature and an anomaly detection model are proposed, which provide support for further research.