Video Anomaly Detection based on Optical Flow Information

Abstract

Although feature descriptors based on optical flow information, are widely utilized in the field of abnormal video anomaly detection at crowd scenes, the performance of descriptors to detection result, with getting rid of the influence of recognition method, is worth to pay more attention. In respect of the accuracy not assured during calculation process, we present a novel feature descriptor in this paper, and name it to be ‘Probability Descriptor based on Optical Flow Orientation and Magnitude’. We firstly do respectively build 1D statistics histogram based on magnitude and orientation of optical flow with the Fuzzy-Partition Mechanism for each pixel; furthermore, build Combinatorial Histogram of Orientation and Magnitude with joint probability distribution for each pixel; finally, build one feature vector for video region through summation operation. For evaluation our proposed descriptor’s performance, we do a series of experiments on publicly available UCSD anomaly detection datasets. The experimental results demonstrate that it outperforms other state-of-the-art descriptors based on optical flow information field.