Online Video Anomaly Detection Methodology With Highly Descriptive Feature Sets

This paper presents a novel methodology for online video anomaly detection. The proposed algorithm divides each video sequence into non-overlapping cuboids, and assigns a state-of-the-art feature vector to each of them. The incoming patterns in testing phase will be then evaluated based on their similarity to the learned patterns. The first achievement of the proposed method is to introduce and apply highly descriptive features and build a histogram of vertical component of optical flow for different regions of the scene. Since the vertical component of optical flow contains both information of magnitude and orientation, it can be considered as an abstract feature rather than using magnitude and orientation, separately. As a result, the dimension of feature vector decreases which leads to reduce the complexity of entire system. The entropy of vertical component is also considered, and hence the differences in velocity and direction of the movements will be monitored. Finally, an efficient technique for anomaly detection search is presented that makes the proposed algorithm an applicable candidate for online performance. The simulation results on UCSD and UMN data sets confirm that the proposed methodology achieves high performance results in case of accuracy and total processing time compared with counterpart approaches.