A low dimensional descriptor for detection of anomalies in crowd videos

In this paper a novel descriptor is proposed for anomaly detection in crowd videos at a global level. Traditional approaches for anomaly detection in crowd videos face the dilemma of trade-off between high accuracy and real time performance. In order to resolve this issue, we propose an efficient descriptor composed of three different features extracted from the optical flow (OF) of a video sequence. The first feature is the sum of the optical flow field magnitude computed after applying a threshold. The second feature is the joint entropy of the OF magnitude of two consecutive frames used to measure the dissimilarity. The third feature is the variance computed from a space–time cuboid constructed using history of the OF field magnitude. Performance of the proposed descriptor is evaluated on the widely used UMN dataset in terms of accuracy and processing time. For UMN dataset, the proposed descriptor provides the highest area under the curve (AUC) compared to the approaches already published in literature.