Oriented Tracklets Approach for Anomalous Scene Detection in High Density Crowd

Abstract

Anomalous scene detection in high density crowd has become a topic of great interest for public safety at the venues of mass gatherings. In this paper, we proposed an anomalous crowd scene detection approach in active contour region. To this end, the oriented tracklets are extracted and analyzed in active contour region. The idea of active contouring helps in minimizing the tracking region. The extracted oriented tracklets are quantized into histogram bins based on their flow direction. For each bin of histogram, an entropy and temporal occupancy measures are computed for every temporal window of some frames. The huge deviation in both measures between temporal windows indicates the occurrence of an abnormal state of motion in the scene. Experiments have been conducted on publicly available UCF Web and Violent Flows datasets, and obtained interesting results. The proposed approach is compared with various stat-of-the-art methods against ROC curve and achieve remarkable accuracy while maintain lower computational complexity.