Research On Crowd Abnormal Event Detection In Video Surveillance

In the field of intelligent video surveillance, crowd abnormal event detection is one of the key issues, whose main task is trying to filter out various unusual events automatically from surveillance videos automatically, and be able to remind security persons to deal with them timely. Occlusion phenomenon occurs easily because the number of moving individuals is large in a crowd scene, which makes the results of traditional anomaly detection based on moving object tracking become poor and impractical. Therefore, we make a new exploration of crowd abnormal event detection from a theoretical and practical view.For the detection of abnormal regions, this paper uses the scan statistical method based on semi-parametric model. The range of abnormal regions can be determined by setting the size of scanning window. The proposed method first exacts optical flow information of moving objects; then optical flow can be represented as flow word histograms by means of bags of visual words; finally, the likelihood ratio test statistic is calculated between inside and outside the region for discovering abnormal regions.The crowd abnormal events can be divided into two categories: global abnormal events and local abnormal events. For the detection of global abnormal events, the size of scanning window is set as the whole image. The event detected in the abnormal region is the global abnormal event. For the detection of local abnormal events, this paper first takes the local nearest neighbor descriptor as a characteristic factor and trains the mixed membership na?ve bays(MMNB) model for every cell in the videos that only have normal behaviors; then the local abnormal region is equally divided into cells with same size and the most probable abnormal cell can be searched through the maximum optical flow energy method, at the same time, the local nearest neighbor descriptor of the corresponding cell is needed to calculate; finally, the cell with the maximum optical flow energy is judged whether it is abnormal under MMNB model, if the cell is abnormal, the 4-neighbourhood cells will be continued to detect, so it achieves the goal of locating anomaly.In addition, the proposed method is tested on the UMN dataset and UCSD dataset. The experimental results show that our method not only suppresses the influence of interference of light, shadow, etc, but also have some improvement in detecting global and local abnormal events compared to some typical methods.