Detection And Analysis Of Crowd Abnormal Behavior In Intelligent Video Surveillance

This paper mainly studies the detection and analysis of crowd abnormal behavior in intelligent video surveillance. The crowd abnormal behavior was divided into many kinds of situations, this paper defines the crowd abnormal behavior as crowd panic to escape, fire evacuation and other unexpected behavior in public places. Abnormal behavior detection was realized through detection and track of moving target, analysis of behavior and recognition of abnormal behavior.Before detecting image motion, the image was became more clearer through preprocessing operations such as gray, denoising, space and frequency domain image enhancement. Then, the mixed gauss model was used to obtain the background image of the video, and the moving object was detected by the difference of the background image and the current image. KLT(Kanade-Lucas-Tomasi)tracking algorithm was used to track the moving people and obtain the velocity、direction and position information of crowd.Next, according to the behavioral similarity of individual particles and other particles in the crowd from the direction and velocity, collectiveness features was extracted to describe the crowd behavior. Collectiveness features was able to describe the crowd behavior very well which integrated the motion information of the whole crowd. According to the average speed of moving people and the grid distribution of individual particles, grid distribution features of particles was extracted to describe the crowd behavior. Grid distribution features of particles was able to describe the crowd behavior well.Finally, the histogram entropy and SVM was used to recognize unexpected abnormal behavior. The collectiveness features was projected into the corresponding histogram, the entropy of the histogram was calculated to compare with trained threshold T and recognize abnormal behavior.Grid distribution features of particles was take into trained SVM behavioral classifier to recognize abnormal behavior.The recognition algorithm of crowd abnormal behavior based on grid distribution features of particles and the recognition algorithm of crowd abnormal behavior based on collectiveness features was validated in UMN and USCD Ped1 database respectively, compared with all kinds of qualitative quantitative result, the proposed algorithm in this paper are better.