The Crowd Analysis Research In Video Surveillance

The traditional video monitoring system, the use of human resources still occupies the main part, which need to have special staff to surveil the video in real time. However, the number of cameras is increasing and people behavior is becoming complex, which leads to monitor staff slack, thus missing key information. Hence, further research of video content becomes the emphasis in current video surveillance system.In the video surveillance system, crowd analysis is the most important part to realize intelligently. The crowd as the most important goal in surveillance has the diversity of behaviors and scales. In individual perspective, pedestrian tracking and compared is the critical content. In the perspective of crowd, the distribution and status should be detected intelligently. Therefore, to analyze the different scales of crowd becomes a problem to be solved.In this paper, pedestrian counting in crowd analysis is studied. 4 algorithms in pedestrian counting are summarized: pedestrian counting algorithm based on individual detection, pedestrian counting algorithm based on division of briquette, pedestrian counting algorithm based on statistical regression and pedestrian counting algorithm based on virtual doors. There are two difficulties among these algorithms. The first, crowd occlusion is unavoidable which leads to counting leakage. The second, the complexity of the algorithm is higher in order to guarantee accuracy.After further research, we focus on accuracy and real-time performance. In order to improve these factors, this paper puts forward a novel algorithm based on flux and occlusion index. Inspired by fluid mechanics we proposed to model the pedestrian flow as time-dependent fluid, and estimate the pedestrian flow using flux. The method avoids detecting pedestrians and reduces the time complexity. In addition, we propose an edge-interval algorithm in high-level crowd situation to estimate the occlusion index, which contribute to the flux. The occlusion index can implement adaptive computation based on different occlusion situation. Consequently, the pedestrian flow is estimated using linear regression by combining the flux and occlusion index. Experiments on PETS2009 and real videos, which include the occluded scenes, elucidate the good performance in pedestrian flow estimation.Finally, according to the requirements of practical application of the public security department, we design and set up the video surveillance system for crowd analysis. In system solutions, hardware, operating system and algorithm scheme, we put forward and choose the suitable framework: Intel framework, Windows operating system and intelligent backend. After that, we design the video surveillance system for crowd analysis and develop application software for monitoring personnel. 4 algorithms of crowd analysis in the video surveillance system are studied and introduced, especially their principles. Person reidentification algorithm can associate the same person in different cameras. The algorithm is based on pedestrian detection, HOG feature extraction and reidentification distance calculation. The reidentification achieves function about target search and target path. The video extraction and retrieval algorithm is based on pedestrian detection and tracking. After that, it condenses the pedestrian data into an image and deposits into database. The algorithm effectively reduces the workload of surveillance personnel and improves the surveillance efficiency. Crowd density estimation algorithm can estimate density level and crowd distribution in the region of interesting. The algorithm realizes the crowd state real-time surveillance. Pedestrian counting algorithm, Based on flux and occlusion index, counts the number of crowd. Finally, the analysis system with 4 crowd algorithms is built in Shanghai Public Security Bureau. applied The test results conform to the requirements of the indicators and the system software has a good press.