Research On Video-based Pedestrian Counting Technology

The rise of video surveillance system is in the1970s, but the system has now beenmore than just a simple scene recording function. As computer vision, pattern recognition,machine learning and other related technologies develop and mature, and in the past manyoperations needed manual analysis have been able to complete tasks through automationof these technologies, therefore more accurate to say that now the video surveillancesystems are intelligent video surveillance systems. The video-based human trafficstatistical technique in intelligent video surveillance system is an important application.Human traffic statistics are an important basis for management and decision-making forshopping malls, subways, airports and other public places, and therefore the human trafficstatistical technique has a very wide range of applied prospects.Video-based human traffic statistical techniques typically include target detection,target tracking, object counting and other related technologies. This paper presents a studyon a human traffic statistical algorithm based on HOG feature and tracking. The algorithmuses head detection as the counting basis for pedestrians to avoid the processing targetocclusion; Using HOG feature to describe the characteristics of the target appearancemakes the algorithm having the ability anti illumination changes and robust; Usingdetecting and tracking integration for target tracking algorithm makes the algorithmhandle the multi-target tracking problems in the scene efficiently; The two ways forcounting, counting across the line and the area, meet the needs of practical application. Forthe scene camera angles problem, we propose a solution based on multi-classifier withoutincreasing the complexity of the algorithm to improve the accuracy of target detection, andgive a selection algorithm from multiple classifiers to select the best classifier. Thepedestrian statistical target sizes are unknown and have a variety of scales in the scene, wepropose an online learning algorithm that can automatically learn the optimal targetdetection scale in the scene, so we can detect target without multi-scale scanning method,which greatly accelerates the time efficiency of human traffic statistics algorithm.