Research On Video Sequence-based Detection Of Pedestrian Flow

With the extensive application of video surveillance system in current society,there are more and more researchers in the area of intelligent control. Especially,the flow detection technology has very good prospects for development in videosurveillance system, and it is widely used in the subway, roads, shopping malls,companies, banks and public transport. The flow detection technology needs imageprocessing, computer vision, and knowledge from other fields, many researchers athome and abroad have got some research achievement. However, the flowdetection technology is still a difficult problem in the theory and application.Based on the fruits of former researchers, this paper studied the pedestriandetection, identification and tracking methods systematically.The flow detection technology mainly includes pedestrian detection,identification and tracking. In detection of pedestrian moving region, the movingtarget detection methods used commonly are introduced in detail, and we comparedthe methods of optical flow, background subtraction and inter-frame difference. Onthis basis, we proposed a Pedestrian Detection method which is the combination ofbackground subtraction method and inter-frame difference method. This method ismore effective in the detection of stationary or slowly moving pedestrians andreduces the false detection rate of pedestrians.In terms of pedestrian recognition, we compared and analyzed the commonlyused methods of pedestrian recognition, considering that for the camera angle, theshape of human head approximates a circle, so, this paper used the Hough methodsto identify human head. For the original Hough algorithm needs large amount ofcomputation and is not effective in practice, this paper combined the shape angelmethod with Hough algorithm to get the position of head and the location ofpedestrian. Our method reduces the amount of computation greatly and improvesthe recognition efficiency.Referring to pedestrian tracking, this paper comprehensively analyzed somecommonly used tracking methods and described the basic principle of Mean Shiftalgorithm and Kalman filter in detail. If the pedestrian moves too fast or blocks each other during the tracking process, the Mean Shift algorithm is liable to losetracking. In order to improve the reliability of tracking, this paper presented a newmethod combining the Mean Shift algorithm and Kalman filter to track thepedestrian, using the information of target location to optimize the tracking.Finally, with the methods of pedestrian detection, recognition and trackingmentioned above, we recorded the information of target location and counted theamount of pedestrian based on the video sequence.