The Study On People Counting Method Based On Head Detection And Tracking

With the rapid development of intelligent video analysis technology, counting people in video has become a research hotspot in the field. Pedestrian counting system can solve the problem of resource allocation in subways, public buses, airports and other places. Meanwhile,it can also provide timely warning for overcrowding and prevent public safety accidents, as well as provide data support for management methods. However, in the actual scene, the complex environment and the interlaced trajectory will bring missing detection and false detection. Therefore, the accurate counting of people flow has a higher practical demand and significance.In this dissertation, the study focuses on foreground segmentation, head detection and target tracking. Firstly, the mixed Gaussian model with dynamic pixel compensation model is used to detect the foreground of video images; Secondly, the improved Random Hough Transform is used to extract the head in the region of interest; Finally, an improved Kalman filter tracking method is used to realize the real-time tracking of people flow, and the counting area is set up for statistics.The main work of this dissertation includes the following aspects:(1) Aiming at the problem that the background error is caused by the abrupt change of illumination in the video, a mixed Gaussian model based on dynamic pixel compensation is proposed. A dynamic pixel compensation model is established by using the center distance between pixel mean and the background pixel in the same position. Then the detection frame is added to the dynamic pixel compensation model. In this way, the influence of illumination change can be incorporated into the detection frame, so the influence of the fast illumination variation on the Gaussian model is weakened, and a better foreground detection result is obtained.(2) In view of the problem of head occlusion and detection interference, this dissertation makes use of a single camera positioned at a certain vertical angle, which can solve the occlusion problem when people come near and contact. Secondly, the region of interest is determined by the result of foreground extraction, and using improved Random Hough Transform to detect circles in the area of interest. This method not only avoids the interference of circular objects in complex background, but also reduces the invalid circle accumulation and improves the efficiency of circle detection(3) In order to reduce the occurrence of missed detection and false detection, this dissertation proposes a method combined Kalman filter with the minimum Euclidean distance.The Kalman filter is used to predict the position of the target in the next frame, then the thesis takes this position as the center to find the matching target and analyzes the problem of target occlusion, separation and stop. This method solves the tracking missing problem. In the counting process, this dissertation adopts double line method to count people.The people counting method is tested on the MATLAB2013b platform. Experimental results show that this dissertation has higher counting accuracy in background mutation,bi-directional people flow and other scenes,and also effectively solves the problem of tracking failure. Thus, this method meets the requirement of real time applications.