Research On The Pedestrian Tracking Across Cameras

With the progress of technology and the development of humans’ life, the security problems have attracted more and more attention. Video surveillance is one of the important means to safeguard the social security. The public security organs build Skynet monitoring system established in the large number of cameras is composed of high streets and back lanes monitoring network. The monitoring network is powerful means of public security agencies to fight crime, and is the strong backing of the law and order in the city. Watching the surveillance video by people and doing manual operation is time-consuming and labor-intensive because the installation of video surveillance is more and more, and the external environment has become increasingly complex. So the intelligent video surveillance system is imminent. The research on the pedestrian tracking across cameras on this paper was to detect and track a same target across multi cameras, which solve the shortage of single camera tracking that the visual field of the scene is limited, pedestrian information is not perfect, etc., to achieve the purpose of intelligent monitoring.This paper focuses on three aspects of the personnel tracking cross cameras algorithm research: 1, pedestrian detection of a single camera; 2, pedestrian tracking of a single camera; 3, target matching cross multi cameras. After pedestrian detecting and tracking of a single camera, the targets across multi cameras were matched up, which completed the transfer of the target in different cameras, also completed the cross camera personnel tracking.We first use the HOG feature and SVM classifier classical collocation and pedestrian detection experiment, pedestrian detection is used for a subsequent cross matching of target camera. Then the pedestrian single camera tracking, based on improved optical flow method, not only calculate optical flow but also reverse back flow calculation, ensure the accuracy and stability of tracking.Finally after pedestrian detection and tracking in a single camera have better results, we focused on the target matching across cameras. Matching the target from a camera handed over to another camera successfully means completing the cross camera tracking. This paper presents two algorithms for target matching.The first algorithm is an improved SIFT feature matching algorithm. First, the tracking object was transformed in various forms, and the multiple states of the target were saved, so that more SIFT feature points can be extracted. And then we used SIFT to match the two goals, to achieve the set threshold that match the success of the target from one camera to another camera. The algorithm even in the camera has far and near transform, the angle of the deviation of the situation can also successfully match the target.Another matching algorithm integrates the detection and tracking of the cascade, continue to use the tracking results to guide the classifier is trained. Then target detection, target detection and adjustment process of tracking, tracking more accurate. In this video data, the detection results and tracking results continue to guide each other, making the detection part of the detection target more sensitive, tracking part of the tracking path is more accurate. In a network of multiple camera networks, the target is being tracked continuously in the system of detection and tracking part of the transmission. Online training increases the accuracy of cross camera personnel tracking.With the analysis of experimental demonstration, the method proposed in this paper can effectively carry out cross camera personnel tracking, and get the accurate results of the tracking target in multiple cameras.