Several Key Issues Studies In Non-overlapping Multi-camera Surveillance System Tracking

On the premise of camera’s imaging area without overlap, non-overlapping multi-camera tracking achieves the target tracking in the whole monitored area. It includes moving object detection and moving object tracking and objectives related. In the actual scenario, because of the target from background interference, the change of light and the angle of the target shift and other issues between cameras, it increases the difficulty of multiple cameras target tracking. Therefore the research of non-overlapping multi-camera tracking system has practical value but equally exits a big challenge.On the basis of mastering the object detection, object tracking and correlation method and related theory, we research the algorithm of monitoring and control system’s object detection, object tracking and correlation. We implement a demo system without overlapping horizon object tracking. In this paper, main work is as follows:1. Moving object detection in static scene:In view of the natural scene illumination change and ghost, we introduce the motion object detection algorithm based on residuals of sparse representation. It uses the first two principal components of PCA to build a dictionary of the background, then uses the dictionary to extract refactoring residuals. It uses the improved Otsu algorithm to obtain the global threshold image to extract foreground. The foreground image combined image edge information to eliminate ghost area. By adopt the idea of conservative with incremental PCA method to update the dictionary. Experiments show that, this algorithm in target detection has a good robustness. It can effectively resist the effects of the change of illumination.2. Multiple objects tracking in static scene:Using the static scene background invariant features, we combine the results of foreground detection to realize multi-target tracking through using foreground matching. Under the assumption that the foreground target has been extracted, by adopting the idea of forward-backward error to realize two-way optimal matching of the foreground area, no match for the target will be eliminated and increased in order to realize the multiple target tracking under static scene. Experiments show that, under the static scene, this algorithm can tracking foreground moving target quickly effectively.3. Object associated:Under the assumption that the light difference between cameras is not the case, according to targeted local shelter and the difference of angle, we propose a target association algorithm based on sparse representation. We use partitioned histogram feature, for one camera targets feature set add a trivial templates to build a dictionary. Using the dictionary to refactoring the target under another camera, the size of the reconstruction error is used to achieve the goal of the match. Through the experiment we prove that the method which under the condition of illumination change is not very strong can achieve more accurate object association without a priori knowledge.4. System builds:To build target extraction system under a single camera through combining the target detection and target tracking. Then, transmit the extracted foreground to the server through the network, and with the use of the target association system on the server side to achieve the target tracking in a region.