Research On Multiple Targets Matching And Tracking Based On Activity Information For Camera Networks

Intelligent camera monitoring has far-ranging prospects and with a high research value in camera network, in which tracking a moving target is the basis. In single camera, the processing target block is the core issue of target tracking. The exterior features of template and movement area are utilized by conventional tracking methods to achieve the goal of matching in adjacent frames and outlines the target trajectory, and they are doomed to fail when the presence of occlusion.Another camera will be provided to supplement additional constraints information to solve the problem that single camera itself can not provide in-depth information about the target. Therefore, the use of multiple cameras for tracking target is faced with the target image matching problem. The routine image matching method of target image based on color, texture, Scale-invariant feature transform as matching features. But in pratical surveillance scenario, due to changes in illumination, the difference camera perspective, the target size accounted for a small proportion of the imaging plane etc., the matching algorithm based on the above characteristics are proved to be applied only under the particular scenario. In particular fixture occlusion and mutual occlusion during exercise will result in decreasing of matching accuracy. Therefore, this article takes matching algorithm scalability on environment into account, We study the following four areas:1. In single camera, we proposed motion region based tracking algorithm. Cost function is constructed, which makes use of size, direction and velocity of motion region and be used to realize association between template and target. Consider the case of merger and separation between the targets, using different template update tactics to handle the problem of occlusion.2. Using MOTP and MOTA as the performance evaluation index of tracking algorithm. The Multiple Object Tracking Precision(MOTP), which shows the tracker's ability to estimate precise object positions, and the Multiple Object Tracking Accuracy(MOTA), which expresses its performance at estimating the number of objects, and at keeping consistent trajectories.3. We proposed robust spatiotemporal trajectory, which considering not only the statistical features of time, and taking the spatial features into account, has geometry of independence and uniqueness. The spatiotemporal trajectory is robust to shape,illumination and camera angle, meanwhile does work without camera calibration.4. Based on the characteristics of spatiotemporal trajectory, we proposed an improved sequence similarity calculation method based on mutual information, the method can effectively eliminate the local similarity between the targets' spatiotemporal trajectory. To obtain optimal matching result, making use of the spatiotemporal trajectories to bulid a bipartite graph, of which the weighted parameters be calculated from similarity of spatiotemporal trajectories.5. This paper regards mismatching rate and the average deviation value as a matching algorithm evaluation for matching results. Simulation and analysis on matching algorithm based on spatiotemporal trajectory under a variety of monitoring scenarios were conducted, meanwhile simulation on matching algorithm based on Scale-invariant feature transform were conducted as a contrast. The results show that the proposed algorithm can achieve a high matching accuracy, while maintaining the value of the average deviation be similar to result of matching algorithm based on Scale-invariant feature transform.