Research Of The Object Matching Algorithm In Non-overlapping Multi-camera

With the rapid development of intelligent video surveillance technology, people’s requirements for smart intelligent video surveillance system are increasing more and more in recent years. Single camera for tracking and target detecting has developed perfectly, but it’s hard to apply to the broadness of the surveillance areas because of its limited scope of monitoring. Due to the consideration of capital and resources, it is impossible for the cameras to cover the whole surveillance areas with a great amount of monitors. Single camera target detection and tracking technology is mature, and how to track and judgment objects through these blind area to another is hard, the no-overlap multi-camera target tracking is a research hotspot.As the difference between the multi-camera, the same targets must have very different characteristics in them, and the time and space of them is separated, so how to match the same target in non-overlapping vision is the key problem in target tracking, we propose a multi-feature data confusion algorithm which combines several object appearance features with the topology relationship among the cameras. The main jobs and contributions in this paper are list as follows:1. About the object detection, the thesis proposes a multiple moving object detection algorithm based on VIBE and HOG feature classification. First, find the possible area of object by VIBE. Then, accurately segment the object by the SVM classifier based on HOG feature. The experimental results show that the algorithm can effectively solve the object shadows and objectives adhesion problems which VIBE algorithm can not solve, and it is better than HOG algorithm in terms of time-consuming or detection accuracy. The algorithm has good robustness, and can accurately detect moving objects.2. For the matching problem of object in non-overlapping multi-camera system, multiply appearance features and their matching algorithm are proposed including color histogram corrected by bright transfer function, LMNN model and SURF feature points. BTF can eliminate differences among multi-camera and compensate for the color histogram disadvantage of the lack of spatial information; The H component of LMNN can increase the distance of target based on different, so that the same target becomes small,easy to match between; the SURF feature point is good robustness of rotation and scale changes, the similarity between the moving targets is obtained by comparing the distance between the feature points. The three matching models can work together to improve the accuracy of the model matching.3. Research the acquisition of camera topology. This paper use a online topology obtain method. We obtain the import and export of a single camera by the objects detection results, and use the Gaussian model to describe the transfer time between different cameras and provide a reliable time-constraint relations of target matching and association. Its effect is good.4. Considering the characteristics of target tracking about overlapping vision, using D-S evidence theory to fuse the above features, and realize tracking the goals which through non-overlapping region, avoid the conflict of the above characteristics, and improve the accuracy of the target tracking.