Research On Object Tracking With Multi-camera

As the development of security industry, video surveillance device has changedthrough the analog device to the digital device and to the network device. Networkingand intelligent video surveillance system is an emergence research field and has alreadypermeated into our daily life. We can see all kinds of video surveillance system inpublic area, for example, airports, schools and other public places. Video surveillancesystem is a complex system including all kinds of technology, such as computer vision，communication and automation.This thesis is focusing on the key technology of multi-camera video surveillancesystem. Based on the multi-camera environment, we make analysis and comparison ofthe detecting algorithm, tracking algorithm，object relay and camera collaborationalgorithm, which are the necessary technology in a multi-camera surveillance system.The main research works are as follows:1. Analysis of the key technology of multi-camera tracking system, includingobject detecting, tracking, matching and multi-camera collaboration technology.2. As the requirement of the detecting and tracking algorithm, we did a research inthe performance of the descriptors (SIFT BRISK SURF ORB) in the small-scalepictures. Experiment demonstrates that SIFT should be the best for feature descriptionin small-scale area.3. We also did a research of histograms of oriented gradients (HOG) and trainedthe small-scale HOG classifier offline, which is more suitable for detecting thesmall-scale pedestrian detecting. Based on the particle filter tracking algorithm, wecreatively proposed a tracking algorithm combining the SIFT and HOG features, whichgreatly improved the robust of the tracking algorithm compared to the traditionalparticle filter.4. We did a research on two difference target handoff scenario. In the overlappingview, we use camera field of view line, which automatically generated by SIFT andhomography matrix. We decide the position by the target’s center-point and land-point.In the non-overlapping view, we fused the SIFT feature, color feature and UV feature based on Dempster-Shafter evidence theory for object matching and handover in amulti-camera environment.5. We made an analysis of the features of multi-camera collaboration frameworks.Based on the game theory, we realized the game theory algorithm with Matlab GUI andverified the effectiveness of this algorithm. The platform we build could be used for thefuture research on camera control framework.