Object Tracking Based On Local Feature Matching

Along with the computer technology, A.I. and image video process development, Video surveillance and computer machine vision was used in many fields. In intelligent surveillance, traffic detection, security monitoring and aerospace, target detection and tracking technology has showed tremendous economic value.Thi's paper aimed to improve the existing tracking algorithm to achieve better performance. The study process can be divided into 3 parts:Firstly, this paper studied of the existing algorithm and then summarized their own advantages and disadvantages respectively; Secondly, we chose the target detection and tracking under the fixed scene as the main research object; Finally, this paper optimized some important algorithms in this field based on theoretical and experimentally analysis, and the result is good. All the work detailed as follows:First of all, we did research on moving target detection technology. In this part, we focued on the interframe difference and background subtraction these two major target detection on existing technology, and after the analysis of its advantages, disadvantages and the suitable scope of each method, we improved the three differential method by combining it with background subtraction method, which performs well.Then we did more analysis of moving target tracking. The research work focued on Kalman filtering, Mean-shift tracking. We did the simulations on Matlab platform and gave the personal viewpoint about object track.At last, we focused on the target tracking method proposed in this paper which based on local feature matching and then build a database about object outline and the improved Shape Context characteristics. In this part, we learned about the shape context and then proposed an improved approach. First, we extend the polar angle domain of shape context feature to accommodate the slight deformation of target boundary to get better matching; then we improved the selection of boundary points; Finally, we implemented target matching by using of target contour information created in the database to instead of the contours of target. The simulation on Matlab proves that the target centroid we tracked by the new approach is very close to the actual target centroid, and when the target is slightly shielded,we still have a better performance than original algorithm.