Research Of Feature Extraction And Object Tracking On A Moving Camera

Today, image processing technology based on computer vision is developing rapidly. Target detection and tracking algorithms receive wide attention for the value of the practical application. And People had realized some of them in public places in security monitoring and intelligent transportation applications results. Although there have been many classic and effective target detection and tracking algorithm, but in real environment, these algorithms are facing the challenge of many aspects, such as the moving camera, the illumination changes, the deformation of target, and complex background. Therefore, the question that designing a robust and fast target detection and tracking algorithm is still need to be resolved.In this paper, shows part of existing target detection and tacking algorithms in detail. And, it analyzes the basic principles and concrete steps of these algorithms, then points out their problems, but also raised related solutions for some of the issues.Firstly, we first study feature extraction method used in target detection and tracking algorithm and analysis both their advantages and disadvantages by comparison. So we get the feature extraction algorithms that easily used in target detection and tracking. Then it proposed a target tracking method by combining feature extraction and target detection.Second, it also researched the traditional Mean-shift tracking algorithm and achieved simulation, then summarized its deficiencies. The main problem is that the target window keeps the same size during the whole tracking which lead inaccurate tracking result and even failure to track. To solve this problem, we suggest a Mean-shift tracking algorithm based on SURF features. This algorithm will change the size of tracking window adaptively according to the result of SURF matching so that we can get a more accurately location.Finally, the suggested algorithm also redesigns an initial point in Mean-shift iterative by the SURF feature matching result so that the algorithm can converge to the real position of the target more steadily and reduce the number of iterations.We used some videos of real-world environment catched by a moving camera in the experiment. And we find that the proposal algorithm achieved a better tracking result and had less iterations by comparing the results of the proposal algorithm and traditional Mean-shift tracking algorithm.