Research And Implementation Of Motion Tracking Algorithms Based On Monocular Video

Motion tracking based on monocular video is one of the hottest topic in the field of computer vision. With comprehensive application prospect in the area of intelligent interface, interactive between human and machine, virtual reality and motion analysis, the research of motion tracking algorithm has both important practical and theoretical value.There are still some unsolved problems in the task of motion tracking, according to the status in quo from abroad and at home. The difficulties come from the following reasons: targets are in complex background or the lighting is varying; there exists part collision or completely block; targets are in stretching, deformation, or rotation. In this paper, the basic idea of current tracking algorithm is summed up, the motion tracking algorithm based on MS theory called MS tracker is realized, and to address the problem of fixed searching window, a continuously adaptive search window algorithm: CAMS tracking algorithm called CAMS tracker is realized.Mean shift is a common iterative algorithm of estimating the probability density function, with simple principle and efficient iteration. With the appropriate kernel function used, it will be able to transform the tracking task into the convergence process of mean shift. MS tracker is proved to complete tracking in approximately real time, and to a certain extent, it can handle part collision. However, it can not deal with varying scale targets, but the CAMS tracker can.Fragment tracker divides tracked object into several fragments and extracts integral histogram for each block. The voting map for each block is computed, and then all the voting maps are combined, and finally, the estimated location of the target is obtained. As the dividing block of object area, and the combined voting maps, it has a good performance to handle part collision in complex background or with some kind of posture change. In the process of formatting blocks, it makes up for missing spatial information in the traditional histogram feature. The complexity of the algorithm keeps approximately invariable from different size of targets.Analysis of the MS, CAMS and Fragment tracker is given with experiment on video test sets of INRIA Lab, France. On handling part collision, CAMS tracker obviously works worse than the other two trackers; in a complex background, Fragment tracker does a better job than MS; Both MS and Fragment trackers have recovery capabilities in the event of part missing target; selecting the suitable search radius is particularly important for Fragment tracker.Lastly, a prototype system of motion tracking is designed and implemented.