Research On Mean Shift Based Object Tracking Algorithms

In recent years, the intelligent video surveillance system has become more and more importantin daily life, and it is also an important research direction in computer vision. Its main task is todetect, identify and track the targets of interest from the video stream, and analyze, understand theirbehaviors. Moving target tracking plays a key role in the intelligent video surveillance system, andit is also an important indicator to measure the quality of a surveillance system. So it has importantresearch value.Mean Shift is a kernel function based non parametric estimation algorithm. For the advantagesof no prior knowledge requirement, fast convergence, and real time tracking ability, it has attrackedlots of interests from researchers. However, there are still many drawbacks to be concerned before itcan be put into practice. Therefore, this thises concentrated on the study of Mean Shift so as toimprove its target tracking performance. The innovations and main contributions of this thesis are asfollows:(1)The multi feature fusion Mean Shift tracking algorithm is studied, which represents a targetthrough LTP texture and color information to increase robustness. Since LTP sets the noisethreshold to a fixed value, it is sensitive to noise variation. To deal with this problem, the LMedSalgorithm is introduced for adaptive threshold eatimation to calculate the LTP texture features.Furthermore, tracking window size is adjusted to tightly enclose the target. Experimental resultsshow the effectiveness and robustness of the algorithm.(2)An improved background weighted Mean Shift tracking algorithm which can effectivelyreduce background’s interference in target localization is proposed. Due to the constant or thethreshold decision changing background histogram can’t effectively represents real time changingbackground information, so the real time update background histogram is proposed, while theKalman filtering algorithm is used for target pre position. Experimental results show that thealgorithm can effectively reduce the background characteristics and enhance the target convexcharacteristics. So the position results will be achieved rapidly and accurately.(3)Mean Shift algorithm has limitions on fast moving targets and non Gaussian noiseenvironment. Particle filtering algorithm is suitable for any state space model, but it needs a largenumber of particles to achieve the effective tracking results. Mean Shift algorithm is induced tomake each particle to local maximum as to reduce the number of particles. At the same time, when the target is seriously occlusioned, only Particle filtering algorithm is used for tracking, and afterthe occlusion disappears, both Particle filtering and Mean Shift algorithm are all used to achieve thetracking result. Further the target window width is estimated, and the template tracking is updated.The experimental results show that the effectiveness of the proposed algorithm.Finally, the summary of this thesis is made, and further research is discussed as well of thisstudy, so future research direction is pointed out.