The Target Tracking Algorithm Research Based On Mean-shift

Target tracking is the process of locating a moving target in a video sequence,which is a key issue in the computer vision community. In recent years, it has avariety of uses, such as cybernation, human-computer interaction, medical imagerecognition, video compression, and so on. Among the various target trackingalgorithms, Mean-Shift tracking algorithm is widespread concerned due to itsadvantages of strict in theory, simple and easy to implementation, and better trackingperformance.This thesis aims to investigate the target tracking algorithm based on theimproved Mean-Shift. Traditional Mean-Shift algorithm is a nonparametric estimationmethod based on density gradient and modeling in a kernel histogram manner. Usually,the target area of interest is manually selected in the first frame of a video sequenceand establishes the target histogram model. Use Bhattacharyya coefficient assimilarity measure and then searches for the most similar areas iteratively insubsequent frames in the video sequence. For the advantages of low computationcomplexity, not sensitive to partial occlusion, rotation, deformation and the movementof the background, etc, it has been widely used in target tracking areas. However,there are also some disadvantages in Mean-Shift algorithm, such as in the situation ofthe change of targets’ size, the kernel bandwidth of traditional Mean-Shift can not beadapted to size change of the target. If the target and the background is too similar todistinguish, the histogram modeling method can not completely meet the needs oftarget feature description, which results in the instability of tracking performance.This thesis first uses traditional Mean-Shift tracking algorithm in different videosequences and analysis the advantages and disadvantages of the algorithm based onthe tracking results. To overcome the problem of kernel bandwidth can not being changed adaptively, two new solution methods which are based on Canny edgedetection and the comparison of Bhattacharyya coefficients are proposed in this thesis.The experiment results validate that the two methods can adapt to the situation ofsimple background. In addition, the latter method is also valid in the condition ofexisting interference in the background.The use of gray histogram as a target character descriptor to deal with thesituation of the similarity between the target and the background is not enough. wepresents a multi-feature fusion based Mean-Shift tracking algorithm. The LBPoperator is introduced to extract the texture feature of targets. The combination ofgrayscale and texture character method can improve the tracking accuracy andenhance the robustness.