Research On Object Tracking Algorithm Based On Mean Shift

Video target tracking is one of the core technologies of the computer vision, itincorporates image processing, automatic control and artificial intelligence, etc, and hasimportant scientific theoretical significance and engineering application value in the fieldsuch as video surveillance, medical diagnosis, biological research, human-computerinteraction, autopilot and robot, etc. This thesis using Mean Shift (MS) as the core algorithmof target tracking, improving the accuracy, robustness and adaptability of target tracking asthe goal, aims to resolve the problems such as partial occlusion, deformation, illuminationvariation, scale variation and tracking rectangle with adjust attitude in video tracking.Traditional color histogram MS algorithm only considers color statistical informationof the object, and doesn’t contain space information, so when the object color closes to thebackground color, the traditional MS algorithm easily causes tracking inaccurately or lost.Aiming at this problem, this thesis proposes a novel MS tracking algorithm with adaptiveblock color histogram, which determines block method by the size of the lastest enclosingrectangle and determines their weight coefficient by the Bhattacharyya coefficient of allblocks. Among them, the adaptive block color histogram contains adaptive block methodand spatial information of the object, and the weighted Bhattacharyya coefficient considersthe influence of different blocks to the overall similarity. Research shows that the proposedmethod can dynamically adjust tracking rectangle scale of the horizontal and verticaldirections, adaptively determining the block method of the object, and has better trackperformance than the traditional MS algorithm and fixed block MS algorithm under somecases such as partial occlusion and deformation, etc.Considering the traditional MS algorithm is vulnerable to the influence of factors suchas illumination change, partial occlusion, image blurring, etc, this thesis introduces theimage local feature points. It presents a comparison of the most popular feature descriptorssuch as BRIEF, LAZY, ORB, RIFF, SIFT and SURF from the six aspects: rotationinvariance, scale invariance, illumination invariance, resisting vagueness, tracking accuracy and extraction time, combining the research content, and finally chooses SIFT as the featuredescriptor.Aiming at the problem of traditional MS algorithm is vulnerable to the influence offactors such as illumination change, partial occlusion, image blurring etc, the thesisproposes a novel object tracking algorithm based on improved MS and SIFT, which iscomposed of proposed MS (initial location) and SIFT feature extraction, matching andtracking (SIFT tracking). The former is the chapter III algorithm in the thesis, and on thebasis of the former tracking results, the latter utilizes SIFT tracking method to obtain SIFTtracking results. Finally, the algorithm utilizes linear weighted method to fuse the trackingresults of improved MS and SIFT tracking, obtaining the final tracking results. Researchshows that the proposed method not only keeps the advantages of chapter III algorithm, butalso further solves the tracking problems under the situation of partial occlusion,illumination change, image blurring, etc.In order to solve the problem which target tracking rectangle can adaptively adjustattitude, the thesis proposes a novel object tracking algorithm fused by MS and SIFT withaffine transformation. The former introduces affine transformation into the target candidatemodel, and translates complex motion tracking into the optimization problem of the costfunction. By calculating the first derivative of the cost function with respect to affineparameters and setting them to be zero, it can find the affine transformation parameters, andget MS affine transformation tracking results. On the basis of the former tracking results, thelatter utilizes the method of SIFT feature extraction, matching and tracking to obtain SIFTaffine tracking results. Finally, the algorithm utilizes linear weighted method to fuse thetracking results of MS and SIFT affine tracking, obtaining the final tracking results. Researchshows that the proposed algorithm extends the chapter III and V algorithms, and canadaptively adjust the attitude of the tracking rectangle.