Research On The Application Of Video Object Tracking Algorithm Based On Meanshift And Particle Filter

As the core topic of the research of computer vision field, detection and tracking ofmoving video objects, is the premise and foundation of intelligent video processingsystem. In actual tracking applications, the randomness of the target types, the complexof the tracking environment, the diversity of system’s functional requirements, all thesefactors making most mature algorithms can only get better tracking results under therange of constraint conditions and specific application scenarios. Therefore, studying anefficient robust video target tracking algorithm which has less restrictions will havepractical significance. This article has an in-depth research on moving target detectionalgorithm, MeanShift correlation tracking algorithm and particle filter tracking algorithm,and summarizes the respective existing problems and proposes the correspondingimprovements. Finally, on the basis of previous research, it proposes an integrated videotracking algorithm with fusion of MeanShift and particle filter, and verify theeffectiveness of the algorithm through a set of comparative experiments.The research of moving target detection algorithm is mainly directed against theproblems of traditional Gaussian mixture model background subtraction method towardsbeing sensitive on the environment mutant and losing information for slowly movingtarget, and it also proposes an improved adaptive target detection method based onGaussian mixture model. First, classify the pixel values before the parameter updating,set model update rates according to the classification results to restrain the slowingmoving foreground being trained into background. Introduce a metrics factor to track theenvironmental changes, make adaptive switching of background subtraction and framedifference when environment mutates, and filter the interference of environmentmutation. Finally, get the accurate moving target through ecological filter. Theexperiments shows that this algorithm has better disturbance resistance against light andother environmental changes, and has better solution of the problems of targets beingundetected because of slow motion or a short stay.The research of MeanShift algorithms mainly analyses the two commoncharacteristics of MeanShift algorithms based on kernel function histogram andprobability distribution. On this basis, since traditional probability distribution algorithmhas robustness difference problem to background color interference, it proposes animproved Camshift target tracking algorithm based on target detection. To each frame, ifirst use the target detection algorithm in this article to extract the objects, then make therectangular area from target detection as an initialization search window, and constantlyrevise the center of the tracking window through the calculation of the window areacentroid. The experiment shows that this algorithm can make a stable tracking on objectswhen it has similar elements in the target color in the scene’s background.The study of the particle filter has an systematic introduction and analysis of particlefilter theory and its implementation methods on target tracking. Towards traditionalparticle filter describes object in its single characteristics and has poor trackingrobustness, this article proposes a multi-observation model particle filter algorithm basedon target color and edge direction features. First, extract the features of the target color and edge direction, set their own observation model through the Bhattacharyya distanceand Euclidean distance. Then, the the multiplicative strategy based on the fusioncharacteristics of particle evaluation criteria as systematic observation model. Finally, bytracking experiment to verify the effectiveness of the algorithm. Then, use the particleevaluation criteria based on fusion characteristics as the system’s observation modelthrough multiplicative strategy. Finally, verify the effectiveness of the algorithm throughtracking experiments.On the basis of in-depth analysis of the advantages and disadvantages of particlefilter and MeanShift algorithm, and because a single algorithm is difficult to meet thereal-time tracking and tracking reliability at same time, so this article proposes a targettracking algorithm(MSPF) based on Meanshift particle filter by an effective combinationof those two algorithm. First, set target description models through color and movementinformation, improve traditional MS algorithm and make it embedded into the particlefiltering process, make the status particles move to the maximum of Posteriorprobability density direction to improve the algorithm efficiency.