Mean Shift And Particle Filter-based Target Tracking Algorithm

Tracking visual objects through image frames has been a hot topic in computer vision field and is widely applied to surveillance, robotics, human machine interface, object based video coding, etc. However, the task of robust tracking is a focus concerned by scholars that regarding fast motion, occlusion, object deformation, illumination variation, background clutters, real-time restriction, etc. It is also a difficult problem must be solved in practice using currently.The author has deeply researched the algorithms of mean shift and particle filter among many object tracking algorithms. The algorithm of mean shift is a non-parametric method based upon climbing gradient. It searches object by iteration to realize object tracking. The obvious merit of mean shift is lesser amount of calculation, simple and easy to realize, so it can meet the need of real-time tracking. But it fails in tracking small and fast moving targets and in recovering a track after a total occlusion. In second chapter, the algorithm of mean shift theory and its using in object tracking are deduced and described in detail. An algorithm of kernel histogram to calculate the distribution of target is proposed through a large number of experiments. The experiments results also show the algorithm is real-time. The analysis of drawbacks of the algorithm of mean shift is in the experiments of the third and fourth chapter.Another algorithm have attracted much attention is particle filter, due to its robust tracking performance in cluttered environments. The particle filter is used to apply a recursive Bayesian filter based on the propagation of sample set over time, maintain multiple hypotheses at the same time and use a stochastic motion model to predict the position of the object. Maintaining multiple hypotheses allows the tracker to handle clutter in the background, and recover from failure or temporary distraction. It has been proved to be a robust method of tracking in non-linear and non-Gaussian case. However, two common problems of the particle filter technique are the degeneracy phenomenon and the huge computational cost. Thus, those problems constitute a bottleneck to the application of particle filter to real-time tracking systems. The algorithm of particle filter theory and its using in object tracking are discussed in detail in third chapter. The method of target template update is put forward. The experiments results show that the algorithm of particle filter is much more robust and much better performance of resisting occlusion and disturbance compare with the algorithm of mean shift. But its computational cost is very huge.In this paper, the author combines the merits and drawbacks of two algorithms and proposes a new idea that is an algorithm uses the mean shift algorithm inside the particle filter. With the help of the mean shift algorithm, we can sample more particles of higher weights, and discard those particles whose contribution to the tracking is almost zero. At the mean time, it reduces the mount of samples that represent the states of the object. The experiments results show the new algorithm reduces the degeneracy problem and the computational cost of the particle filter. In the experiment of tracking one same object, this tracking algorithm not only maintain the high robustness and better performance of resisting occlusion and disturbance of the algorithm of particle filter, but also the computational cost is less than one third of the algorithm of particle filter's. The real-time characteristic of this algorithm increases dramatically.