Moving Object Tracking Based On Particle Filter

Object tracking is one of the active areas in computer vision research field. With the development of software and hardware, object tracking has been applied in many fields and the researches of related algorithms have deepened in recent years. As an algorithm with multi-hypothesis estimation based on Bayesian estimation, particle filter has been predominant on tracking nonlinear moving target.The feature of stochastic of particle filter avoids it trap into local optimization, but plenty of samples must be needed to preserve the precision of samples estimation, which leads to high computation. In this paper, we embed mean-shift into the particle filter framework and proposed a new algorithm, which based mean-shift and particle filter, named The K-means and Mean-Shift Embedded Particle Filter (KMSEPF).As particles with adjacent state will move to the same fitness peak, performing deterministic search for each particle is computational costly and unnecessary. In order to overcome the disadvantage mentioned above, the KMSEPF algorithm improves the general mixture algorithms which are based on particle filter and Mean-shift. Firstly it clusters the multi-hypothesis using K-means, and then the Mean-shift determination search is carried out on cluster centers. By those processes, the tracker can find the peak of posterior estimation without repeated process on the samples with similar feature, which leads to less computation. Experiments show that the new algorithm reduces the computation complexity, while maintains the high precision and the ability to control the occlusion, comparing with the MSEPF algorithm.Another adaptive particle filter algorithm is also introduced, which includes an adaptive online appearance model and an adaptive motion model with high ability to deal with the changing appearance and occlusion problem; and in this algorithm the samples number can be change based on the scale of the tracking error, which improves the efficiency of particles and reduces the computation cost.