Real-time Object Tracking with Multi-feature Particle Filter

A new multi-feature Particle Filter is presented in this paper to improve sampling efficiency and performance of tracker in real-time object tracking. First, the edge information of the target is extracted using Sobel algorithm in order to obtain the profile feature. Secondly, we find that the most rigid targets such as vehicle, tank, airplane, etc always move along its own profile. So, we combine the profile cue and motion cue of the target to sample the particles instead of generating particles randomly in conventional particle filter. A new Local Optimal Searching algorithm is adopted in the particles' initialization procedure, so we call this particle set the Local Optimal Particle Set. Then we only apply these "high quality" particles to participate into the tracking algorithm. Because of the number of the particles in our algorithm is not fixed, usually greatly smaller than conventional particle filter, the computational has been decreased dramatically. Then, we use the weighted gray histogram feature as the state model to describe the target. By combining the profile cue and motion cue properly, the computation is put on the more effective particles reasonably, and the utilization ratio of particles is increased. Because number of more-contribution particle is increased and the less-contribution one is decreased, so the computation of the algorithm is reduced dramatically, and the degeneracy and the impoverishment phenomenon are overcome. At the same time, the multi-hypothesis of the particle filter is still maintained. Experimental results on tracking visual targets in long video sequences show the wonderful tracking performance such as strong anti-occlusion and anti-disturb using the proposed particle filter, and the computing time has been reduced within 25ms per frame. It makes the computational load no more the bottleneck issues which prevent the Particle Filter from using into the actual application.