A Stochastic Resampling Based Selective Particle Filter for Robust Visual Object Tracking

In this work, a new variant of particle filter has been proposed. In visual object tracking, particle filters have been used popularly because they are compatible with system non-linearity and non-Gaussian posterior distribution. But the main problem in particle filtering is sample degeneracy. To solve this problem, a new variant of particle filter has been proposed. The resampling algorithm used in this proposed particle filter is derived by combining systematic resampling, which is commonly used in SIR-PF (Sampling Importance Resampling Particle Filter) and a modified bat algorithm; this resampling algorithm reduces sample degeneracy as well as sample impoverishments. The measurement model is modified to handle clutter in presence of varying background. A new motion dynamics model is proposed which further reduces the chance of sample degeneracy among the particles by adaptively shifting mean of the process noise. To deal with illumination fluctuation and object deformation in presence of complete occlusion, a template update algorithm has also been proposed. This template update algorithm can update template even when the difference in the spread of the color-histogram is especially large over time. The proposed tracker has been tested against many challenging conditions and found to be robust against clutter, illumination change, scale change, fast object movement, motion blur, and complete occlusion; it has been found that the proposed algorithm outperforms the SIR-PF (Sampling Importance Resampling Particle Filter),\ud bat algorithm and some other state-of-the-art tracking algorithms