Robust object tracking by variable rate kernel particle filter

Abstract

Robust moving video object tracking under illumination variations, occlusion, object scale and appearance changes is a challenging problem. Bayesian filtering in particular particle filtering is conventionally used for nonlinear and non-Gaussian object state estimation problems because of its high performance. In this paper we extend the color based variable rate particle filter (VRCPF) existing in the literature by employing a kernel based filtering density function. The idea behind integrating a kernel into the model is it enables us to converge to the filtering density function smoothly resulting in improved object tracking accuracy. Video object tracking performance of the proposed filtering, K-VRCPF; has been tested on commonly used BoBoT and OTB datasets. Tracking accuracy reported in terms of center pixel error, and root mean square error (RMSE) demonstrate that, as a result of the regularized sampling of the posterior distribution, K-VRCPF with Gaussian kernels reduces the center pixel error and RMSE.