Object tracking using particle filter with back projection-based sampling on saliency

Abstract

The computation cost is always the big problem for particle filter because the number of samples and iterations until convergence. It is decreased by using back projection-based sampling method, which applied the concept of corresponding between 3D world space and 2D image plane. Size of search space is reduced by sampling the particles in 2D image plane then will be back projected to 3D world space. Although back projection-based sampling method can reduce the search space, the search space is extended larger and more samples are necessary if the objects appear far away from each other. This paper applied object detection algorithm as saliency segmentation using RGB-D information. It is used to obtain the object saliency before sampling the particles. The required number of samples is more decreased because the samples are not generated into the background boundary. In additional, the modified Augmented MCL is adapted to increase occasion of particles sampling around the target object region, which makes algorithm rapidly successful.