Robust Object Tracking by Particle Filter with Scale Invariant Features

Abstract

Few literatures employ SIFT (scale-invariant feature transform) for tracking because it is time-consuming. However, we found that SIFT can be adapted to real-time tracking by employing it on a subarea of the whole image. In this paper the particle filter based method exploits SIFT features to handle challenging scenarios such as partial occlusions, scale variations and moderate deformations. As proposed in our method, not a brute-force feature extraction in the whole image, we firstly extract SIFT keypoints in the object search region only for once, through matching SIFT features between object search region and object template, the number of matched keypoints is obtained, which is utilized to compute the particle weights. Finally, we can obtain an optimal estimate to object location by the particle filter framework. Comparative experiments with quantitative evaluations are provided, which indicate that the proposed method is both robust and faster.