Mean shift based object tracking with accurate centroid estimation and adaptive Kernel bandwidth

The object tracking algorithms based on mean shift are good and efficient. But they have limitations like inaccuracy of target localization and sometimes complete tracking failure. These difficulties arises because of the fact that in basic kernel based mean shift tracking algorithm, the centroid is not always at the centre of the target and the size of tracking window remains constant even if there is a major change in the size of object. It causes introduction of large number of background pixels in the object model which give localization errors or complete tracking failure. To deal with these challenges a new robust tracking algorithm based on edge based centroid calculation and automatic kernel bandwidth selection is proposed in this paper. This approach includes relocation of the track window on the middle of the target object in every frame and automatic size adjustment of tracking window so that minimum background pixels will be introduced in object model. The proposed algorithm show good results for almost all the tracking challenges faced by basic mean shift kernel tracking method.