PD-Shift: Patch Detector Shift based Tracker

We identify two important directions of algorithm development in the tracking literature - first, the older and more established mean-shift based approaches; and second, more recent formulations of viewing tracking as detection by parts (aka tracking-learning-detecting). Introductory versions of both these approaches had their own weaknesses due to the main challenges of tracking like illumination variations, abrupt motions, deformations/articulations and occlusions. Both these approaches have developed independently with a number of follow-ups in the literature and even the matured versions were never integrated. We propose to integrate these two different viewpoints of single object tracking where the contribution lies in the construction of the target model as a collection of object part/patch detectors integrated with a kernel. Similarly, the kernel weighted patch detector responses form the candidate model. The Bhattacharyya coefficient based similarity measure between the target and candidate model is maximized in a gradient ascent framework to deduce a mean-shift like centroid update rule to iteratively track the target in the subsequent frames. We call this unified one the Patch Detector Shift based Tracker (PD-Shift). Experimental results of PD-Shift based tracking performance analysis is presented for the BoBoT dataset.