Dictionary Learning and Confidence Map Estimation-Based Tracker for Robot-Assisted Therapy System

In this paper, we propose a new tracker based on dictionary learning and confidence map estimation for a robot-assisted therapy system. We first over-segment the image into superpixel patches, and then employ color and depth cues to estimate the object confidence of each superpixel patch. We build two Bag-of-Word (BoW) models from initial frames to encode foreground/background appearance, and compute object confidence at superpixel level using BoW model in both foreground and background. We further refine target confidence by depth-based statistical features to mitigate noise interference and the uncertainty of visual cues. We derive the global confidence of each target candidate at bag level, and incorporate the confidence estimations to determine the posterior probability of each candidate within the Bayesian framework. Experimental results demonstrate the superior performance of the proposed method, especially in long-term tracking and occlusion handling.