Evaluation of Multi-part Models for Mean-Shift Tracking

Abstract

Mean-shift tracking is a data-driven technique for tracking objects through a video sequence. We propose an innovation to mean-shift tracking that combines the background exclusion constraint with multi-part appearance models. The former constraint prevents the tracker from moving to regions where no foreground objects are present, while the multi-part nature of the models enforces a spatial structure on the tracked object. We also use a simple formula to determine the scale of the object in each video frame, and note the importance of setting an appropriate convergence condition. An evaluation of our proposed tracker and several existing trackers is performed using a ground truth dataset. We demonstrate that our innovation yields more accurate tracking than existing mean-shift techniques.