Tracking as Segmentation of Spatial-Temporal Volumes by Anisotropic Weighted TV

Tracking is usually interpreted as finding an object in single consecutive frames. Regularization is done by enforcing temporal smoothness of appearance, shape and motion. We propose a tracker, by interpreting the task of tracking as segmentation of a volume in 3D. Inherently temporal and spatial regularization is unified in a single regularization term. Segmentation is done by a variational approach using anisotropic weighted Total Variation (TV) regularization. The proposed convex energy is solved globally optimal by a fast primal-dual algorithm. Any image feature can be used in the segmentation cue of the proposed Mumford-Shah like data term. As a proof of concept we show experiments using a simple color-based appearance model. As demonstrated in the experiments, our tracking approach is able to handle large variations in shape and size, as well as partial and complete occlusions.