Visual Tracking by Hypothesis Testing

Abstract

A new approach for tracking a non-rigid target is presented. Tracking is formulated as a Maximum A Posteriori (MAP) segmentation problem where each pixel is assigned a binary label indicating whether it belongs to the target or not. The label field is modeled as a Markov Random Field whose Gibbs energy comprises three terms. The first term quantifies the error in matching the object model with the object’s appearance as given by the current segmentation. Coping with the deformations of the target while avoiding optical flow computation is achieved by marginalizing this likelihood over all possible motions per pixel. The second term penalizes the lack of continuity in the labels of the neighbor pixels, thereby encouraging the formation of a smoothly shaped object mask, without holes. Finally, for the sake of increasing robustness, the third term constrains the object mask to assume an elliptic shape model with unknown parameters. MAP optimization is performed iteratively, alternating between estimating the shape parameters and recomputing the segmentation using updated parameters. The latter is accomplished by discriminating each pixel via a simple hypothesis test. We demonstrate the efficiency of our approach on synthetic and real video sequences.