Visual Face Tracking: a Coarse-to-Fine Target State Estimation

摘要

Keypoint-based methods are used in visual tracking applications. These methods often model the target as a collection of keypoint descriptors. Target localization on subsequent frames is thus a complex task that involves de-tecting keypoints, computing descriptors, matching features, and checking match consistency to update the reference model adequately and avoid tracker drifts. This work aims to boost keypoint tracking efficiency while reducing complexity by a coarse-to-fine state estimation to track human faces. In this context, we present a novel face tracking algorithm combining color distribution and keypoints to model the target. Our tracking strategy is based on a color model to predict a coarse state where the target search should be performed using keypoints. The fine estimation of the target state is then made by matching candidate keypoints with those of a reference appearance model that evolves during the tracking procedure. Qualitative and quantitative evaluations conducted on a number of challenging video clips demonstrate the validity of the proposed method and its competitiveness with state of the art trackers.