Real-time active visual tracking with level sets

This paper presents a new real-time active visual tracker which improves standard mean shift tracking by using level sets to extract contours from the target. We use colour and the disparity map computed from a stereo camera pair which prove to be powerful features for tracking in an indoor surveillance scenario. To combine the features in the level sets process, we enhance Chen´s et al appearance model of [5] by using a probabilistic model determined via Expectation-Maximization (EM) clustering. The level set result is used as the weighting kernel which improves the accuracy of the similarity measurement in the mean shift method. Finally a Kalman filter deals with complete occlusions.