Adaptive covariance tracking with clustering-based model update

We propose a novel approach to track nonrigid objects using the recently proposed adaptive covariance descriptor [1] with clustering-based model update mechanism. The adaptive covariance descriptor represents an object of interest according to its characteristics in a small-dimensional covariance matrix and possesses higher discriminative power with respect to the original covariance descriptor. A clustering-based update mechanism is then conducted on the target model to adapt to the object appearance changes during the tracking process. We show that by updating with a carefully selected cluster, the update mechanism can efficiently deal with significant appearance deformations and severe occlusions. Comparative experimental results on challenging video sequences demonstrate the effectiveness of the proposed approach.