Bayesian Covariance Tracking with Adaptive Feature Selection

Effective appearance models are one important factor for robust object tracking. In this paper, a more elaborate object representation model via a simultaneous online feature selection and feature fusion algorithm is proposed, in which extended variance ratio is used to select the most discriminative power features, and thereby account for appearance model using region covariance descriptor which takes into account feature correlation information during tracking. Fusing all selected features, we get a more discriminative appearance model. Furthermore, our simultaneous online feature selection and feature fusion method is integrated into particle filter framework for robust tracking. Experimental results show that this proposed method is robust in heavy occlusions scenes and is able to handle variations in illumination and scale.