Appearance-based subspace learning model using incremental PCA in object tracking

Visual tracking is still a challenging subject due to the targeted object´s change in direction and size, stochastic disturbance, and drastic lighting change under complicated scene. Based on the subspace´ updating and real-time learning, a visual tracking framework is proposed in the work. The Incremental PCA algorithm and the new measurement on subspace´s similarity in computing particles´ weights are introduced in our tracking processes under Condensation algorithm. Not based on the trained database in advance, our method updates the subspace about the moving target by continuously discarding the old frame and adopting the new one. Differed from conventional PCA method, the Incremental PCA method adaptively updates the subspace which can reflect appearance variation of the moving target over long period of time. Compared with Condensation algorithm using color histogram, the tracker proposed in this paper can effectively track the target under complicated surrounding and it is being incrementally updated with new frames. Challenging experimentations on standard testing videos demonstrate the proposed tracker´s effectiveness and accurateness in actual tracking processes.