Local random Mean-hash classifier based adaptive visual tracking

Target visual tracking has been a central task in computer vision applications, meanwhile has also been facing with a variety of difficulties in tracking with target varying state and complex scenes. In this paper, a novel local Mean-hash classifier based tracking algorithm is presented. The local Mean-hash classifier is defined using the Poisson distribution based model as the target reference model, which is the main novelty in this research. The multi-instance learning is achieved using the detected target as the on-line training sample, and is combined with the current target reference model for model updating purposes. The updated target model is used for target detection and adaptive tracking in the next frame. The proposed scheme is evaluated by comparing with 6 other relatively newly published and commonly used tracking algorithms on a given number of standard video sequences under various complex scenes. The results are provided indicating that the proposed approach can adaptively track the target with good robustness in complex circumstances.