文摘
Appearance model in visual tracking is a key component to attain robustness and efficiency. In the last decades, many complex appearance models have been proposed to improve performance of tracking algorithm. However, these models are difficult to maintain accuracy and efficiency simultaneously. In this paper, we observe that data-dependent hashing method could improve processing speed by generating compact representation for the visual object. But applying the method to visual tracking is still a challenging task. To reinforce the performance of hashing technique, a novel hashing method called two dimensional ensemble hashing is proposed. In our tracker, image samples are hashed to binary matrices, and the Hamming distance is used to measure their confidences. Moreover, for adapting situation change, the hash functions are updated by the learning model at each frame. Experimental results not only demonstrate the accuracy and effectiveness of our tracker, but also show that the tracking algorithm outperforms other state-of-the-art trackers.