In this paper, we propose a novel tracking framework based on a sparse and discriminative hashing method. Different from the previous work, we treat object tracking as an approximate nearest neighbor searching process in a binary space. Using the hash functions, the target templates and the candidates can be projected into the Hamming space, facilitating the distance calculation and tracking efficiency. First, we integrate both the inter-class and intra-class information to train multiple hash functions for better classification, while most classifiers in previous tracking methods usually neglect the inter-class correlation, which may cause the inaccuracy. Then, we introduce sparsity into the hash coefficient vectors for dynamic feature selection, which is crucial to select the discriminative and stable features to adapt to visual variations during the tracking process. Extensive experiments on various challenging sequences show that the proposed algorithm performs favorably against the state-of-the-art methods.

In this paper, we propose a novel tracking framework based on discriminative supervised hashing algorithm. Different from previous methods, we treat tracking as a problem of object matching in a binary space. Using the hash functions, all target templates and candidates are mapped into compact binary codes, with which the target matching is conducted effectively. To be specific, we make full use of the label information to assign a compact and discriminative binary code for each sample. And to deal with out-of-sample case, multiple hash functions are trained to describe the learned binary codes, and group sparsity is introduced to the hash projection matrix to select the representative and discriminative features dynamically, which is crucial for the tracker to adapt to target appearance variations. The whole training problem is formulated as an optimization function where the hash codes and hash function are learned jointly. Extensive experiments on various challenging image sequences demonstrate the effectiveness and robustness of the proposed tracker.

No abstract

No abstract