Visual Tracking via Supervised Similarity Matching

Supervised learning algorithms have been widely applied in tracking-by-detection based methods for object tracking in recent years. Most of these approaches treat tracking as a classification problem and solve it by training a discriminative classifier and exhaustively evaluating every possible target position; problems thus exist for two reasons. First, since the classifier describes the common feature of samples in an implicit way, it is not clear how well the classifier can represent the feature of the desired object against others; second, the brute-force search within the output space is usually time consuming, and thus limits the competence for real-time application. In this paper, we treat object tracking as a problem of similarity matching for streaming data. We propose to apply unsupervised learning by Locality Sensitive Hashing (LSH) and use LSH based similarity matching as the main engine for target detection. In addition, our method applies a Support Vector Machine (SVM) based supervised classifier cooperating with the unsupervised detector. Both the proposed tracker and several selected trackers are tested on some well accepted challenging videos; and the experimental results demonstrate that the proposed tracker outperforms the selected other trackers in terms of the effectiveness as well as the robustness.