Tracking-recommendation-detection: A novel online target modeling for visual tracking

In this paper, we propose a novel object tracking algorithm which combines a tracking-by-detection approach with a recommender system. The proposed algorithm consists of two steps: (1) to match detected features with a three-layer feature dictionary, i.e., target appearance model, which is initially built on the first frame and being adaptive frame-by-frame; (2) to add detected features into the aforementioned dictionary and use a recommender system to optimize the dictionary simultaneously. Among the two steps, while the former determines a new location for the target, the latter updates the target appearance model. Extensive experiments, which are done on 20 benchmark video sequences, demonstrate that the tracking results for the proposed novel method outperforms the state-of-the-art trackers both in terms of accuracy and robustness.