Robust tracking by accounting for hard negatives explicitly

In this paper, we present a method of robust tracking by accounting for hard negatives (i.e., distractors) of the tracking target explicitly. Our method extends the recently proposed Tracking-Learning-Detection (TLD) approach [7] in two aspects: (i) When learning the on-line fern detector, instead of using a set of features which are first randomly generated and then fixed throughout the tracking, we utilize a feature selection stage which constantly improves the performance of the detector, especially in tracking articulated objects (e.g., pedestrians); (ii) To address the diversity of distractors, instead of tracking a target against the whole set of collected negative examples, we account for the hard negatives explicitly, so that tracking drifts are largely prevented when multiple resembled targets appear in videos (e.g., people with white skirts and jeans). Experiments on a series of diverse videos show that our method outperforms TLD.