Trajectory Smoothing Constraint and Hard Negative Mining for Distractor-Aware Regression Tracking

Recently, convolutional regression networks have drawn great attention in the tracking community. Convolutional regression trackers formulate the regression network as one convolutional layer and take advantages of end-to-end learning. However, existing convolutional regression trackers regress the input feature to Gaussian-like soft labels, which still assign a large label to semantic backgrounds in online model fine-tuning. As a result, in the presence of background distractors, convolutional regression trackers tend to drift toward regions, which exhibit a similar appearance compared to the object of interest. In this paper, we propose to achieve distractor-aware regression tracking with trajectory smoothing constraint and hard negative mining. The trajectory smoothing constraint measures motion distance and motion direction between the adjacent frames to discard distractors that fail to meet trajectory smoothness. On this basis, distractors are fed into convolutional regression networks as hard negative samples, which boosts the discriminability of convolutional regression trackers against background distractors. The experimental results on the OTB100 and VOT2016 benchmarks show that the proposed algorithm performs favorably in terms of both accuracy and robustness when compared with state-of-the-art trackers.