Dual-Force Metric Learning for Robust Distracter-Resistant Tracker

摘要

In this paper, we propose a robust distracter-resistant tracking approach by learning a discriminative metric that adaptively learns the importance of features on-the-fly. The proposed metric is elaborately designed for the tracking problem by forming a margin objective function which systematically includes distance margin maximization and reconstruction error constraint that acts as a force to push distracters away from the positive space and into the negative space. Due to the variety of negative samples in the tracking problem, we specifically introduce the similarity propagation technique that gives distracters a second force from the negative space. Consequently, the discriminative metric obtained helps to preserve the most discriminative information to separ rate the target from distracters while ensuring the stability of the optimal metric. We seamlessly combine it with the popular L1 minimization tracker. Our tracker is therefore not only resistant to distracters, but also inherits the merit of occlusion robustness from the L1 tracker. Quantitative comparisons with several state-of-the-art algorithms have been conducted in many challenging video sequences. The results show that our method resists distracters excellently and achieves superior performance.