Comparative Object Similarity Learning-Based Robust Visual Tracking

Tracking-by-detection for visual object tracking is the most popular and successful framework at present. It treats the tracking problem as a classification task and learns information about the target from each tracking result online. Accurate model learning of the classifier requires numerous positive samples. However, it is difficult to obtain numerous positive training samples at the beginning of visual tracking. In this paper, we propose a novel comparative object similarity learning method to strengthen the training samples set. The core of our approach is that the comparative object similarity information between the candidate objects is taken into account when training classifiers. In addition, the classifier model is updated with the image information of the target to be predicted by further exploring the temporal context between successive image frames. According to the Bayesian inference theorem, the tracking results, which are estimated from the posterior probability distribution of target, are more accurate. We implement two versions of the proposed tracker with the representations from both conventional hand-crafted and deep convolution neural networks-based features to validate the effectiveness of the algorithm. The quantitative and qualitative experimental results demonstrate that the proposed method performs superiorly against several state-of-the-art algorithms on large-scale challenging benchmark datasets.