Robust and Real-Time Visual Tracking with Triplet Convolutional Neural Network

In this paper, we propose a new visual object tracking which realizes robustness against object occlusion and deformation. In the proposed visual tracking, triplet convolutional neural network (triplet-CNN) structure is devised. The three inputs for the triplet-CNN come from current query frame, tracked object in a previous frame, and reference object. Object location in the query frame is predicted by fusing latent features from the three inputs. Moreover, predicted object is compared with reference object by using a Siamese CNN, so that object occlusion and deformation are detected and search range of tracking object is found adaptively. Comprehensive experimental results on a large-scale benchmark database showed that the proposed method outperformed state-of-the-art tracking methods in terms of precision and robustness with real-time tracking (about 25 fps).