SVTN: Siamese Visual Tracking Networks With Spatially Constrained Correlation Filter and Saliency Prior Context Model

Abstract

Recently, Siamese network based trackers have been greatly developed and achieved state-of-the-art performance on multiple benchmarks. However, the decision-making mechanism needs to be studied more deeply in order to obtain higher accuracy. In this paper, we propose a novel Siamese network based visual tracking method, which enhances decision-making ability by Spatially Constrained Correlation Filter (SCCF) and Saliency Prior Context (SPC) model. We use the deep features extracted from Siamese networks to train the SCCF via the efficient Alternating Direction Method of Multipliers (ADMM), and our SCCF applies a penalizing matrix to suppress the boundary effect well. Meanwhile, we regard the end-to-end output of Siamese networks as a priori probability and utilize the spatio-temporal relationship to establish the SPC model. The SPC model can handle the various cases of feature distributions generated from different targets and their contexts. Further, we also take measures to solve some challenging problems in visual tracking, such as target scale change and target occlusion. We conduct extensive experiments to demonstrate the effectiveness of the proposed method, which obtains currently the best results on three large tracking benchmarks, including OTB-2013, OTB-2015, and VOT-2016.