STL_Siam: Real-time Visual Tracking based on reinforcement guided network

Abstract In recent years, deep visual tracking algorithms based on Siamese has made great breakthrough in both speed and accuracy. However, due to the dependence of Siamese network on the target template, these trackers are prone to drift or even fail to track in a complex tracking environment. In this work, we navigate the Siamese network with a STLNet model and propose the STL_Siam method. The STLNet, trained offline with the dataset enhanced by ROMIX method, is introduced into the SiamGrad algorithm to infer the movement of the target during online tracking, and guides the SiamGrad network to track. In order to evaluate the reliability of the proposed algorithm, we conducted experiments on the OTB2013 and VOT2016 two benchmarks. The algorithm achieved excellent performance, and the performance of the sequences with complex obstruction interference has been slightly improved. Experiments suggests that the proposed architecture reaches 40+fps and gets 0.865 precision on OTB2013, which is higher than results of the ACT. Meanwhile, compared with the baseline algorithm ACT, the A, R, and EAO of the proposed approach is increased by 5.6%, 1%, and 0.3% respectively on VOT2016 respectively.