Fast and Efficient Target Appearance Variations Consistency Tracking

Long Xu,Ying Wei,Zhiyuan Yao

DOI: https://doi.org/10.2139/ssrn.4016128

2022-01-01

Abstract:In the single target tracking field, big target appearance variations usually lead the tracking fail. To address this issue, we propose an appearance variations consistency tracking algorithm based on SiamRPN++. It brings considerable improvements in target accurate localization. In the proposed algorithm, the target appearance variation information is captured by a conditional spatial transformer neural network. Compared with the baseline algorithm, the proposed algorithm brings better performance in addressing the target appearance variations issue. To improve the tracker’s robustness of multi-scale targets, a feature pyramid fusion module is adopted, which also brings additional baseline improvements. In addition, to make the proposed algorithm more generalization, the tracker training process is transformed into a multi-task learning problem, including classification, bounding box regression, and feature learning tasks. In this framework, different task losses will be used to calculate their corresponding gradients, and the major updating direction is determined by the classification and regression losses. Finally, the proposed algorithm is evaluated on some prevalent tracking benchmarks including GOT-10k, LaSOT, VOT2018, etc. Compared with the baseline, the proposed algorithm improves the tracking performance a lot in these benchmarks. The proposed algorithm also demonstrates competitive performance when compared with the state-of-the-art algorithms.