Anchor-Free Ranking-Based Localization Optimized Siamese RGB-T Object Tracking Network

Abstract

The fast tracking speed of Siamese-based RGB-T tracking has garnered significant attention. However, current Siamese-based RGB-T trackers still face certain limitations, including insufficient bounding box estimation, neglecting the interaction between positive and negative samples, and the complexity associated with adjusting anchors hyperparameters. To address these challenges, a novel anchor-free ranking-based localization optimized Siamese RGB-T tracking network is proposed. Firstly, to enhance bounding box estimation accuracy, a boundary localization optimization module (BLOM) is proposed to optimize the bounding boxes of the visible and thermal infrared (TIR) branches. Subsequently, a fully convolutional anchor-free RGB-T tracking network is proposed to avoid tricky hyperparameter tuning of anchors and to minimize human intervention. Finally, to improve the collaborative learning between positive and negative samples and to eliminate any mismatch between classification and localization, the ranking-based optimization algorithms (RAR) is devised. Extensive experiments conducted on two RGB-T benchmark datasets, which conclusively validates the effectiveness of our proposed network, showcasing the outstanding performance of the proposed tracker.