SiamDL: Siamese Dual-Level Fusion Attention Network for RGBT Tracking

Abstract: Most of the existing dual-mode tracking algorithms rely on feature fusion design.We propose a Siamese Dual-level Fusion Attention Network (SiamDL) for RGBT Tracking by combining dual-level balance module and multi-domain aware module. Dual-level balance module (DLBM) introduces a new dual-level fusion attention mechanism to utilize the two modal information at decision-level and feature-level, which is used to provide a more reasonable way to balance the two modal features’ weight ratio. Multi-domain aware module (MDAM) introduces a new cross domain siamese attention mechanism to make mode-domain (referring to visible and infrared modal branches) and time-domain (referring to template and search branches) information interact with each other, which is used to enhance feature expression ability of network and adaptively update template features. The average tracking speed on rtx3060ti is about 45fps, which suggests that SiamDL has achieved state-of-the-art performance by carrying out experiments on three RGB-T tracking benchmark datasets.