An efficient Siamese triple-stream network with collaborative channel-spatial aggregation for RGBT tracking

As an effective approach for obtaining accurate target location in complex scenarios, RGBT tracking methods have recently attracted significant attention. However, the fundamentally different physical imaging mechanisms of visible and infrared modalities induce substantial discrepancies in appearance characteristics and feature distributions, causing semantic misalignment in feature representations and consequently challenging Siamese-based trackers that depend on a shared embedding space and direct similarity matching. To this end, we propose a novel triple-stream channel-spatial collaborative aggregation network for efficient RGBT tracking, named SiamCCA, which contains two parallel feature extraction streams, and one feature fusion stream. First, a dynamic gating scale awareness (DGSA) module is designed to adaptively adjust unimodal feature representations through dynamic gating mechanism and multi-scale adaptive fusion without significantly increasing computational overhead. Second, a channel-spatial collaborative aggregation (CSCA) module is constructed to accurately capture long-range channel and spatial dependencies, facilitating the model to better extract cross-modal information. Third, a region proposal selection (RPS) module is established to obtain accurate tracking results according to confidence scores of the fusion response map. Finally, comprehensive experiments have been demonstrated on three RGBT benchmark datasets. The results illustrate that SiamCCA can sufficiently handle different challenging scenarios while maintaining real-time processing at 56 FPS, outperforming other state-of-the-art trackers.The code is available at https://github.com/Mrxing-abc/SiamCCA/tree/master.