多维空间感知和全局特征调制的鲁棒RGBT目标跟踪OA

To address the susceptibility of RGBT object tracking to extreme lighting conditions and cluttered backgrounds in complex scenes,alongside existing methods'inadequate modelling of shared structural information across modalities,this paper proposes an RGBT object tracking approach based on multi-dimensional spatial perception and global feature modulation(multi-dimensional spatial perception network,MSPNet).Firstly,within a dual-stream multi-domain convolu-tional neural network framework,this paper constructs a multi-directional scene awareness(MSA)module.By introduc-ing direction-aware convolutions along horizontal,vertical,and diagonal orientations,this module explicitly models cross-directional spatial dependencies between RGB and infrared features.This overcomes the limitations of traditional convolu-tions in directional modelling,enhancing the interaction and alignment of structural information across modalities.Sec-ondly,a dynamic channel aggregation(DCA)module is designed.Utilizing a branched architecture and decomposable convolutions,it separately models and adaptively fuses channels carrying low-frequency structural information and high-frequency detail information within the dual-modal features.This enhances the discriminative power of multi-scale fea-ture representations.Finally,a global feature modulation(GFM)mechanism is established.This employs adaptive horizon-tal and vertical pooling to obtain direction-aware global contextual representations.Combined with large-kernel striped convolutions,it generates spatial modulation weights that effectively suppress background interference features during the fusion stage,thereby improving foreground object localization accuracy.Experimental results on the GTOT and RGBT234 benchmark datasets demonstrate that MSPNet achieves PR/SR of 90.4%/73.9%and 80.0%/55.8%respectively,exhibiting superior robustness and tracking accuracy compared with existing state-of-the-art trackers.