AWRUT: hyperspectral video tracker based on attention weights and response map union

In addressing the impact of background clutter on tracking performance in hyperspectral target tracking, a novel hyperspectral target tracking method, to our knowledge, based on union decision conditions is proposed. This method introduces a genetic function to reduce the dimensionality of the hyperspectral data, utilizes ViT to model the relationships between local and global features, and generates corresponding attention weight maps and response maps through a prediction head. Additionally, union decision conditions are proposed to evaluate the degree of background interference in the image. Based on this criterion, the response maps in the near-target and target regions are selectively reconstructed using sorting reconstruction to suppress interference regions. Unlike traditional trackers that directly use response maps for tracking, the specific advantage of the proposed algorithm lies in combining sorting and reembedding to generate a reconstructed response map, which can accurately reflect the position of the target area, suppress the interference of near-target pixels, and improve the ability to cope with background noise challenges. Experimental results demonstrate that the tracker based on attention weights and response map union outperforms state-of-the-art hyperspectral video trackers on existing datasets.