SiamTU: Hyperspectral video tracking against appearance changes using improved SiameGAT with adaptive template updating

Abstract

Siamese-based tracking algorithms are gaining popularity in hyperspectral tracking because of their effectiveness in feature matching. Nevertheless, these techniques require refinement to handle significant variations in the target’s appearance, including deformation, occlusion, and small size, more effectively. Therefore, we propose the SiamTU (SiameGAT with Template Updating) tracker, which includes band selection, an improved graph attention based Siamese model, and an adaptive template update mechanism, using public hyperspectral videos. Initially, a band selection technique employing the neighborhood group normalized filter is utilized to obtain three high-significance bands. The resultant synthesized image is then used as the input for the tracker. Following this, a module for efficient feature refinement is developed to improve the features that have been encoded. The model has the ability to simultaneously concentrate on both local details and overall structures by extracting features from the input at various scales. Moreover, an embedded template update network enhances the SiamTU’s capability referring to appearance changes effectively. Results from tests on the hyperspectral dataset reveal that SiamTU is more effective than comparable algorithms, attaining a success rate of 0.617 and a precision value of 0.939. The code will be accessible at https://github.com/ctb2/SiamTU.