BS-SiamRPN: Hyperspectral Video Tracking based on Band Selection and the Siamese Region Proposal Network

The spectral information of hyperspectral data is helpful to improve the performance of target tracker. Limited by the number of training samples, most video tracking methods based on deep learning cannot effectively utilize the semantic information of hyperspectral data. Therefore, this paper proposes a hyperspectral video target tracking method based on Band Selection and the Siamese Region Proposal Network (BS-SiamRPN). Firstly, an intelligent optimization algorithm is utilized to determine three bands with the largest joint entropy. Based on limited hyperspectral information, Transfer Learning (TL) is performed on the pre-trained siamese networks to obtain a good parameter. Here, index. Then, the dataset in Generic Object Tracking (GOT) benchmark is utilized to retrain the SiamRPN method. Finally, the information of three specific bands referring to hyperspectral videos is regarded as the input of the migration network, which can obtain matching results. Experimental results on public hyperspectral data show that the proposed siamese network based tracking method effectively utilize spectral features, and do well in visual effects and objective evaluation.