An Anchor-Free Siamese Target Tracking Network for Hyperspectral Video

Hyperspectal target tracking is aimed at taking advantage of the spectral and spatial information in the target tracking. However, due to the limited training samples, the existing hyperspectral target trackers cannot exploit semantic information of hyperspectral image. In this paper, in order to solve this problem, we propose an anchor-free Siamese network for hyperspectral video target tracking (HA-Net). A spectral classification branch is introduced to the anchor-free Siamese network to increase the network's ability to identify objects. This branch exploits all the bands of the hyperspectral video for end-to-end training, to obtain more discriminative features. By fusing the classification response map of the spectral classification branch with the classification response map of the anchor-free Siamese network, the ability of the network to distinguish foreground and background can be enhanced. At the same time, the anchor-free tracking network can reduce the calculation time of the network. The experiments conducted on hyperspectral video showed that HA-Net can effectively exploit the spectral features and significantly improve the performance of the tracking network.