Hyperspectral Video Target Tracking Based on TensorSSA and Deep Features
Hongmin Lou, Shenghui Rong, Haoran Guo, Jiankang Ma · 2023
Most target tracking is performed on visible video. This conventional RGB video lacks spectral information, leading to the tracking process being susceptible to interference from background clutter, illumination changes, etc. Hyperspectral images (HSIs) have the advantage of spectral integration, which can interpret the features of a target in both spatial and spectral dimensions. Therefore, HSIs with continuous spectral information have unique advantages in the field of target tracking. In this paper, a correlation filter-based HSIs tracking method is proposed. Feature maps are generated using tensor singular spectrum analysis (tensorS SA) feature and depth features to jointly exploit spatial contextual information and deep semantic information. Then a kernelized correlation filtering framework is used to generate feature responses on HSIs, and the largest region in the final output response is the center of the target in the current frame. Experiments conducted on hyperspectral images show that the proposed method can effectively utilize spatial-spectral information and outperforms the hyperspectral video tracking methods that partially use single features and the traditional color video tracking methods.