Target-aware and spatial-spectral discriminant feature joint correlation filters for hyperspectral video object tracking

Abstract

Visual tracking has been considered a promising task in computer vision. Most existing trackers construct tracking frameworks based on color video which provides information in limit visible spectrums, while hyperspectral video gives more material-based information for targets and distractors in background. Although hyperspectral video contains abundant spectral information, high-dimensional data brings negative influence for visual tracking due to redundant information. To exploit the intrinsic characteristics in hyperspectral video, a novel hyperspectral video-based tracking algorithm is proposed in this paper. A target-aware band selection (TABS) method is designed to select discriminative information which is beneficial to distinguish a target from complex background. To take advantage of the spatial–spectral relationship in hyperspectral video, an adaptive spatial–spectral discriminant analysis method (ASSDA) is designed to embed high-dimensional hyperspectral data into low-dimensional space. In the tracking process, two false-color video branches generated from TABS and ASSDA are put into correlation filters-based tracker, respectively. After that, the output responses of two branches are combined to obtain a joint estimation in hyperspectral video. Extensive experimental results illustrate the effectiveness of our method compared with those state-of-the-art color and hyperspectral trackers.