SATMtracker: Hyperspectral object tracking based on scale-adaptive tensorSSA feature and motion evaluator

Abstract

Hyperspectral videos contain rich spectral information, enabling precise material identification and effective object-background differentiation. However, object tracking in hyperspectral videos remains challenging in scenarios involving occlusion and scale variations. To achieve robust object tracking in hyperspectral videos, object trackers needs to be designed with multi-scale and occlusion handling capabilities. In this paper, we propose SATMtracker, a novel hyperspectral object tracker based on scale-adaptive tensor singular spectrum analysis (tensorSSA) and a motion evaluator, designed to enhance tracking accuracy in complex environments. Specifically, the tensorSSA feature with different scales is employed to improve the tracker’s robustness to scale variations. Additionally, deep convolutional features are leveraged to integrate semantic and spatial information and fused with the multi-scale tensorSSA features to enhance feature representation and tracking reliability. In scenarios like severe occlusion, the extracted features may not be always reliable. We further introduce the related confidence evaluator (RCE), which assesses feature reliability using a new metric, the related confidence level (RCL). If the feature is unreliable, the motion cues are used to handle these challenge scenarios. Experimental results demonstrate the effectiveness of our proposed SATMtracker, achieving more accurate tracking performance.