Robust RGB-T Object Detection via Quality-Aware Decision Fusion Based on Dempster–Shafer Theory

RGB-T object detection has achieved substantial progress by harnessing the complementary characteristics of diverse spectra for all-weather detection, particularly in critical security applications. However, most state-of-the-art models are based on a feature intermediate fusion structure, which may inadvertently create modality-dependent shortcuts. When a certain location of the input image in the shortcut branch is disturbed, these methods often fail to make accurate decisions, which limits their generalization ability in real-world application scenarios. In this paper, we aim to present a robust multispectral object decision fusion method (called QDFM) to perform effectively on both interferencefree datasets and those affected by sensor interference or failure data. We incorporate evidential deep learning into one-stage object detection and modify the loss function to mitigate the negative effects of class imbalance, allowing the network to naturally express uncertainty in its predictions. Subsequently, quality-aware Generalized D-S evidence theory is employed to reassign background confidence within the belief framework, effectively addressing the one-vote veto problem commonly encountered in traditional D-S combination rules. Experiments conducted on two publicly available datasets demonstrate the effectiveness of our approach in handling both sensor failure scenarios and interference-free data, showcasing significant improvements in robustness and accuracy. QDFM achieves a 10% mAP improvement compared to other advanced methods under modal failure conditions, while outperforming the baseline D-S method by over 2% across various interference scenarios.