MHP-DETR: A Vehicle Detection System Based on Multimodal Attention and Hypergraph Fusion for Complex Weather Conditions

Vision sensor-based vehicle detection systems face significant challenges when operating under adverse weather conditions in intelligent transportation systems (ITSs) and autonomous driving applications. Traditional sensor data-processing methods demonstrate limited capabilities in handling degraded visual information affected by meteorological factors such as rain, fog, and snow. To address the limitations of conventional approaches regarding feature representation capability, cross-scale feature fusion, and environmental robustness, an end-to-end vehicle detection sensor data-processing framework called MHP-DETR is presented. The framework incorporates three collaborative sensor data enhancement modules: the multiscale synergistic attention block (MSAB), which adaptively processes multidimensional features across channel, pixel, and spatial dimensions; the hypergraph-based multiscale feature fusion network (HMF-Net), which constructs high-order feature association representations in non-Euclidean space, overcoming the limitation of traditional methods that focus solely on local associations; and the polar-frequency-enhanced encoder layer (PFEL), which optimizes feature representation simultaneously in both temporal and frequency domains through polar decomposition linear attention and frequency modulation feed-forward networks (FMFFNs). Comprehensive experiments on complex weather vision sensor datasets demonstrate that MHP-DETR achieves significant performance improvements of 4.19% and 2.87% over the RT-DETR baseline processing algorithm in mAP50 and mAP50-95 metrics, respectively, reaching detection accuracies of 65.68% and 46.35%. Cross-dataset generalization experiments further confirm the broad applicability of the framework.