Abstract:As the core supporting technology for intelligent transportation systems, intelligent real-time traffic sign detection plays an irreplaceable strategic role in constructing autonomous driving decision-making and high-precision dynamic maps. Given the real-time detection needs in complex and variable road scenarios, especially the limitations of existing DETR models in complex scenarios and small object detection, this study proposes an object detection model. To solve the problem of traffic sign feature degradation in harsh conditions such as strong illumination, rain and snow, the OmniFocus-RT-DETR model is introduced. The model introduces the dynamic-range histogram self-attention (DHSA) mechanism, which adaptively handles pixel patterns degraded by weather conditions and combines dynamic range convolution to effectively alleviate feature distortion caused by severe illumination changes. Additionally, spatially separated and deformable convolution (SPDConv) and the CSP-OmniKernel module are introduced, and combined with joint learning in both spatial and frequency domains, the robustness in complex environments is significantly enhanced. Experimental results show that compared to the baseline model RT-DETR, OmniFocus-RT-DETR improves mAP@50 by 8.9% and 7.7% on the TT100K_aug and CCSTDB_aug datasets respectively, and small object detection accuracy increases by 9.1% and 4%. Additionally, it outperforms current mainstream methods in accuracy, robustness, and real-time performance. Experimental results reveal that the improved method can achieve higher object detection accuracy of traffic signs in complex scenarios.