To address the issue of decreased vehicle detection accuracy in rainy conditions due to reduced visibility, raindrop obstruction, and insufficient contrast, this paper proposes a rain-based vehicle detection method (R-RTDETR) based on an improved RT-DETR-R18 model. The original backbone network is replaced with a combined Conv and C2f-wConv network, and weighted convolutions are used to adaptively adjust features, enhancing the representation of key region information. In the encoder, a Global-Local Spatial Attention (GLSA) module is introduced to achieve efficient fusion of global semantics and local details. In the decoder, a P2 shallow feature branch is introduced to optimize the multi-scale fusion path and strengthen vehicle perception capabilities. Experimental results show that R-RTDETR improves mAP by 1.52% and recall by 1.25% on the test set compared to the baseline RT-DETR-R18; on the ACDC rainy dataset, mAP@0.5 is improved by 2.33%, demonstrating improved stability and robustness of rain-based vehicle detection while maintaining computational efficiency, validating its application potential in intelligent transportation scenarios. 针对雨天环境下车辆检测易受能见度降低、雨滴遮挡及对比度不足等因素影响而导致检测精度下降的问题，本文提出一种基于改进RT-DETR-R18模型的雨天车辆检测方法（R-RTDETR）。采用Conv与C2f-wConv组合网络替换原主干网络，通过加权卷积实现特征自适应调节，以增强关键区域信息表达；在编码器，引入全局-局部空间注意力模块（Global-Local Spatial Attention, GLSA）以实现全局语义与局部细节的高效融合；在解码器，引入P2浅层特征分支，优化多尺度融合路径，强化车辆感知能力。实验结果表明，R-RTDETR在测试集上的mAP较基线RT-DETR-R18提升1.52%，召回率提升1.25%；在ACDC雨天数据集上的mAP@0.5提升2.33%，在保证计算效率的同时提高了雨天车辆检测的稳定性与鲁棒性，验证了其在智能交通场景中的应用潜力。
Yang et al. (Fri,) studied this question.