【摘要】 针对交通标志目标小、模型体积大的问题，提出一种基于改进RT-DETR(real-time detection transformer)的轻量化交通标志检测方法.首先，在骨干网络中引入PConv(partial convolution)卷积优化BasicBlock结构，构建PC＿Block模块，增强特征提取能力，减少参数量和计算量；其次，提出DyASF(dynamic attentional scale sequence fusion)模块，通过自适应方式融合多尺度特征，提高模型的空间感知能力和特征表达能力；而后，使用Focaler-MPDIoU损失函数，通过聚焦不同回归样本以及引入最小点距离，准确评估预测边界框与真实边界框之间的相似度，提升模型的检测性能；最后，引入P2检测层，提高对小目标的检测能力，增强模型鲁棒性.在公共数据集TT100K上进行对比实验，结果表明，改进后的模型参数量降低了53.8%,计算量降低了22.8%,检测精度提高了2.4%,可以更好地满足交通标志检测的要求.更多还原

【Abstract】 Aiming at the problem of small traffic sign targets and large model volume, a lightweight traffic sign detection based on improved RT-DETR(real-time detection transformer)is proposed. Firstly, the PConv(partial convolution)convolution optimization BasicBlock structure is introduced into the backbone network, and the PC＿Block module is constructed, which enhances the feature extraction capability and reduces the number of parameters and computation. Secondly, the DyASF(dynamic attentional scale sequence fusion)module, which fuses multi-scale features in an adaptive way to improve the model’s spatial perception and feature expression ability. Then, using the Focaler-MPDIoU loss function, the similarity between the predicted bounding box and the real bounding box is accurately evaluated by focusing on different regression samples and introducing the distance of the minima, which improves the detection performance of the model. Finally, the P2 detection layer is introduced that improves the detection of small targets and enhances the model robustness. Comparison experiments are conducted on the public dataset TT100K,and the results show that the improved model reduces the number of parameters by 53.8%,the computation amount by 22.8%,and the detection accuracy by 2.4%,which can better meet the requirements of traffic sign detection.更多还原