【摘要】 针对车辆检测领域中现有的小目标检测算法面临的复杂度高、特征提取不足以及检测率低的问题，提出了一种改进模型ARR-DETR，旨在提升小目标车辆检测的效率和精度。借助CSP的思想，通过卷积加性自注意力机制（convolutional additive self-attention,CATM）和卷积门控线性单元（gated linear unit,GLU）构建CSP-ADD-CGLU模块，以改进RT-DETR的骨干网络，在降低计算复杂度的同时增强关键特征的提取能力，从而提升模型的表现力和训练稳定性。构建AIFI-RepBN模块，采用渐进重参数化方法，在训练过程中逐渐从LayerNorm过渡到BatchNorm，实现更高效的计算，同时保持模型性能。在融合模块中引入显式空间先验和注意力分解，强化对小目标位置的感知能力，同时提升对不同尺度特征的选择性关注能力，提高模型对小目标的识别精度，还增强了整体特征融合的效果。实验结果表明，改进的RT-DETR模型在BDD100K数据集上的P、Recall和FPS指标分别为74.4%、66.1%和67.4%。与原始RTDETR模型相比，分别提升了1.6%、2.1%和3.3%，表明该方法可以更快速、更准确地检测到小目标车辆。更多还原

【Abstract】 To address the issues of high complexity,insufficient feature extraction,and low detection rates in existing small object detection algorithms for vehicle detection,proposes an improved model,ARR-DETR,aimed at enhancing the efficiency and accuracy of detecting small vehicle targets.Firstly,leveraging the concept of CSP,the CSP-ADD-CGLU module is constructed using the convolutional additive self-attention mechanism(CATM)and the convolutional gated linear unit(GLU)to improve the backbone network of RT-DETR.This enhancement reduces computational complexity while boosting the extraction of key features,thereby improving the model’s expressiveness and training stability.Secondly,the AIFI-RepBN module is developed using a progressive re-parameterization approach to gradually transition from LayerNorm to BatchNorm during training,achieving more efficient computation while maintaining model performance.Finally,explicit spatial priors and attention decomposition are introduced in the fusion module to strengthen the perception of small target locations and improve selective attention to features of different scales,thereby enhancing the model’s accuracy in recognizing small targets and improving overall feature fusion effectiveness. Experimental results show that the improved RT-DETR model achieves precision,recall,and FPS metrics of 74.4%,66.1%,and 67.4% on the BDD100K dataset,representing improvements of 1.6%,2.1%,and 3.3% over the original RT-DETR model,respectively,indicating that the proposed method can detect small vehicle targets more quickly and accurately.更多还原