摘要： 针对现有的行人检测仍存在密集行人或小目标行人检测精度低的问题，本文提出了一种基于YOLOv5的综合改进算法模型YOLOv5_Conv-SPD_DAFPN(Asymptotic Feature Pyramid Network)。首先，针对小目标或密集行人的特征信息易丢失这一问题，本文在骨干网络中引入了Conv-SPD网络模块替代原有的跨步卷积，从而有效地缓解了特征信息丢失的问题。其次，针对非相邻特征图不直接融合从而引起特征融合率较低的问题，本文提出了全新的双层渐进金字塔网络DAFPN，显著提高了行人检测的准确性和精度。最后，本文基于EIoU-Loss和CIoU-Loss引入EfficiCIOU-Loss的定位损失函数，以调整和加速帧回归率，从而促进了网络模型的更快收敛。本算法模型在CrowdHuman和WiderPerson行人数据集上相比于原YOLOv5模型算法， AP50、AP50-95分别提升了3.9、5.3个百分点和2.1、2.1个百分点；引入EfficiCIOU-Loss之后，本模型收敛速度分别提升了11%、33%。这些创新性的改进使得基于YOLOv5的密集行人检测在特征信息保留、多尺度融合和损失函数优化等方面都取得了显著进展，提高了其在实际应用中的性能和效率。

Abstract: In view of the problem of low accuracy in existing pedestrian detection methods for dense pedestrians or small target pedestrians, this article proposes a comprehensive improved algorithm model called YOLOv5_Conv-SPD_DAFPN (Asymptotic Feature Pyramid Network) based on YOLOv5. Firstly, to address the issue of feature information loss for small targets or dense pedestrians, this article introduces a Conv-SPD network module in the backbone network to replace the original skip convolution, effectively mitigating the problem of feature information loss. Secondly, to solve the problem of low feature fusion rates caused by non-adjacent feature maps not directly merging, this article proposes a new two-layer gradual pyramid network called DAFPN, which significantly improves the accuracy and precision of pedestrian detection. Finally, based on EIoU-Loss and CIoU-Loss, this article introduces the EfficiCIOU-Loss location loss function to adjust and accelerate the frame regression rate, promoting faster convergence of the network model. The algorithm model improves AP50 and AP50-95 by 3.9 and 5.3 percentage points, respectively, compared to the original YOLOv5 model algorithm on the CrowdHuman and WiderPerson pedestrian datasets. After introducing EfficiCIOU-Loss, the model convergence speed has improved by 11% and 33%, respectively. These innovative improvements have made significant progress in dense pedestrian detection based on YOLOv5 in terms of feature information retention, multi-scale fusion, and loss function optimization, enhancing its performance and efficiency in practical applications.