基于YOLOv8的轻量化多尺度交通目标检测算法  

Lightweight multi-scale traffic object detection algorithm based on YOLOv8

作　　者：吴冲 汪烨(指导)[1] WU Chong;WANG Ye(College of Mechanical Engineering,Shanghai Dianji University,Shanghai 201306,China)

机构地区：[1]上海电机学院机械学院,上海市201306

出　　处：《上海电机学院学报》2025年第1期39-44,51,共7页Journal of Shanghai Dianji University

摘　　要：针对交通场景下目标尺度变化大和检测算法复杂度高的问题,提出了一种轻量化多尺度检测算法(DMB-YOLOv8)。首先,通过引入扩张残差块构建C2f-DWR,在主干网络高层扩展合适的感受野,增强多尺度目标识别能力;其次,构建多协调注意力更全面地捕捉目标位置特征,聚焦关键检测内容;最后,设计轻量级多融合双向网络(MBN),改进YOLOv8的颈部结构,提升多尺度特征融合能力,保留更多浅层的小目标信息特征。实验结果表明:相较于基础模型,DMB-YOLOv8的平均精度m AP50提升了1.11%,参数量降低了25.8%,在提升检测精度的同时实现了轻量化,适用于实际交通场景需求。In traffic scenarios,large target scale variations and the high computational complexity of detection algorithms pose significant challenges.To address these issues,a lightweight multi-scale detection algorithm named DMB-YOLOv8 is proposed.The algorithm introduces a dilation-wise residual block(C2f-DWR)to expand the receptive field in the high-level backbone network,enhancing multi-scale target recognition.A multi-coordination attention mechanism is employed to capture positional features of targets comprehensively and emphasize critical detection regions.Additionally,a lightweight multi-fusion bidirectional network(MBN)is designed to optimize the neck structure of YOLOv8,improving multi-scale feature fusion and preserving shallow-layer small-target information.Experimental evaluations on the BDD1ooK dataset demonstrate that DMB-YOLOv8 achieves a 1.11%increase in mean average precision(mAP50)and a 25.8%reduction in parameter count compared to the baseline model.The proposed algorithm not only enhances detection accuracy but also maintains a lightweight design,making it well-suited for real-world traffic applications.

关 键 词：交通目标检测 扩张残差 注意力机制 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]