面向复杂场景的多尺度行人和车辆检测算法  

Multi⁃scale pedestrian and vehicle detection algorithm for complex scenes

作　　者：王娟敏 皮建勇[1] 黄昆 胡伟超 胡倩 WANG Juanmin;PI Jianyong;HUANG Kun;HU Weichao;HU Qian(State Key Laboratory of Public Big Data,College of Computer Science and Technology,Guizhou University,Guiyang 550000,China)

机构地区：[1]贵州大学计算机科学与技术学院公共大数据国家重点实验室,贵州贵阳550000

出　　处：《现代电子技术》2025年第9期143-153,共11页Modern Electronics Technique

摘　　要：为了解决行人和车辆检测任务中由于多尺度和遮挡现象导致漏检的问题,文中提出一种基于YOLOv8的改进检测算法RDRFM-YOLO。对于主干网络,设计RFDRep模块替代卷积和C2f模块,以加强网络对于不同尺度特征的捕获能力;对于颈部网络,设计SFMS模块进行优化,以提升模型对遮挡目标的特征提取能力。在自制的行人和车辆数据集上的实验表明,改进的RDRFM-YOLO相较于原始算法有更好的性能表现,同时保持了高效的检测效率。mAP@0.5达到了56.7%,mAP@0.5:0.95达到了37.3%,相比于原始算法分别提高了2.8%和2.3%,参数量和浮点运算量为3.3×10^(6)和9.2×10^(9),相比于原始算法仅增加了0.1×10^(6)和0.3×10^(9)。同时,模型在多个数据集上均有较好的性能表现。A YOLOv8-based improved detection algorithm named RDRFM-YOLO is presented to address the issue of missed detections caused by multi-scale cases and occlusion in pedestrian and vehicle detection tasks.For the backbone network,the RFDRep module is designed to replace the convolution and C2f modules,enhancing the network′s capability to capture features at different scales.For the neck network,the SFMS module is designed for optimization,improving the model′s ability to extract features of occluded objects.Experiments on a custom pedestrian and vehicle dataset show that the algorithm RDRFM-YOLO outperforms the original algorithm,maintaining high detection efficiency.The mAP@0.5 of the RDRFM-YOLO reaches 56.7%,and its mAP@0.5:0.95 reaches 37.3%,which are improvements of 2.8%and 2.3%,respectively,over the original algorithm.Its parameter count and floating-point operations are 3.3×10^(6) and 9.2×10^(9),only increasing by 0.1×10^(6) and 0.3×10^(9),respectively,in comparison with those of the original algorithm.Additionally,the model shows good performance across multiple datasets.

关 键 词：行人和车辆检测 多尺度 遮挡 RDRFM-YOLO RFDRep模块 SFMS模块 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]