摘要
为解决复杂交通环境中小目标、遮挡目标检测能力不足,泛化能力弱等问题,设计了一种基于YOLOv8的优化模型,命名为YOLOv8-MS。提出了轻量感受野增强模块MGSmodule,基于此模块对C2f模块进行优化,提高特征提取的效率和准确性。引入了分离和增强注意力模块SEAM,该模块有效聚焦于受群体集中影响的区域,从而提升模型对遮挡目标的检测能力。设计专门针对小目标的下采样特征提取器,旨在降低小目标的误检率和漏检率,进一步提高检测准确率。在KITTI数据集上进行验证,改进算法的P、R、mAP50、mAP50-95较基准模型YOLOv8n分别提高0.9%、6.3%、5.7%、4.8%。同时,模型在VisDrone数据集验证,改进后的模型的P、R、mAP50、mAP50-95分别提高2.5%、2.3%、2.6%、1.6%,显示出良好的泛化性和鲁棒性。
To address the insufficient detection and weak generalization of small and occluded targets in complex traffic environments,an optimized model based on YOLOv8,named YOLOv8-MS,is developed.First,a lightweight receptive field enhancement module MGSmodule is proposed to optimize the C2f module,improving the efficiency and accuracy of feature extraction.Then,the separation and enhanced attention module SEAM is introduced,which effectively focuses on the areas affected by the concentration of the group,thereby improving the model’s detection of small targets.Finally,a downsampling feature extractor is specifically designed for small targets,aiming to reduce the false detection and missed detection rates of small targets and further improve detection accuracy.Verified on the KITTI dataset,the improved algorithm improves P,R,mAP50,and mAP50-95 compared to the benchmark model YOLOv8n by 0.9%,6.3%,5.7%,and 4.8% respectively.Meanwhile,the model is validated on the VisDrone dataset.It improves P,R,mAP50,and mAP50-95 by 2.5%,2.3%,2.6%,and 1.6%respectively,demonstrating fairly good generalization and robustness.
作者
张瑞乾
秦慧军
陈勇
袁旭浩
周若轩
ZHANG Ruiqian;QIN Huijun;CHEN Yong;YUAN Xuhao;ZHOU Ruoxuan(Beijing Information Science and Technology University,Beijing 100192,China;Bejing Laboratory for New Enerey Vehicles,Beijing 100192,China)
出处
《重庆理工大学学报(自然科学)》
北大核心
2025年第9期23-30,共8页
Journal of Chongqing University of Technology:Natural Science
基金
国家自然科学基金面上项目(52077007)
新能源汽车北京实验室建设项目(PXM2020_014224)。
