基于改进YOLOv8n算法的城市空间混行交通参与者检测模型  

Urban spatially mixed traffic participants detection model based on improved YOLOv8n

作　　者：周军超 陈鑫 高建杰[3] 章杰 ZHOU Junchao;CHEN Xin;GAO Jianjie;ZHANG Jie(School of Mechanical Engineering,Sichuan University of Science&Engineering,Zigong Sichuan 643000,China;Chengdu-Chongqing Economic Circle(Luzhou)Advanced Technology Research Institute,Luzhou Sichuan 646000,China;Intelligent Policing Key Laboratory of Sichuan Province,Sichuan Police College,Luzhou Sichuan 646000,China;College of Automotive and Mechanical Engineering,Changsha University of Science&Technology,Changsha Hunan 410114,China)

机构地区：[1]四川轻化工大学机械工程学院,四川自贡643000 [2]成渝地区双城经济圈(泸州)先进技术研究院,四川泸州646000 [3]四川警察学院智能警务四川省重点实验室,四川泸州646000 [4]长沙理工大学汽车与机械工程学院,湖南长沙410114

出　　处：《中国安全科学学报》2024年第12期178-186,共9页China Safety Science Journal

基　　金：智能警务四川省重点实验室开放课题资助(ZNJW2023KFQN005,ZNJW2022KFQN004);泸州市科技计划资助(2023JYJ066)。

摘　　要：为提升智能网联汽车和交通监测系统对交通参与者的识别精度和检测速度,及时应对城市空间混行交通环境中的安全隐患,提出一种基于改进YOLOv8n算法的城市空间混行交通参与者检测模型。首先,在数据输入阶段利用几何变换和像素变换策略来防止过度拟合,提高鲁棒性和泛化性;其次,使用空间到深度的非跨行卷积(SPD-Conv)模块代替所有原始卷积层,提高对低分辨率小目标的特征提取能力;同时,在颈部网络融合结构中加入轻量级坐标注意力机制(CA)模块,在几乎不添加任何计算开销的同时提高对关键信息的识别精度;然后,用边界框损失函数有效交并比(EIoU)替代原损失函数,使模型获得更卓越的收敛速度与识别稳定性;最后,利用平台公开和自建融合的交通参与者数据集(BNS)进行消融和对比试验,运用自动驾驶试验平台进行实景实时检测。结果表明:相比于基线模型YOLOv8n,改进模型SEC-YOLO的每秒传输帧数(FPS)和平均精度均值(mAP)分别提高了7.3%和3.2%;与主流模型对比,mAP和FPS性能值最佳;在自动驾驶试验平台上的实景检测平均准确率为95%。SEC-YOLO算法模型对城市交通参与者的检测准确率更高,鲁棒性和实时性更强。In order to improve the recognition accuracy and detection speed of traffic participants by intelligent networked vehicles and traffic monitoring systems so that they can timely respond to the safety hazards in the mixed traffic environment in urban space,a mixed traffic participant detection model in urban space based on the improved YOLOv8n algorithm was proposed.Firstly,geometric transformation and pixel transformation enhancement strategies were employed in the data input stage to prevent overfitting and improve robustness,and generalization.Secondly,the SPD-Conv module was used to replace all original convolution layers of the YOLOv8n algorithm,which enhances the feature extraction capability for low-resolution small targets.Meanwhile,the CA module was added to the fusion structure of the neck network of the YOLOv8n algorithm to improve the recognition accuracy of key information with almost no additional computational overhead.Then,the boundary box loss function EIoU was used to replace the original loss function,enabling the model to achieve superior convergence speed and recognition stability.Finally,the ablation and comparison experiments were carried out with the public and self-built integrated traffic participant dataset,and the real-time detection experiment was carried out with the automatic driving experiment platform.The experimental results show that compared to the YOLOv8n model,the improved SEC-YOLO model has increased mAP and FPS by 3.2%and 7.9%respectively.The SEC-YOLO model outperforms mainstream models in terms of mAP and FPS as well.The average accuracy of real-scene detection on the automatic driving experimental platform is around 95%.The SEC-YOLO algorithm model achieves higher detection accuracy for urban traffic participants,with stronger robustness and real-time performance.

关 键 词：YOLOv8n 空间混行 交通参与者 检测模型 空间到深度的非跨行卷积(SPD-Conv) 坐标注意力机制(CA) 

分 类 号：X951[环境科学与工程—安全科学]