基于YOLOv8s模型改进的道路交通目标检测方法研究  

An improved YOLOv8s method and its application in road traffic target detection

作　　者：桑嘉更 张志佳[1] 肖传民 罗海波[2] 张俊摇 SANG Jiageng;ZHANG Zhijia;XIAO Chuanmin;LUO Haibo;ZHANG Junyao(College of Artificial Intelligence,Shenyang University of Technology,Shenyang 110870,China;Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110169,China;The Third Militray Representative Office of the Air Force Equipment Department,Shenyang 110144,China;China Academy of Machinery Shenyang Research Institute of Foundry Co.,Ltd.,Shenyang 110022,China)

机构地区：[1]沈阳工业大学人工智能学院,辽宁沈阳110870 [2]中国科学院沈阳自动化研究所,辽宁沈阳110169 [3]空装驻沈阳地区第三军事代表室,辽宁沈阳110144 [4]中国机械总院集团沈阳铸造研究所有限公司,辽宁沈阳110022

出　　处：《红外与激光工程》2024年第11期294-307,共14页Infrared and Laser Engineering

基　　金：辽宁省应用基础研究计划项目(2023JH2/101300237)。

摘　　要：红外图像目标检测在交通领域中有很重要的应用价值,然而,由于红外图像存在分辨率低、缺乏颜色信息、对比度差、特征模糊的特点,导致现有模型在检测红外车辆与行人时精度不高。为此,文中对YOLOv8s进行了改进,首先对特征融合机制进行改进,在网络中添加小目标检测层,充分利用目标的浅层特征信息,提高对小目标检测的准确性。其次引入了SPD(Space to Depth)细粒化模块来代替YOLOv8s中的3×3卷积进行下采样,避免了3×3卷积下采样导致红外图像细粒度信息丢失。并且还设计了一个新的混合注意力机制,使网络更好地聚焦感兴趣的区域,减少背景对行人和车辆检测的干扰,增强模型对目标特征的关注度。最后使用Focal EIOU损失函数代替CIOU损失函数,改善了CIOU在特殊情况失效和正负样本不平衡的问题。在交通场景红外图像数据集FLIR_ADAS_v2上进了行实验,验证了算法的有效性。与YOLOv8s相比,改进后的模型mAP@0.5从83.4%提升到了89.3%。Objective Infrared image target detection has significant application value in the field of transportation,as it can help people promptly detect targets and respond in special conditions such as strong light at night or in rainy and foggy weather.However,due to the characteristics of infrared images,such as low resolution,lack of color information,poor contrast,and blurred features,existing models do not achieve high average detection accuracy when detecting infrared vehicles and pedestrians.The main issue is the problem of missing detection for overlapping targets and small targets in traffic scenes.Therefore,this paper aims to design an infrared pedestrian and vehicle detection model based on YOLOv8s(You only look once version 8),which is crucial for improving the safety of intelligent driving.Methods YOLOv8s,an advanced object detection model in recent years,is categorized into five distinct versions—n,s,m,l,and x—according to the network's depth and breadth to cater to diverse requirements.YOLOv8s,ensuring a certain level of detection precision with a moderate model size,is chosen as the base model.The manuscript introduces four improvements to the YOLOv8s architecture(Fig.2).Firstly,the network architecture is re-engineered with the incorporation of a small target detection layer to improve detection capabilities for distant pedestrians and vehicles(Fig.3).Secondly,the SPD(space-to-depth)module replaces the original network's 3×3 downsampling convolution in the backbone and neck networks(Fig.4),to safeguard the fine-grained details within the image.Thirdly,a hybrid attention mechanism(Fig.5)is crafted to bolster the network's attentiveness to pedestrians and vehicles.Fourthly,the Focal EIOU loss function is utilized,which not only addresses the deficiencies of the CIOU loss function that may become ineffective under certain circumstances but also mitigates the issue of imbalance between positive and negative samples.Results and Discussions The dataset utilized in this study is the FLIR ADAS(Advanced Dri

关 键 词：YOLOv8s 目标检测 注意力机制 损失函数 细粒化卷积 红外图像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]