基于LE-YOLOv8的轻量级交通标识线检测研究  

Study on Lightweight Traffic Sign Line Detection Based on LE-Yolov8

作　　者：孔宸 胡海霞[1,2] 向岑阳 KONG Chen;HU Haixia;XIANG Cenyang(College of Mechatronics Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China;Key Laboratory of Mine Intelligent Equipment and Technology,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学机电工程学院,安徽淮南232001 [2]安徽理工大学矿山智能装备与技术重点实验室,安徽淮南232001

出　　处：《兰州工业学院学报》2025年第2期38-45,共8页Journal of Lanzhou Institute of Technology

基　　金：安徽省教育厅重点项目(KJ2020A00282);淮南市科技计划项目(2023A3113)。

摘　　要：针对智能辅助驾驶在环境感知中对交通道路标识线实时检测问题,提出了一种轻量高效的LE-YOLOv8道路标识线检测算法。该算法应用高效“幽灵”卷积EGConv来代替传统卷积部分,通过简洁的线性运算生成更多特征映射,增大感受野;利用重参数化动态卷积DyConv重构C2f模块,使网络增加参数量的同时保持较低的FLOPs;颈部使用跨尺度融合模块CCFM,以增强模型对于尺度变化的适应性和对小尺度目标的检测能力;结合轻量级的自注意力检测头Head_SA,来获取全局特征信息,最大限度地提高检测效率;利用边界框回归损失函数Focaler-IoU,通过线性映射来重构IoU损失,解决现有边界框回归不足问题。改进后的LE-YOLOv8算法计算量、参数量、权重大小仅有3.3 G、1.1 M、2.4 M,相对于YOLOv8n基础算法分别减少了59.8%、65.6%、61.3%;FPS达到220帧/s,提升了39%;mAP值达到90.1%,较于基础算法提升0.2个百分点。Aiming at the problem of real-time detection of traffic road signs in environment perception by intelligent assisted driving,a lightweight and efficient LE-YOLOv8 road sign line detection algorithm is proposed.In this algorithm,efficient"ghost"convolution EGConv is used to replace the traditional convolution part,and more feature maps are generated by simple linear operations to increase the receptive field.The C2f module is reconstructed by using reparameterized dynamic convolution DyConv to increase the number of parameters while maintaining low FLOPs.In the neck,cross-scale fusion module CCFM is used to enhance the adaptability of the model to scale changes and the detection ability of small-scale targets.Combined with lightweight self-attention detection Head SA,the global feature information is obtained to maximize the detection efficiency.Using the bounding box regression loss function Focaler-IoU,the IoU loss is reconstructed by linear mapping to solve the problem of insufficient bounding box regression.The calculation amount,parameter number and weight of the improved LE-YOLOv8 algorithm are only 3.3G,1.1M and 2.4M,which are reduced by 59.8%,65.6%and 61.3%respectively compared with the basic YOLOv8n algorithm;mAP value reached 90.1%,which is 0.2 percentage points higher than that of the basic algorithm.

关 键 词：辅助驾驶 YOLOv8n 交通标识线 EGConv 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]