基于改进YOLOv7-tiny的轻量化道路目标检测算法  

An Improved Lightweight Road Target Detection Algorithm Based on YOLOv7-tiny

作　　者：何泽江 蒋淑霞[1] 柳霞 He Zejiang;Jiang Shuxia;Liu Xia(College of Mechanical and Intelligent Manufacturing,Central South University of Forestry and Technology,Changsha 410004)

机构地区：[1]中南林业科技大学机械与智能制造学院,长沙410004

出　　处：《汽车技术》2025年第2期9-16,共8页Automobile Technology

基　　金：教育部产学合作协同育人项目(230826311807159);湖南省自然科学基金项目(2022JJ50012);湖南省研究生科研创新项目(CX20220735)。

摘　　要：针对目标检测算法对算力和存储空间的高要求限制其在边缘设备中检测功能实时性的问题,提出了一种基于YOLOv7-tiny改进的轻量化道路目标检测算法。首先,通过K-means++聚类算法生成适合道路目标检测的先验锚框;其次,改进ELAN结构轻量化主干网络,同时提出轻量型多尺度特征(LMS)模块优化颈部网络;最后,使用西格玛线性单元(SiLU)激活函数加速模型收敛,采用MPDIoU损失函数进一步提高检测精度。试验结果表明:改进后的模型参数量减少18.3%,计算量降低15.0%,且所有类别平均检测精度上升1.1%。在Jetson TX2中,使用TensorRT加速后的检测速度达到48帧/s,基本满足道路目标检测的实时性要求。To address the high computational and storage demands of object detection algorithms that limit real-time performance of edge devices,this study proposes an improved lightweight road object detection algorithm based on YOLOv7-tiny.First,prior anchor boxes optimized for road object detection are generated using the K-means++clustering algorithm.Secondly,the backbone network is streamlined by modifying the ELAN structure,while a Lightweight Multi-scale Feature(LMS)module is introduced to optimize the neck network.Finally,the Sigmoid Linear Unit(SiLU)activation function is adopted to accelerate model convergence,and the MPDIoU loss function is employed to further improve detection accuracy.Experimental results demonstrate that the improved model achieves an 18.3%reduction in parameters,a 15.0%decrease in computational complexity,and a 1.1%increase in mean average precision across all categories.When deployed on Jetson TX2 with TensorRT acceleration,the detection speed reaches 48 frames per second,essentially meeting real-time requirements for road object detection applications.

关 键 词：自动驾驶 YOLOv7-tiny 道路目标检测 轻量化 Jetson TX2 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]