草坪环境下基于YOLO-CGO的上下半身行人检测研究  

Research on Upper and Lower Body Pedestrian Detection in Lawn Based on YOLO-CGO Environment

作　　者：林传渠 曾新华 龙伟 蒋林华[1] 胡灵犀 LIN Chuanqu;ZENG Xinhua;LONG Wei;JIANG Linhua;HU Lingxi(School of Information Engineering,Huzhou University,Huzhou 313000,China;Academy for Engineering&Technology,Fu-dan University,Shanghai 200433,China;Yiwu Research Institute,Fudan University,Jinhua 322000,China)

机构地区：[1]湖州师范学院信息工程学院,浙江湖州313000 [2]复旦大学工程与应用技术研究院,上海200433 [3]复旦大学义乌研究院,浙江金华322000

出　　处：《软件导刊》2024年第12期213-219,共7页Software Guide

基　　金：国家自然科学基金项目(62373148,61972016);上海市科学技术委员会研究基金项目(21JC1405300)。

摘　　要：草坪环境下的行人检测模型存在识别率低、模型体积大、参数多、识别速度慢等问题,难以部署到计算能力有限的机器人平台上。为解决这一问题,对YOLOv5s进行改进,并提出了更轻量化和高精度的YOLO-CGO模型。采用轻量级网络MobileNetv3重置模型的特征提取网络,减少了模型参数量,提高了检测速度;结合CA注意力模块改进颈部网络的C3模块,将颈部网络的卷积层替换为GSConv卷积层,最后一层卷积替换为ODConv卷积层以减轻模型复杂度并保持准确性。实验结果表明,YOLO-CGO模型在自建数据集上比原始模型模型的参数量降低38%,模型体积降低38%,计算量GFLOPS减少50%,取得了明显的轻量化;与原始模型相比,该模型在mAP@0.5上提高了1个百分点,在mAP@0.5:0.95上提高了1.7个百分点。研究表明,YOLO-CGO模型在轻量化的情况下可获得优良的精准度,为算力资源有限的割草机器人自动化、智能化提供了实际可行的途径。There are some problems in lawn environment pedestrian detection model,such as low recognition rate,large model size,multiple parameters,and slow recognition speed,which make it difficult to deploy to robot platforms with limited computing power.A more lightweight with high-precision YOLO-CGO model depending on YOLOv5s is proposed to solve the above problems.First,the feature extraction network of the model was reset using the lightweight network MobileNetv3,reducing the number of model parameters and improving detection speed.Then improve the C3 module of the neck network by combining CA(Coordinate Attention)attention module.In the end replacing convolutional layers of the neck network with GSConv convolutional layers,and the last convolution layer was replaced by the ODConv convolution layer reduces the complexity of the model while maintaining accuracy.The experimental results show that the YOLO-CGO model designed in this paper on the self-built dataset reduces the parameter count by 38%,model volume by 38%,and computational load GFLOPS by 50%compared to the original model,achieving significant lightweighting;And compared with the original model,the model proposed in this article is superior in map@0.5 Up by 1 percentage point map@0.5 Increase by 1.7 percentage points above 0.95.This study indicates that the YOLO-CGO model proposed in the article can achieve excellent accuracy in extremely lightweight situations,providing a practical and feasible approach for the automation and intelligence of lawn mowing robots with limited computing resources.

关 键 词：行人检测 轻量化 GSConv ODConv CA注意力模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]