基于深度学习的草坪障碍物检测研究  

Research on Grassland Obstacle Detection Based on Deep Learning

作　　者：黄东辉 向宇[2] 高巧明 HUANG Donghui;XIANG Yu;GAO Qiaoming(School of Mechanical and Automotive Engineering,Guangxi University of Science and Technology,Liuzhou 545006,China;Guangxi Automotive Parts and Vehicle Technology Key Laboratory,Guangxi University of Science and Technology,Liuzhou 545006,China)

机构地区：[1]广西科技大学机械与汽车工程学院,广西柳州545006 [2]广西科技大学广西汽车零部件与整车技术重点实验室,广西柳州545006

出　　处：《拖拉机与农用运输车》2022年第4期42-45,57,共5页Tractor & Farm Transporter

摘　　要：草坪障碍物的检测是智能割草机械不可避免的问题,它关系到割草机能否安全、稳定、高效的运行。考虑到障碍物检测的实时性以及嵌入式平台的应用,本文对现有的YOLOv4目标检测模型做出改进。采用MobileNetV2网络作为YOLOv4模型中的主干网络,并通过K-means算法进行聚类以获得更适应于检测草坪障碍物的先验框。实验结果表明,在自制数据集上本文所设计的轻量化YOLOv4较原模型在参数量方面减小了39%,检测速度提升了49%,mAP为94.40%,为割草机的实际应用提供了一种思路。The detection of lawn obstacles is an inevitable problem of intelligent mower,which is related to the safe,stable and efficient operation of mower.Considering the real-time performance of obstacle detection and the application of embedded platform,this paper improves the existing YOLOv4 target detection model.The MobileNetV2 network is used as the backbone network in the YOLOv4 model,and the K-means algorithm is used for clustering to obtain priors more suitable for detecting lawn obstacles.The experimental results show that the lightweight YOLOv4 designed in this paper on the self-made dataset has 39%reduction in the number of parameters and 49%improvement in the detection speed with 94.40%mAP compared with the original model,which provides an idea for the practical application of lawn mower.

关 键 词：深度学习 机器视觉 割草机 YOLOv4 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]