基于改进YOLOv4的三维点云导盲系统设计  

DESIGN OF 3D POINT CLOUD BLIND GUIDE SYSTEM BASED ON IMPROVED YOLOV4

作　　者：杜龙龙[1] 陆学斌 罗孝[2] Du Longlong;Lu Xuebin;Luo Xiao(School of Information Engineering,Huzhou Vocational and Technical College,Huzhou 313000,Zhejiang,China;School of Information Engineering,Huzhou University,Huzhou 313000,Zhejiang,China)

机构地区：[1]湖州职业技术学院物流与信息工程学院,浙江湖州313000 [2]湖州师范学院信息工程学院,浙江湖州313000

出　　处：《计算机应用与软件》2025年第2期94-101,共8页Computer Applications and Software

基　　金：黑龙江省自然科学基金项目(F2018018);浙江省教育科研项目(Y201942338);浙江省访问工程师项目(FG2019129);浙江省大学生科技创新活动计划暨新苗人才计划项目(2020R443003)。

摘　　要：针对传统导盲系统存在抗环境干扰能力较低,识别准确率低等问题,提出一种基于YOLOv4的激光点云导盲系统。运用高精度激光扫描仪采集路况点云信息,将点云信息转换成包含特征信息的投影图像并建立路况数据集,搭建基于DARKNET的网络模型训练框架识别复杂路况。采用K-means++算法改进YOLOv4模型中原有的聚类算法,提高模型多尺度检测的适应性。实验结果表明,系统识别复杂路况的平均精度为98.12%,与同类产品相比能够准确、稳定识别路况障碍。Aimed at the problem that traditional blind guide system is susceptible to environmental interference and has low recognition accuracy,a 3D point cloud blind guide system based on YOLOv4 is proposed.This design used the high precision laser scanner to collect point cloud road information which was sent to the cloud server and converted into projection image containing feature information.The mode established road condition data set and built a network model based on DARKNET training framework to identify the complex road conditions.K-means++algorithm was used to improve the original clustering algorithm of YOLOv4 model and improve the adaptability of multi-scale detection of the model.The experimental results show that the average accuracy of the system for identifying complex road conditions is 98.12%.Compared with similar products,it can accurately and stably identify road conditions and obstacles.

关 键 词：激光扫描 深度学习 YOLOv4 嵌入式应用 导盲系统 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]