应用YOLO v4模型的赛道锥桶检测与识别方法  

Track Cone Barrels Detection and Recognition Method Based on YOLO v4 Model

作　　者：李强 陶立波 杨爱喜 Agyei PHILIP LI Qiang;TAO Li-bo;YANG Ai-xi;Agyei PHILIP(School of Mechanical&Energy Engineering,Zhejiang University of Science and Technology,Zhejiang Hangzhou 310023,China;Polytechnic Institute,Zhejiang University,Zhejiang Hangzhou 310015,China)

机构地区：[1]浙江科技学院机械与能源工程学院,浙江杭州310023 [2]浙江大学工程师学院,浙江杭州310015

出　　处：《机械设计与制造》2024年第10期20-24,33,共6页Machinery Design & Manufacture

基　　金：浙江省自然科学基金项目(LY21E050001);汽车新技术安徽省工程技术研究中心开放基金项目(QCKJ202105)。

摘　　要：为快速检测与准确识别赛道锥桶,提出了一种基于YOLO v4模型的赛道锥桶检测与识别方法。首先依据复杂多变的赛道场景采集了多张锥桶图像作为数据集原始数据,在工控机上进行锥桶数据集制作、训练和模型选取;然后搭建基于YOLO v4模型的锥桶检测与识别系统,选择三种较为常见赛道场景进行实车试验。试验结果表明,所提出的方法在不同光照条件下仍能快速检测并准确识别目标锥桶,特别是在锥桶较为密集且多个锥桶目标重叠的场景下,置信度达到0.91以上,具有较强的鲁棒性,且实时检测的平均帧率达到35f/s,能够满足无人驾驶方程式赛车对感知系统准确性和实时性的需求。In order to quickly detect and accurately identify the track cone barrel,a track cone barrel detection and recognition method based on YOLO v4 model is proposed.Firstly,according to the complex and changeable track scene,multiple cone barrel images are collected as the original data of the data set,and the cone barrel data set is made,trained and selected on the industrial computer;Then build a cone barrel detection and recognition system based on YOLO v4 model,and select three common track scenes for real vehicle test.The experimental results show that this method can still quickly detect and accurately identify the target cone barrel under different lighting conditions,especially in the scene where the cone barrel is dense and multiple cone barrel targets overlap,the confidence is more than 0.91,has strong robustness,and the average frame rate of real-time detection is 35f/s,it can meet the needs of driverless formula racing for the accuracy and real-time performance of the sensing system.

关 键 词：无人驾驶方程式赛车 YOLO v4 赛道锥桶 目标检测 锥桶识别 

分 类 号：TH16[机械工程—机械制造及自动化] U462.1[机械工程—车辆工程]