基于虚拟现实技术的自动驾驶仿真训练研究
Research on Automatic Driving Simulation Training Based on Virtual Reality Technology
李成林 1贺鹏飞 1范鹏飞 1徐康 2聂荣3
作者信息
- 1. 烟台大学物理与电子信息学院 烟台 264005
- 2. 青岛阳氢集团有限公司 青岛 266113
- 3. 郑州航空工业管理学院智能工程学院 郑州 450046
- 折叠
摘要
随着自动驾驶技术的发展,现实中的路测数据已经难以满足相关算法的测试需求.为了缩短自动驾驶算法的开发周期,减少开发成本,论文结合数字孪生思想,搭建了自动驾驶仿真训练平台.仿真平台采用Unity3D引擎进行自动驾驶仿真,使用车辆内置激光雷达和摄像头采集仿真数据;通过Python建立服务器读取摄像头数据,并同步仿真画面进行道路分析,其中采用DeepLabV3+神经网络检测道路,OpenCV检测周围环境;同时编译制图软件连接服务器,通过内存共享机制,读取激光雷达数据,进行实时道路绘制.平台生成的数据集与现实路况生成的数据集,经过10次训练产生的训练误差相似度为93.3%,验证误差相似度为90.9%.
Abstract
With the development of autonomous driving technology,the realistic road test data has been difficult to meet the testing requirements of relevant algorithms.In order to shorten the development cycle of autonomous driving algorithms and reduce the development cost,this paper combines the digital twin idea to build an autonomous driving simulation training platform.The sim-ulation platform adopts Unity3D engine for autonomous driving simulation,and uses vehicle built-in LiDAR and camera to collect simulation data.It establishes a server through Python to read camera data and synchronize simulation screen for road analysis,in which DeepLabV3+ neural network is used to detect the road and OpenCV to detect the surrounding environment.At the same time,its compiles mapping software to connect to the server.Through the memory sharing mechanism,the LIDAR data is read and re-al-time road mapping is performed.The similarity between the data set generated by the platform and that generated by the real road conditions after 10 training sessions is 93.3%for the training error and 90.9%for the validation error.
关键词
自动驾驶/数字孪生/Unity3D/DeepLabV3+/OpenCVKey words
autopilot/digital twins/Unity3D/DeepLabV3+/OpenCV引用本文复制引用
基金项目
2022年河南省科技攻关项目(222102220048)
烟台市2021年校地融合发展项目(1521001-WL21JY01)
出版年
2024
国家科技期刊平台
NSTL
万方数据