摘要
针对现有自动驾驶课程缺乏软硬件相结合、稳定性强且易于开发的车道保持实验,提出了一种基于深度学习端到端算法的车道保持系统。引入轻量化的无人驾驶小车硬件平台与华为云ModelArts软件平台,设计了包括数据采集处理、模型设计、云端训练、车端部署测试的完整开发流程;从教学场景出发,对深度学习端到端算法进行轻量设计,较好捕捉图像与底盘角度的映射关系,保证了高准确率与泛化性。同时,进行了模型训练及实验测试,训练了200个Epoch后成功收敛,在U型弯、直转弯、S型弯严格不压线通过率均值为93.05%(高速)与97.22%(低速)。实验结果表明,该系统在实测中表现出较强的鲁棒性,且实验环境易于搭建,弥补了教学实验中学生计算资源不足的现状;通过端云协同开发帮助学生有效提升工程实践能力。
In the existing driverless courses,we lack strong stability and easy-to-developed lane keeping experiments which combine software and hardware.Therefore,a good experiment based on deep learning end-to-end algorithm is proposed.A lightweight driverless car platform and a Huawei Cloud platform are introduced to complete data acquisition and processing,model design and training,vehicle deployment and testing in teaching area.The improved end-to-end algorithm can better capture the mapping relationship between the image and the chassis angle,and ensure high accuracy and good generalization performance.In the experiment,the model converges successfully after training 200 epochs.The mean value of strict stability rate in U-bend,straight turn and S-bend sections are 93.05%(high-speed mode)and 97.22%(low-speed mode),respectively.The experiment is easy to build environment and show its strong stability in testing.Besides,it makes up for the shortage of students’local computing resources.Moreover,the end-cloud collaborative development helps students improve their engineering practice ability.
作者
肖雄子彦
楚朋志
梁晓妮
薛万坤
任桐鑫
XIAO Xiongziyan;CHU Pengzhi;LIANG Xiaoni;XUE Wankun;REN Tongxin(Student Innovation Center,Shanghai Jiao Tong University,Shanghai 200240,China)
出处
《实验室研究与探索》
CAS
北大核心
2022年第12期27-33,共7页
Research and Exploration In Laboratory
基金
教育部产学合作育人项目(202002142023)。
关键词
车道保持
无人驾驶
深度学习
教学实验
lane keeping
unmanned driving
deep learning
teaching experiment
