基于深度学习的无人驾驶汽车车道跟随方法  

Lane Following Method of Autonomous Vehicle Based on Deep Learning

作　　者：高扬[1] 陈士伟 刘进渊 王书棋 Gao Yang;Chen Shiwei;Liu Jinyuan;Wang Shuqi(Chang’an University,Xi’an 710054)

机构地区：[1]长安大学,西安710054

出　　处：《汽车技术》2022年第3期14-20,共7页Automobile Technology

基　　金：陕西省自然科学基金项目(2019JLP-07)。

摘　　要：为了改善车道跟随算法由于人工进行的复杂模型设计而导致的不利于实现机器的自学习与自设计,并且在一些极端行驶工况下易导致性能下降的问题,将深度学习的注意力机制与循环神经网络相结合提出一种基于深度时空注意力的车道跟随模型,在自制的真实车道跟随数据集上进行测试,所得误差为1.43%。为了验证该模型在黑夜、阴影、无车道线等困难场景下的跟随效果,结合一种时空信息与深层信息融合的车道线检测模型在仿真车道跟随数据集上进行测试,所得误差为2.27%,迁移学习对比测试结果表明,该模型能够在困难场景下有效地实现无人驾驶车辆的车道跟随。In order to improve the lane following algorithm,which is not conducive to the realization of the selflearning and self-design of the machine due to the manual design of the complex model,and is likely to cause performance degradation under some poor driving conditions,this article proposes a lane following model which combines deep learning attention mechanism and recurrent neural network.And the model is trained on the self-made actual lane following data set obtaining a test error rate of 1.43%.In order to verify the following effect of the model in poor driving scenes such as dark night,shadow,no lane,etc.,a lane detection model combing spatial-temporal property with deep information fusion is tested on the simulated lane following data set,a test error of 2.27%is obtained.Finally,the transfer learning comparison experiment results show that the model can effectively achieve lane following for autonomous vehicle in poor driving scenarios.

关 键 词：车道跟随 深度学习 注意力机制 循环神经网络 无人驾驶 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]