辅助车道线检测的端到端自动驾驶

End to end Autonomous Driving Model Using Lane Detection Auxiliary

基于深度学习的端到端自动驾驶有着简洁高效的优势,尤其在车道保持上有着良好表现,但是面临路况复杂时存在极大的不稳定性,表现为车辆偏离车道现象。针对此问题,文章首先在虚拟环境下利用神经网络可视化方法分析了车道偏离的原因,然后在方法上将方向盘转角序列作为神经网络输入,同时根据车道线检测的方法求出车辆所在车道的面积作为辅助任务。文末分析对比了文章方法和递归神经网络(RNN,LSTM)方法在平稳性上的差异,最后通过虚拟实验和实车实验验证文章中的方法的有效性。结果表明,本文中的方法能有效改善车辆行驶平稳性问题,和LSTM方法相比稳定性效果相近,但本方法操作应用简单,节省计算资源。

End to end self-driving has drawn increasing attention due to its simplicity and efficiency.Especially,this approach has achieved good performance in lane following.But its safety is still a concern with existing problem like sudden lane departure under complex road conditions.In view of this problem,firstly,we analyze the example accounted for such decision making problem by visualizing attention.Secondly,advocate sequences of steering angles as the input to Convolutional Neural Network(CNN),then,calculate the road area of the car in its lanes using lane detection as auxiliary task.Meanwhile,the contrast difference of stability is analyzed compared to the recurrent neural network in deep learning(RNN,LSTM).Finally,the results show that this method enhances the stability closed to LSTM and computation-ally efficient verified by virtual and real vehicle test.