基于端到端深度学习的智能车自动转向研究

Research on automatic steering of intelligent vehicle based on end-to-end deep learning

为解决由图像直接计算出控制量的端到端深度学习算法中感知器和控制器难以区分的问题,对其网络结构进行了改进。通过预训练一个自编码器,得到良好的道路特征编码后,将编码器作为感知器和和转角预测控制器一起进行端到端的训练。训练结果表明,改进后的自动转向网络模型收敛得更快,预测的角度在测试集上能较好地跟随实际角度变化而变化。利用解码器和特征图反向传播法分别还原出道路图片,可视化了该自动转向模型重点关注的道路特征。

In order to distinguish the perceptron and controller of end-to-end automatic steering models,this paper improved the structure of the deep learning network.By pre-training an autoencoder and getting a good road feature after encoding,it used the encoder as a perceptron and trained with the steering wheel angle prediction controller model by end-to-end approach.The training results show that the improved automatic steering network model converges faster,and the predicted angle can change well with the change of the actual angle.By using the decoder and the feature map reverse propagation method respectively,this paper restored the road pictures and visualized the road characteristics that the automatic steering model focused on.