基于深度学习的虚拟到现实车道保持控制

Deep-Learning-Based Lane-Keeping Control Framework: From Virtuality to Reality

深度学习由于其强大的非线性拟合能力,已经被广泛应用于无人驾驶控制器训练领域.然而,由于其训练过程需要大量标注数据,耗费大量人力物力,且人为采集的数据很难覆盖危险工况,导致训练的模型泛化能力较差,影响了深度学习控制器的性能提升.本研究提出一种从虚拟世界采集样本,将训练模型向真实世界泛化的端对端卷积神经网络(CNN)控制器训练框架.为缩小虚拟和真实世界的差距,本研究以语义分割图像作为媒介,将虚拟和真实图像分别转化为语义分割图像用于训练和测试.结果表明,虚拟到现实训练得到的控制器可以较好地跟随道路变化趋势,经权值微调后预测输出与人类驾驶员操作相近,最大平均绝对误差和均方根误差分别为1.693 9°和2.885 0°,平均绝对百分比误差在5%以内.

Deep-learning-based approaches have been widely used for training controllers for autonomous vehicles due to their powerful ability to approximate nonlinear functions or policies.However,the training process usually requires large amount of labeled real-world driving dataset,and the data collection and labeling process is always time and money-consuming.Corner-case data collecting is difficult for human,which may lead to poor generalization ability of the training model and hamper the promotion of the deep learning controller performance.Thus a framework was proposed to train an end-to-end convolutional neural network(CNN) controller in virtual game engine TORCS and generalized the trained controller into reality.To narrow the gap between virtual and real environment,semantic segmentation was used as medium.The training dataset from game engine and the testing data from reality are all translated into semantic segmentation style for both training and validation.Validation result shows that the trained lane keeping controller can be deployed to solve the real-world driving task and get a good performance.Compared with human driver operation,the changing trends of the fine-tuned CNN controllers output is consistent with human driver,the maximum MAE and RMSE are 1.693 9°and 2.885 0°,and the average MAPEs are all under 5%.