基于多源传感器与导航地图的端到端自动驾驶方法  

End-to-end autonomous driving method based on multi-source sensor and navigation map

作　　者：朱波 张纪伟 谈东奎 胡旭东 ZHU Bo;ZHANG Jiwei;TAN Dongkui;HU Xudong(Automotive Research Institute,Hefei University of Technology,Hefei 230009,China;Intelligent Manufacturing Institute,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]合肥工业大学汽车工程技术研究院,合肥市230009 [2]合肥工业大学智能制造技术研究院,合肥市230051

出　　处：《汽车安全与节能学报》2022年第4期738-749,共12页Journal of Automotive Safety and Energy

基　　金：安徽省科技重大专项(201903a05020016);安徽省重点研究与开发计划(202004b11020002);安徽省自然科学基金(2208085QE153)。

摘　　要：提出了一种基于多源传感器与导航地图的多端输入单端输出(端到端)自动驾驶决策控制模型,以弥补现有端到端自动驾驶方法中基于深度神经网络(DNN)的PilotNet模型在主动避障行驶和交叉路口通行方面的不足。该模型的传感器数据输入端包括:单目前视摄像头、360(°)多线激光雷达(LiDAR)所得二维俯视图、精准定位的局部导航地图等3部分;车辆控制命令输出端为方向盘转向角。进行了多工况仿真和实车试验。结果表明:与PilotNet模型相比,该模型的方向盘转向角均方根误差(RMSE)值下降了37%;因而,该模型具备主动避障和交叉路口通行的能力。This paper proposed a multiple-input single-output(end-to-end) autonomous driving decision control model based on multi-source sensors and navigation map to make up for the shortcomings of existing end-toend autonomous driving methods in active obstacle avoidance and passing through intersection by using the PilotNet model based on Deep Neural Network(DNN). The sensor data in model’s input end consisted of three parts: a monocular front view camera, a 2-D top view obtained by 360° multi-layer Light-Laser Detection and Ranging(LiDAR), and a local navigation map based on accurate positioning. The model’s output end generated the steering wheel angle, which was a vehicle control command. Some multi-condition simulations and real vehicle tests were conducted. The results show that the root mean square error(RMSE) of steering wheel angle decreases by 37% compared with that by using the PilotNet model. Therefore, the model has the ability of active obstacle avoidance and intersection passing.

关 键 词：自动驾驶车辆 端到端自动驾驶 神经网络 摄像头 导航地图 激光雷达(LiDAR) 

分 类 号：U467.13[机械工程—车辆工程]