基于改进PilotNet的端到端无人车控制

Based on improved PinlotNet for end-to-end control of unmanned vehicle

端到端的驾驶决策目前是无人驾驶领域的研究热点之一。针对一款无人小车,首先使用一个9层的卷积网络PilotNet对车辆进行横向控制,纵向的前进速度为一个恒定值,在训练的过程中发现该网络存在过拟合的现象。在此基础上,设计了一个新的深度学习网络用于提取图像中的特征,并将预测的结果应用于小车的纵向和横向运动控制。进行训练之前,遥控无人小车沿地图上标志的路径行进,将车载摄像头获得的图像数据以及该时刻遥控器对应的前进速度和转向角给定值记录下来,作为观察值,将所设计的网络输出作为预测值,使用MSE作为目标函数,以便进行训练,所设计的网络参数较PilotNet减小了25%。最终,通过在验证集上的对比发现,所设计的网络较PilotNet误差更小,将设计的网络部署在小车上,通过实验发现,采用速度调节的小车,比未采用速度调节的小车跑完跑道的时间少5 s,表明所建立的模型具有较好的控制效果。

End-to-end driving decision-making is currently one of the research hotspots in the field of unmanned driving.For an unmanned car,a 9-layer convolutional network PilotNet is first used to control the vehicle horizontally,and the longitudinal forward speed is a constant value.During the training process,it is found that the network has an over-fitting phenomenon.On this basis,a new deep learning network is designed to extract the features in the image,and the predicted results are applied to the longitudinal and lateral motion control of the trolley.Before training,the remote-controlled unmanned car travels along the path marked on the map,and records the image data obtained by the on-board camera and the given value of the forward speed and turning angle corresponding to the remote control at that moment,as the observation value,and the designed value.The output of the network is used as the predicted value,and MSE is used as the objective function for training.Finally,through the comparison on the verification set,it is found that the designed network has a smaller error than the PilotNet.The designed network is deployed on the car.It is found through experiments that the car with speed adjustment is better than the car without speed adjustment.The time to finish the track is less than 5 s,which indicates that the established model has a better control effect.