基于时空特征融合的端到端无人车控制

基于深度学习的端到端车辆控制器多是由二维卷积神经网络(2D convolutional neural netuark,2D CNN)训练得到的,因未考虑时间维度上的帧间运动信息,使得控制器的可解释性与泛化能力较差,而三维卷积神经网络(3D CNN)可以从连续视频帧中学习时空特征。深度确定性策略梯度强化学习(depth deterministic policy gradient,DDPG)常用于连续动作的控制优化,但DDPG算法仍存在采样方式不合理而导致的样本利用率低的问题。基于此,采用3D CNN与改进DDPG算法相结合的方法对车辆方向盘转角和速度进行预测。通过实车实验实现了车辆在所设置轨道上的自主驾驶,为基于深度学习和强化学习方法解决自动驾驶中的端到端控制问题提供了科学方法。

End-to-End unmanned vehicle control based on spatio-temporal feature learning

The end-to-end vehicle controller based on deep learning was mostly trained by 2D convolutional neural network (2D CNN). The controller performance is general because the inter-frame motion information in the time dimension is not considered. The 3D CNN can learn the spatio-temporal features from continuous video frames and achieve better results. Depth deterministic policy Gradient reinforcement learning (DDPG) was often used for control optimization of continuous motion. However, the DDPG algorithm still has the problem of low sample utilization caused by unreasonable sampling mode. Therefore, the combination of 3D CNN and improved DDPG algorithm is used to predict the steering angle and speed of the vehicle. Through the actual vehicle experiment, the autonomous driving on the set track was realized. It provides a scientific method for solving end-to-end control problems in autonomous driving based on deep learning and reinforcement learning methods.