耦合深度学习-运动学的自动驾驶一体化换道研究

Lane Changing Model Coupling Deep Learning and Kinematics for Autonomous Vehicles

现阶段主流算法都面临着数据量不足导致的模型无法完全应对所有场景,从而造成换道失败或者影响换道过程中的安全性。提出了一种新的耦合深度学习-运动学的自动驾驶换道一体化模型。该模型基于经典BP神经网络,对神经元中的激活函数进行替换,从而克服已有模型的不足。通过CarSim仿真来研究了本模型产生的自动驾驶车辆换道轨迹的可跟踪性,结果显示本模型可以更好地学到人类的安全换道行为,可推广使用。

The mainstream algorithms are faced with the problem that the model could not cope with every driving scenarios for lane change due to insufficient data. This paper proposes a novel lane-changing model coupling deep learning and kinematics for automatic driving. Based on the classical BP neural network, the activation function in neurons is replaced to improve the model performance. The trajectories generated by this model are simulated by using CarSim. The results show that the smooth lane changes are achieved efficiently and safely. The trajectory output of this model can be widely used in automatic driving.