Abstract: Human-imitative autonomous driving based on deep neural networks provided a new idea for advanced autonomous driving under complex environments. However, it was difficult for upper driving instructions to enter the network due to the closeness of the network. Therefore, a switched deep neural network was designed, which consisted of convolutional neural network layers and long short-term memory network layers. A feature selection network layer was embedded in the two sub-networks, and then different feature branches were activated by different driving instructions to complete the corresponding driving tasks. Vehicle test results show that the switched deep neural network does not greatly increase the inference time and the different driving tasks will be accomplished according to the different driving instructions, such as the left-turn, right-turn, straight-go at the intersection and obstacle-bypass in the roads.

Key words: switched deep neural network, human-imitative, autonomous driving, decision-making algorithm

摘要： 基于深度神经网络的拟人化自动驾驶为复杂环境下的高级自动驾驶提供了新的思路，但网络的封闭性会导致上层驾驶指令很难进入网络，为此，设计了一种开关式深度神经网络，该网络由卷积神经网络层和长短期记忆网络层组成，在两个子网络间嵌入了具有开关性质的特征选择层，并通过不同的驾驶指令选择激活不同的特征分支，从而完成了相应的驾驶任务。实车测试结果表明，开关式深度神经网络不会显著增加模型的推理时间，同时该网络实现了根据输入的不同驾驶指令完成车辆在路口的左转、右转、直行和绕障行为。

关键词: 开关式深度神经网络, 拟人化, 自动驾驶, 决策算法