基于深度神经网络的自动驾驶平行运动规划的研究

Motion planning is one of the core technologies for autonomous driving, whose purpose is to acquire the motion parameters or available paths of autonomous vechiles from the current state to the target state. How to utilize the advantages of deep neural networks and big data, build an end-to-end motion planning model from perceived data to motion parameters, and adapt the motion planning model for complicate and emergency traffic scenes are the difficult issues in the motion planning field. The project studies parallel motion planning for autonomous driving based on deep neural networks and parallel systems. In this study, an end-to-end deep cascaded motion planning model is proposed based on multiple deep neural networks. By combing the generative adversarial network and the variational auto-encoder, we propose a deep generative model of virtual emergency traffic scenes and a motion planning method for potential emergency traffic scenes. A parallel learning method for deep motion planning model is developed based on artificial traffic scenes and real traffic scenes. Big data including both artificial scenes and real scenes are used to test and analyze the proposed model. The research of this project can provide more reliable and safer motion planning schemes for autonomous driving, lay a new theoretical foundation and provide technical support for autonomous unmanned systems such as automated vehicle, mobile robot, and unmanned aerial vehicle, etc.

运动规划是自动驾驶的核心技术之一，其目的是获取从自主车辆的当前状态到目标状态的运动参数或可行路径。如何充分利用深度神经网络和大数据的优势，构建从感知数据到运动参数的端到端的运动规划模型，以及如何让规划模型能够适应复杂和紧急的交通场景是运动规划的难点。本课题以深度神经网络为理论基础，结合平行系统的思想，研究自动驾驶中的平行运动规划问题。其研究内容包括基于多种深度神经网络，提出端到端的深度级联运动规划模型；通过生成对抗网络和变分自编码器相结合的方法，提出虚拟紧急交通场景的深度生成模型，以及针对潜在的紧急交通场景的运动规划方法；基于人工交通场景和真实交通场景，探索面向深度运动规划模型的平行学习方法；利用基于人工场景和真实场景的大数据对模型进行测试和分析。本课题的研究成果不仅为自动驾驶提供更安全、可靠的规划方案，而且为无人车、移动机器人和无人机等自主无人系统提供新的理论基础和技术支撑。

运动规划是自动驾驶汽车、无人机、移动机器人、智能机械臂等自主无人系统的关键技术。如何让运动规划算法具备自主学习能力，在复杂和紧急的驾驶场景中做出有效的规划决策，提高规划的鲁棒性和泛化能力，是值得深入研究的问题。本课题以深度神经网络为理论基础，研究如何构建端到端级联运动规划模型，以及面向该模型的神经网络训练方法，其结果都达到了业界领先的效果。同时，综合利用人工场景和真实场景，构建了平行数据集，提出了针对紧急驾驶场景的平行规划新方法，以及虚拟到真实场景的翻译模型，该方法在构建的数据集和公开数据集上都取得了较好的效果。