智能汽车路径规划与自主决策理论及关键技术

Intelligent vehicles (IVs) are the trend and core of traffic development in future due to the great potential of easing road congestion, improving traffic safety and ride experience, and even very likely to change the mode of human travel. This project aims to realize the "human-like driving brain" of IV on the basis of environmental perception ability. The research follows the ideas of the cross-disciplinary "vertical integration" and the "horizontal collaboration" of the enterprises, universities and research institutes by playing to the strengths of ground traffic discipline chain of the university. First of all, the global path planning method is established by deep learning of massive trajectory data of commercial passenger vehicles, in which the intelligent level of IV and the knowledge of experienced drivers are considered. In addition, based on the natural driving big data which is collected for more than 3 years reaching a mileage of over 500,000 km and self-developed microscopic traffic simulation system (which already has the driver behavior decision-making ability based on the logical model), integrated multi-dimensional deep learning is used for realizing decision-making behavior and local path planning of IV. Furthermore, based on the cognitive theory, the relationship between the physiological factors behind the driver's behavior and the synergistic effect of the human, vehicle, road and environment is explored to achieve a high degree of anthropomorphic driving. Finally, fast and full life-cycle test is conducted by the accelerated testing method in the test platform which is developed by the team. The project is going to provide strong support for the further application of the IV technology.

由于在解决交通安全、道路拥堵以及改善驾乘体验上具有巨大潜力，甚至极有可能改变人类的出行模式，智能汽车（IV）将是未来交通的发展趋势及核心。课题拟在IV具备环境感知能力基础上，以实现IV“类人驾驶脑”为目标，发挥本校地面交通学科链优势，按照跨学科“纵向整合”、产学研“横向协同”思路展开研究。课题首先利用海量商用车运行轨迹数据进行深度学习，建立考虑IV智力水平、经验驾驶人知识的全局路径规划方法；然后利用已开展3年有余，行驶超50万公里的自然驾驶大数据和自主研发的微观交通仿真系统（已具备基于逻辑模型的自主决策能力）对IV决策行为和局部路径规划进行深度多维集成学习；进一步基于认知行为理论，探究驾驶人行为动作背后的心生理影响因素和人车路环境的协同作用，以实现高度拟人化自主驾驶。最后利用团队研发的IV测试平台和加速测试方法对其进行快速、全生命周期测试。课题研究将为IV技术深度应用提供强有力支撑。

由于在解决交通安全、道路拥堵以及改善驾乘体验上具有巨大潜力，甚至极有可能改变人类的出行模式，智能汽车将是未来交通的发展趋势及核心。课题在智能汽车具备环境感知能力基础上，以实现智能汽车“类人驾驶脑”为目标，按照跨学科“纵向整合”、产学研“横向协同”思路展开研究。课题按自动驾驶相关任务自上而下开展了全局路径规划、自主决策、局部路径规划、驾驶人认知过程与拟人化驾驶建模、基于场景的自动驾驶虚拟测试方法等5大研究内容。. 在全局路径规划方面，现有研究无法做出综合考虑交通流复杂性及风险规避的路径推荐，课题综合考虑先验知识、交通分配以及自动驾驶“智力水平”提出最优全局路径规划方法。在自主决策方面，针对我国道路交通流具有混合交通、高密度交互、路权不明确以及违章现象突出等典型特征，基于数据驱动和多模型驱动两种方法分别构建了复杂交通流环境下智能汽车自主决策方法，并通过深度强化学习方法融合以上两种方法的优势，提升智能汽车在复杂环境下的自主决策能力。由于车辆实时运动状态、外部环境都在持续更新，且外部环境具有不可预知等特性，课题提出在考虑避障前提下的在线实时轨迹规划方法，实现了多约束条件、多优化目标下的在线实时局部路径规划。进一步，基于认知行为理论，探究驾驶人行为动作背后的心生理影响因素和人车路环境的协同作用，提升拟人化自主驾驶能力。最后系统构建了基于场景的自动驾驶虚拟测试与评价理论、方法和技术体系，突破场景智能化生成、加速测试、测试工具链、驾驶能力评价等全栈技术，形成场景高覆盖、测试可加速、工具可协同、评价可互信的自动驾驶虚拟测试仿真服务系统。. 课题注重研究成果落地与应用，不仅将上述理论方法用于课题组自主研发的智能汽车中，也大力推动产学研合作合，与百度、华为、腾讯、上汽、宇通等行业领军企业开展自动驾驶技术合作，为智能汽车规划决策技术深度应用提供强有力技术支撑。