基于人类驾驶知识的无人驾驶车辆智能决策系统研究

At present, the decision-making capability of unmanned ground vehicle (UGV) has made a lot of technology progresses, but in real and complex driving scenarios it is still difficult to be comparable with the human drivers. To solve these problems with UGV, such as the decision-making ability is low and the decision system cannot adapt to the unknown environment, the project proposes to develop an intelligent decision system for UGV based on human’s driving knowledge, such as traffic rules, human driving experience, etc. The project mainly focuses on four aspects to develop the intelligent decision system: knowledge extraction, knowledge modeling and expression, knowledge reasoning and knowledge sharing. To extract the traffic rules knowledge from traffic laws, regulations and related documents, the analysis methods are used; and deep learning method is applied to extract human’s driving experience knowledge from human’s driving behaviors in different driving scenarios. A semantic hypergraph is proposed for knowledge modeling and expression to build driving knowledge base. To obtain decision information more efficiently, a multi-reasoning mechanism based cooperative reasoning method is developed, which can transform the UGV’s perception information into decision outputs. And to solve the learning limits of single UGV, a cloud computing based UGV cloud service platform is proposed to provide knowledge sharing services among UGVs. Through studying of the above four key technologies, the project is expected to develop an intelligent decision system for UGV, which can provide more reasonable and safer decision information to UGV.

近年来，无人驾驶车辆决策技术已经取得了一定的研究成果，但是在真实复杂行驶环境下其自主决策能力仍然难以与人类驾驶员相媲美。针对目前无人驾驶车辆智能决策水平低、难以适应未知环境等问题，本项目拟建立基于交通规则以及人类驾驶经验等驾驶知识的无人驾驶车辆智能决策系统。通过对交通法律法规的形式化描述和人类驾驶行为的深度学习，提取出交通规则知识和驾驶经验知识；利用基于语义超图的驾驶知识建模、表达和存储方法，构建无人驾驶决策系统知识库；开发基于多种推理机制协同的知识推理方法，将无人车的感知输入转化为决策输出，同时通过案例存储达到增量学习的目的；利用云计算技术，构建无人驾驶云服务平台，实现无人车之间知识的共享与集成，以达到共同进化的目的。本项目预期通过研究人类驾驶知识的提取、表达、推理以及共享等关键技术，最终实现一个无人车驾驶智能决策系统，从而为无人驾驶车辆提供更加合理、安全的决策信息。

近年来，无人驾驶车辆决策技术已经取得了一定的研究成果，但是在真实复杂行驶环境下其自主决策能力仍然难以与人类驾驶员相媲美。针对目前无人驾驶车辆智能决策水平低、难以适应未知环境等问题，本项目建立了基于交通规则以及人类驾驶经验等驾驶知识的无人驾驶车辆智能决策系统。首先基于概念分析等技术，对交通法律法规进行关键术语提取、抽象，基于本体论方法，对交通规则进行形式化描述；然后，基于SHRP2自然驾驶数据，对人类的驾驶行为进行分析，并且基于LSTM网络模型，实现基于深度学习的驾驶行为学习与建模方法；接着，构建了基于知识的无人驾驶车辆智能决策系统框架，利用语义超图技术，对驾驶环境、驾驶行为等关键概念及其关系进行建模，在此基础上，建立了包括交通规则以及驾驶经验的无人驾驶车辆行为决策知识库；开发了定性与定量相结合的知识推理方法，利用产生式规则推理机制，实现基于知识的行为决策，利用知识推理与马尔科夫决策过程相结合的技术，实现他车行为的预测；利用云计算技术，构建基于Storm的无人驾驶实时云服务平台，实现无人车之间知识的共享与集成，以达到共同进化的目的。. 本项目借鉴人类驾驶员驾驶车辆的思维过程，研究基于人类驾驶知识的无人驾驶车辆智能决策系统，其科学意义在于为无人驾驶车辆行为决策提供通用、统一的理论基础与框架，从而进一步提高无人车的智能与自主水平。将人类驾驶知识如交通规则、驾驶经验编码并保存入知识库，具有以下优点：便于决策知识的增量更新；便于未来交通法律法规的变更；便于无人驾驶系统的移植与更新，等等。基于上述优点，基于人类驾驶知识的无人车智能决策系统具有重要的应用前景，特别是对于无人车产业化的发展具有重要的意义。