城区不确定环境下无人驾驶车辆行为决策方法研究

发布时间：2018-05-05 21:41

本文选题：无人驾驶车辆 + 行为决策　；参考：《中国科学技术大学》2017年博士论文

【摘要】：无人驾驶车辆在现代交通系统和极端危险场景下具有重大的应用价值,并且可以作为验证相关新兴技术的综合试验平台,近年来吸引了国内外广泛的关注。作为无人驾驶车辆的"大脑",行为决策系统决定了其行驶的安全性与合理性,提高行为决策系统的智能化水平一直以来都是无人驾驶领域的研究重点和难点。与高速公路等交通环境相比,城区交通环境具有高度的复杂性和不确定性:复杂性主要体现在道路拓扑结构的复杂性、道路元素和交通参与者类型的多样性、交通参与者之间以及交通参与者与道路元素之间交互的复杂性;不确定性主要体现在感知信息的不确定性、他车运动状态的难以预测性等。由于这些复杂性和不确定性,无人驾驶车辆需要其行为决策系统能够实时给出高可靠性、高安全性的行为决策结果,但目前大多数行为决策方法只能适应简单、确定性的驾驶环境,无法满足上述需求。为此,本文提出了一种基于本体论和马尔科夫理论的行为决策方法,对其中的驾驶场景建模、他车运动状态预测和驾驶动作生成方法进行了深入研究。具体研究内容如下:(1)针对驾驶场景中的多源异构信息和先验驾驶知识表达不充分所导致的行为决策准确性差的问题,提出了一种基于本体论的驾驶场景建模方法。首先,通过深入研究驾驶场景信息的层次性和交互性,提出一种考虑场景元素之间语义关联性的结构化本体模型,实现了对驾驶场景的有效建模。进而,基于驾驶场景的语义描述构建了驾驶知识库和在线推理系统,实现了先验驾驶知识的高效利用。该模型解决了传统建模方法(如栅格分解法)可理解性差、对场景元素间的关联性描述不充分的问题,克服了先验知识与推理过程高度耦合导致的知识增量式更新困难的难题,为他车运动状态预测和驾驶动作生成奠定了坚实的基础。(2)针对现有行为决策方法缺乏对他车运动状态的预测而导致的行为决策不合理的问题,提出了一种基于驾驶意图的他车运动状态预测方法。首先,以驾驶意图为突破口,提出基于隐马尔科夫模型(HMM)和知识推理的他车驾驶意图预测方法,实现了动态环境下场景自适应的他车驾驶意图预测。进而,以驾驶意图为引导信息提出一种基于三阶贝塞尔曲线的快速轨迹预测方法。最后,在考虑驾驶行为不确定性的情况下,基于轨迹预测结果对他车的运动状态进行建模。实验结果表明:所提出的预测方法增加了驾驶意图预测的预测步长和轨迹预测的实时性,提高了预测结果的准确性和模型的场景适应性,为动态环境下的驾驶动作生成提供了前提条件。(3)针对现有行为决策方法对不确定性考虑不足和知识推理能力差导致的行为决策不合理的问题,提出了一种混合部分可观察马尔科夫决策过程(POMDP)和逻辑推理的驾驶动作生成方法。首先,为了考虑感知信息的不确定性,提出一种基于POMDP概率模型的驾驶动作生成模型。针对该模型存在的知识推理能力差和"维度灾难"问题,借鉴人类驾驶员视觉行为的注意力分配机制,根据其目标驱动性和分布粒度不均匀性,构建了基于交通规则知识和驾驶导向信息的逻辑推理系统,实现了对驾驶场景的高效推理和对状态空间的有效压缩。通过概率模型和逻辑推理系统的有机协作实现了驾驶动作的生成。实验结果显示:该方法在安全性和行车效率方面优于主流的基于有限状态机的行为决策方法。本文以真实城区道路为测试环境,以"智能先锋II"无人驾驶车辆为实验平台,对所提出的行为决策方法进行了验证。实验结果表明,本研究提高了无人驾驶车辆在城区复杂、不确定环境下自主行为决策的合理性、安全性和实时性,有效促进了无人驾驶车辆智能化水平的提升。
[Abstract]:Unmanned vehicle (unmanned vehicle) has great application value in modern traffic system and extreme dangerous scene, and it can be used as a comprehensive test platform to verify the related new technologies. In recent years, it has attracted wide attention both at home and abroad. As the "brain" of unmanned vehicle, the behavior decision system determines the safety and rationality of its driving. The intelligent level of the high behavior decision system has always been the focus and difficulty of the research in the unmanned field. Compared with the highway environment, the urban traffic environment has high complexity and uncertainty: complexity is mainly reflected in the complexity of road topology, the diversity of road elements and the types of traffic participants, The complexity of the interaction between the traffic participants and the traffic participants and the road elements; the uncertainty is mainly reflected in the uncertainty of the perception information, the unpredictability of the vehicle movement, and so on. Because of these complexity and uncertainty, the unmanned vehicle needs its behavior decision system to give high reliability and high safety in real time. The result of sexual behavior decision, but most behavior decision methods can only adapt to the simple, deterministic driving environment, and can not meet the above requirements. Therefore, a behavioral decision method based on ontology and Markoff theory is proposed in this paper, in which the driving scene is built, the motion state of his car is predicted and the driving action is generated. The specific research contents are as follows: (1) a driving scene modeling method based on ontology is proposed in view of the poor accuracy of multi source heterogeneous information and poor prior driving knowledge in driving scenario. Firstly, the level and interaction of driving scene information are studied. A structured ontology model, which considers the semantic relevance between the scene elements, is proposed to achieve effective modeling of the driving scene. Then, a driving knowledge base and an online reasoning system are constructed based on the semantic description of the driving scene, and the efficient use of a priori driving knowledge is realized. The model solves the traditional modeling method (such as grid decomposition). The problem of poor comprehensibility and inadequate description of the correlation between the scene elements has overcome the difficult problem of the incremental updating of knowledge caused by the high coupling of prior knowledge and reasoning process, which lays a solid foundation for the motion state prediction and driving action generation of his car. (2) the current behavior decision method lacks the motion of his car. The problem of irrational behavior decision caused by the prediction of state of state, a method of moving state prediction based on driving intention is proposed. First, taking driving intention as a breakthrough, the driving intention prediction method based on Hidden Markov model (HMM) and knowledge reasoning is proposed to realize the adaptive scene of his car driving under the dynamic environment. Then, a fast trajectory prediction method based on the three order Bessel curve is proposed with the driving intention as the guidance information. Finally, the motion state of the vehicle is modeled based on the trajectory prediction results under the uncertainty of driving behavior. The experimental results show that the proposed prediction method increases the driving intention. The prediction step length and the real time of the trajectory prediction can improve the accuracy of the prediction results and the adaptability of the model scene, and provide the precondition for the driving action generation in the dynamic environment. (3) the problem that the existing behavior decision method is not reasonable in the uncertainty of the uncertainty and the poor ability of knowledge inference, is put forward. A hybrid part can be used to observe the driving action generation method of the Markov decision process (POMDP) and logical reasoning. First, in order to consider the uncertainty of the perceptual information, a driving motion generation model based on the POMDP probability model is proposed. The attention distribution mechanism of driver visual behavior, based on its target driving and distribution granularity inhomogeneity, constructs a logical reasoning system based on traffic rules knowledge and driving guidance information, which realizes the efficient reasoning of driving scenes and effective compression of the state space. The organic collaboration through the probability model and the logic reasoning system is carried out. The experimental results show that the method is superior to the mainstream behavior decision making method based on the finite state machine in the aspects of safety and driving efficiency. This paper takes the real urban road as the test environment and the "intelligent pioneer II" unmanned vehicle as the experimental platform, and verifies the proposed behavior decision method. The results show that the study improves the rationality, security and real-time performance of the autonomous behavior decision of unmanned vehicles in the complex and uncertain environment, which effectively promotes the intelligent level of unmanned vehicles.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：U463.6

【相似文献】