【摘要】 在大数据、云计算和机器学习等新一代人工智能技术的推动下,自动驾驶的感知智能在近年来得到显著的提升与发展.然而,与人类驾驶过程中隐含的以自我目的实现为引导的自探索性和自主性相比,现阶段自动驾驶技术主要以辅助驾驶功能为主,还停留在以被动感知、规划与控制为主的初级智能自动驾驶阶段.为实现车辆智能从数据驱动的环境感知、辅助决策、被动规划到知识驱动的场景认知、推理决策、主动规划的提升,亟需增强车辆自身对复杂外界信息归纳提炼、推理决策、评价估计等类人能力.首先回顾自动驾驶关键技术演化及其应用发展历程;随后分析测试对车辆智能评估的效用;然后基于平行测试理论,提出自动驾驶车辆认知智能训练、测试与评估空间的构建方法,并设计基于平行测试的认知自动驾驶智能训练框架.该项研究工作预期能为推动自动驾驶从感知智能向认知智能的升级提供可行的技术支撑与实现路径.更多还原

【Abstract】 Driven by the new generation of artificial intelligence technologies such as big data, cloud computing and machine learning, the perceptive intelligence of autonomous driving has been significantly improved and progressed in recent years. However, compared with the self-purpose driven human driving process, the current autonomous driving technologies are mainly focusing on the auxiliary driving functions, and still stay in a primary-intelligence stage which is dominated by passive perception, planning and control. In order to cross the “cognition gap” of vehicle intelligence, from data-driven environment perception, assisted decision making and passive planning to knowledge-driven scenario cognition, reasonable decision making and active planning, it is important to enhance the humanoid abilities of the vehicles, including but not limited to summarize and extract complex external information from the environment, reasoning, evaluation and estimation. This paper reviews the evolution and application of key technologies in autonomous driving, analyzes the effectiveness of testing on vehicle intelligence and performance evaluation. After then, based on the parallel test theory, it puts forward the space construction method for training, testing and evaluating the cognitive intelligence of autonomous vehicle, and proposes an intelligent training framework for cognitive autonomous driving. The work is expected to provide a feasible and possible path for autonomous vehicle cognitive intelligence.更多还原