智能汽车复杂动态环境的深度层级感知与理解方法研究

Perceiving and understanding complex dynamic environment accurately and efficiently is the key technical problem for intelligent vehicles. Focusing on the diversity of traffic scenes and the lack of labelled data, the difficulty of analyzing key elements, the complex relationship between environment factors and the ambiguity of their interaction rules, this project aims to conduct the following research. 1) By investigating the cross domain perceptual enhancement theory based on collaborative transfer learning, exploit the knowledge transfer mechanism for dynamic perceptual data, extract the domain-invariant features, and perceive the complex dynamic environment information precisely from the mixed traffic flow. 2) By exploring the perceptual mechanism of deep transfer learning based on multi-task and multi-level integration, mine the coupling relations between multi-perceptual tasks and scene understanding, and analyze the key elements and their features accurately under the complex dynamic environment. 3) By understanding and predicting the traffic force between environment factors based on structural characteristic analysis, learn the interaction and evolution mechanisms between semantic targets, and advance the dynamic environment understanding and sequence prediction for restricted traffic scenes. The research of this project can help improve the perception and understanding techniques for Intelligent Vehicles in dynamic and complex environment, as well as providing the theoretical foundations and technical supports for the development of intelligent vehicles in China.

准确高效地对复杂动态场景进行感知与理解，是智能汽车必须解决的关键技术难题。针对交通环境下场景多变且有效标注数据匮乏、关键目标解析困难、交通要素关系复杂且相互作用规律不明确等问题，本项目拟开展以下研究工作：1）通过基于协同迁移的跨域感知增强理论研究，探究面向动态数据的知识迁移机理，提取跨域不变性特征，从而精确感知混合交通流下的复杂动态环境信息；2）通过基于多任务层级联合的深度迁移感知方法研究，挖掘多感知任务与场景透彻解析的耦合关系，准确解析复杂动态环境下的关键目标及其特征；3）通过基于结构特性分析的交通要素动力学理解与预测方法研究，分析语义目标间的动力学作用与演变机理，解决交通受限模式下复杂动态环境理解与预判的准确性问题。本项目的研究将推动智能汽车复杂动态环境深度解析技术的进步，为我国智能汽车发展提供重要的理论基础与技术支撑。

针对智能汽车准确高效地对复杂动态场景进行感知与理解这一技术难题，项目在感知数据增强与跨域模型迁移、关键交通目标深度层级感知、交通场景动力学预测及决策控制等三个方面形成了创新性理论或关键技术的突破。首先，针对传感数据质量退化与跨域感知带来的信息失真与感知失效问题，项目组分析了环境对传感器设备数据采集的影响机理，研究了动态交互场景下的域不变信息表征机制，提出了多维数据不完备信息预处理融合策略与跨场景域迁移模型，有效解决了强非线、强时变、强混杂场景下感知算法泛化能力不足的问题，提高了感知任务的鲁棒性；其次，在前述感知数据增强与跨域模型迁移的基础上，项目组进一步探究了差异性目标的感知局限性，挖掘了动态交通场景多层级感知任务的耦合关系，构建了多目标的精准检测、鲁棒识别框架模型及多任务学习模型，实现了复杂动态环境下交通参与目标的可信解析；最后，基于关键运动交通目标的解析结果，项目组通过创新交通目标运动序列建模与空间交互更新方法，攻克了交通目标行为意图捕获、时空制约关系重构以及历史信息重利用的问题，进而缓解了动态环境下目标行为理解的稳定性不足，并通过挖掘传感器信息与油门、方向盘和制动等控制信息间的映射关系，提出多模态多任务的端到端自动驾驶行为决策模型，有效实现对多类控制信息的精准决策。项目组积极与企业深度合作，联合攻关多项关键技术，形成了从基础理论创新、核心技术突破到工程应用落地的闭环研究体系，创新性成果对推动我国智能汽车的科技进步具有重要意义。