无人驾驶类人决策的图谱建模与迁移泛化

Aim at three problems, unclear expression of learning content, uneven learning object and weak generalization ability in the human-like decision-making for the autonomous vehicle, the key of this study is to transfer Learning of the characteristic graph, through describing the coupling relationship between driver's perception and decision behavior, on the basis of big data of driving behavior. This study mainly aims to: 1)study the interactive dynamic influence of perceptual map, strategy of decision diagrams and behavior planning spectrum. 2)propose the similarity distance measure of the characteristic graph, using pairwise constrained semi-supervised clustering algorithm to find the ideal "perfect driver". 3)propose transfer learning of the characteristic graph. According to the sharing characteristics of the domain, the graph of source samples (perfect driver) is asymptotically converged using fine-grained correlation and then is transferred to the target sample(the autonomous vehicle).Finally, a learning system, considering the whole process of the depiction of the decision map (sharpening the content) - perfect driver extraction (perfecting the object) - graph transferring learning (improving the generalization ability), is formed. It will lay the foundation of the popularization and application of the autonomous vehicle' learning, and further expect that the autonomous vehicle would have excellent driving quality, flexible decision-making ability, high driving skills, and green driving behavior. It can be of great significance for building a healthy, harmonious, efficient and green driverless traffic environment.

针对目前无人驾驶“类人决策”中学习内容表达不清晰、学习对象参差不齐、学习泛化能力弱三个问题，以驾驶行为大数据为背景，以驾驶员感知-决策-行为的隐性耦合关系刻画为特点，以图谱迁移学习为核心，重点研究特征图谱建模表征中注意感知图、策略决策图和行为规划谱的交互式动态影响；提出图谱相似性距离测度，采用成对约束半监督聚类算法发现理想的“完美驾驶员”；提出图谱迁移学习算法，根据域空间共享特征对源样本（完美驾驶员）的决策图谱进行细粒度渐近关联聚合将其学习迁移到目标样本（无人驾驶）。最终形成涵盖决策图谱刻画（内容清晰化）-完美驾驶员提取（对象完美化）-图谱迁移学习（提升泛化能力）全过程的学习体系，为无人驾驶类人学习的推广应用奠定基础，进而期望无人驾驶具备优良的驾驶素质、灵活的决策能力、高超的驾驶技能和绿色的驾驶行为，对营造健康、和谐、高效、绿色的无人驾驶交通环境具有重要意义。

“类人学习”是提高无人驾驶行车安全性、操作顺畅性、模型泛化性的有效途径。项目针对目前无人驾驶“类人学习”中学习内容表达不清晰、学习对象参差不齐、学习泛化能力弱三个问题，按照无人驾驶“学什么——向谁学——怎么学”为主线展开研究。提出了驾驶人感知-决策-操作的特征图谱表征法，采用半监督和分段聚合法对驾驶人换道过程的扫视行为进行建模，构建了驾驶人注意感知图；采用深度神经网络和CART决策树组合算法提取换道决策规则，构建了驾驶人策略决策图；基于两点视觉理论提出视觉远引导区引导换道轨迹规划，构建了换道过程的决策规划谱，明确了无人驾驶的学习内容。采用双向门控循环(BiGRU)网络对车辆换道轨迹进行评价分类；基于换道轨迹特性、决策规划谱、换道评价图谱等构建不同类别的种子集和成对约束；提出种子-成对约束K-medoids半监督聚类算法对换道轨迹进行聚类，提取出“完美驾驶人”的优秀换道轨迹，为无人驾驶树立了学习榜样。研究了基于类别分布与相似性的源样本选择机制，采用改进的最长公共子序列算法计算源样本与目标样本之间的相似度，提出了小样本情况下Tradaboost-LSTM雾天跟驰模型；基于图谱特征提出学习驾驶人的视觉远引导区确定无人驾驶换道终点的搜索区域，建立了自由流下无人车换道轨迹的类人规划模型；提出了自由流下无人驾驶换道轨迹规划到稳定流下换道轨迹规划的迁移泛化模型，从而提升了无人驾驶的泛化能力。最终形成了涵盖“特征图谱刻画-优秀换道轨迹提取-泛化迁移”全过程的学习体系，为无人驾驶类人学习的推广应用奠定了基础，进而期望无人驾驶具备优良的驾驶素质、灵活的决策能力、高超的驾驶技能和绿色的驾驶行为，对营造安全、高效、绿色的无人驾驶交通环境具有重要意义。