基于异质边增强时空图注意力网络的自动驾驶换道轨迹规划  

Heterogeneous Edge-enhanced Spatial-temporal Graph Attention Network for Autonomous Driving Lane-changing Trajectory Planning

作　　者：董晴 中野公彦 杨波 季学武[1] 刘亚辉[1] DONG Qing;NAKANO Kimihiko;YANG Bo;JI Xue-wu;LIU Ya-hui(School of Vehicle and Mobility,Tsinghua University,Beijing 100084,China;Institute of Industrial Science,University of Tokyo,Tokyo 153-8505,Japan)

机构地区：[1]清华大学车辆与运载学院,北京100084 [2]东京大学生产技术研究所,东京153-8505

出　　处：《中国公路学报》2024年第3期147-156,共10页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52221005);国家留学基金委项目(202206210211)。

摘　　要：轨迹规划仍然是自动驾驶技术大规模应用所面临的关键难题之一。例如,自动驾驶中的换道轨迹规划算法通常被构建为一个针对代价函数的优化过程。然而,为适应多样化的交通场景而手动调整代价函数中的特征权重,是一项极具挑战性的任务。针对这一问题,提出了一种基于异质边增强时空图注意力网络(Heterogeneous Edge-enhanced Spatial-Temporal Graph ATtention network, HEST-GAT)的新型换道轨迹规划方法。首先,采用逆强化学习技术,从大量专家级换道示范中提取代价函数的特征权重向量,构建了一个专家级换道示范数据集。然后,将交通场景构建为一个异质有向图,其中,交通参与者的位置定义为节点属性,交通参与者间的相对位置作为边属性,交通参与者之间的关联类型则定义为边类型。边的属性和类型组合,形成了边的特征表示。为捕获交通场景中的空间和时间信息,采用HEST-GAT网络进行特征提取,并计算了各场景下代价函数的特征权重。接着构建了一个结合轨迹特征和特征权重的代价函数,并通过优化过程生成最终的换道轨迹规划。为验证所提出方法的实用性,在真实驾驶数据集上进行了多轮换道轨迹规划测试和评估。研究结果显示:与基于时空图卷积网络方法相比,基于HEST-GAT的换道轨迹规划在模拟专家示范轨迹方面的误差显著减少。纵向舒适性、纵向效率、横向舒适性和安全性4项关键指标的误差分别降低了5.5%、5.4%、1.4%和6.0%。结果表明提出的方法能够生成与人类驾驶行为高度一致的换道轨迹,具有卓越的场景适应能力。Trajectory planning remains one of the key challenges in the large-scale application of autonomous driving technology.For instance,in autonomous driving,lane-changing trajectory planning algorithms are typically built as an optimization process targeting the cost function.However,manually adjusting the feature weights in the cost function to suit diverse traffic scenarios is a highly challenging task.To address this issue,our study proposed a new lane-changing trajectory planning method based on the heterogeneous edge-enhanced spatial-temporal graph attention network(HEST-GAT).Initially,we employed inverse reinforcement learning techniques to extract feature weight vectors of the cost function from a multitude of expert lane-changing demonstrations,thereby constructing an expert-level lane-changing demonstration dataset.Subsequently,traffic scenarios were modeled as heterogeneous directed graphs,where the locations of traffic participants were defined as node attributes,their relative positions as edge attributes,and the types of connections between them as edge types.These attributes and types were combined to form the edge feature representation.To capture the spatial and temporal information within traffic scenes,we utilized the HEST-GAT network for feature extraction,calculating the feature weights of the cost function for each scenario.We then constructed a cost function that integrates trajectory features and feature weights,generating the final lane-changing trajectory plan through an optimization process.To validate the practicality of our proposed method,multiple rounds of lane-changing trajectory planning tests and assessments were conducted on real driving datasets.The results demonstrate that,in comparison to spatial-temporal graph convolutional network methods,lane-changing trajectory planning based on HEST-GAT significantly reduces errors when emulating expert demonstration trajectories.Specifically,errors in longitudinal comfort,longitudinal efficiency,lateral comfort,and safety are reduced by 5.5%

关 键 词：汽车工程 换道轨迹规划 逆强化学习 图注意力网络 代价函数权重 自动驾驶 

分 类 号：U461.6[机械工程—车辆工程]