【作者】 李涛；

【导师】 潘之杰； 杨国青； 李红；

【作者基本信息】 浙江大学 ， 软件工程， 2025， 硕士

【摘要】 近年来,以高度有序化为特征的多车智能网联自动驾驶逐渐成为解决城市交通通行效率低、能耗高等问题的核心着力点。在仿真环境中对智能网联自动驾驶全过程管控、并探索不同的路网和场景对其性能的影响,对加速系统迭代优化、推动其落地具有重要意义。基于此,本文面向自动驾驶仿真,针对路网和场景生成技术展开深入研究,取得诸多创新性研究成果。本文主要内容和贡献如下:(1)在路网生成方面,针对数据集缺乏、现有方法生成路网拓扑连通性差、难以同时建模空间位置和拓扑特征的问题,本文构建了包含72400组样本(覆盖范围达800006)8)~2)的路网数据集,创新地提出了基于拓扑感知增强的路网生成模型。该模型包含以Transformer为核心的生成对抗网络和基于图卷积的奖励子网络,从全局拓扑层面对生成过程监督,以生成满足空间位置约束、期望拓扑属性的路网。(2)在场景生成方面,针对数据集缺乏、现有生成方法难以基于给定边界和属性进行可控生成的问题,本文构建了包含461666个子区域的场景布局数据集,提出了基于条件变分学习的图动态注意力模型。该模型将场景边界和布局属性编码为条件向量,并使用虚拟结点以增强属性编码在条件控制中的强度,实现场景平面图的可控生成。此外,本文提出基于CLIP模型的三维矢量场景生成策略,以生成可供车载传感器感知的路侧场景。(3)为支持路网和场景在运行过程中迭代优化,本文实现多车自动驾驶仿真运行系统。基于该仿真系统,在考虑全天出行需求的前提下,本文定量分析了路网对车辆运行效率的影响、场景对自动驾驶感知算法的影响,并为算法优化提供训练数据,以实现运行在环、数据闭环的迭代优化。实验结果表明,本文方法所生成的路网交通便捷性高、场景多样性强,所实现的仿真系统能有效支持路网运行和场景感知,为自动驾驶感知算法提供优化方向。更多还原

【Abstract】 In recent years,multi-vehicle intelligent connected autonomous driving,characterized by high-level organization,has gradually become a key focus for addressing issues such as low urban traffic efficiency and high energy consumption.Conducting comprehensive control of the process within simulation environments and exploring the impact of different road networks and scenes on its performance is of significant importance for accelerating system iteration and optimization.Based on this,this paper focuses on autonomous driving simulation and conducts in-depth research on road network and scenario generation technologies,achieving numerous innovative research outcomes.The main contents and contributions of this paper are as follows:(1)Road Network Generation.In response to the lack of datasets,poor connectivity of generated road networks by existing methods,and the difficulty in simultaneously modeling spatial and topological features,this paper constructs a road network dataset containing 72,400 samples(covering an area of 80,0006)8)~2).An innovative topology-aware enhanced road network generation model is proposed,which includes a Transformer-based generative adversarial network and a graph convolution-based reward subnetwork.This approach supervises the process from a global topological perspective to generate road networks that meet spatial constraints and desired topological attributes.(2)Scene Generation:Addressing the lack of datasets and the challenge of controllable generation based on given boundaries and attributes,this paper builds a scene planar map dataset comprising 461,666 city blocks.A graph dynamic attention model based on conditional variational learning is proposed.This method encodes scene boundaries and layout attributes into conditional vectors and uses virtual nodes to enhance the strength of attribute encoding in conditional control,enabling controllable generation of scene planar maps.Additionally,a 3D vector scene generation strategy based on the CLIP model is proposed to generate roadside scenes perceivable by vehicle sensors.(3)To support iterative optimization of road networks and scenes during operation,this paper implements a multi-vehicle autonomous driving simulation operation system.Based on this simulation system,this paper quantitatively analyzes the impact of road networks on vehicle operation efficiency and the influence of scenes on autonomous driving perception algorithms,considering travel demand throughout the day.It also provides training data for algorithm optimization,achieving loop-closure in operation and data for iterative optimization.Experimental results demonstrate that the road networks generated by the proposed method exhibit high traffic accessibility and diversity in scenes.The implemented simulation system effectively supports road network operation and scenario perception,providing an optimization direction for autonomous driving perception algorithms.更多还原