| 摘要: |
| 深度学习的快速发展带动着自动驾驶技术的迅速进步. 深度学习感知模型在识别准确率逐步提升的同时, 也存在鲁棒性和可靠性不足等隐患, 需要在大量场景下进行充分测试以确保达到可接受的安全标准. 基于场景的仿真测试是自动驾驶技术的核心和关键, 如何描述和生成多样化仿真测试场景是需要解决的关键问题之一. 场景描述语言能够描述自动驾驶场景并在虚拟环境中实例化场景获取仿真数据, 但现有的场景描述语言大都缺少对于场景道路结构的高层抽象和描述. 提出路网属性图来表示路网中抽象出的实体及他们的关系, 并设计能简洁描述场景路网结构的语言SceneRoad. SceneRoad 可以基于描述的场景道路结构特征构建路网特征查询图. 这样, 在路网中搜索符合描述的场景道路特征的问题被抽象为路网图上的子图匹配问题, 该问题可用VF2算法求解. 进一步地, 将SceneRoad作为扩展集成到Scenic场景描述语言中. 使用拓展后的语言随机生成大量多样的静态场景并构建仿真数据集. 仿真数据集的统计信息表明生成的场景具有丰富的场景多样性. 不同感知模型在真实和仿真数据集上的训练测试结果表明, 模型在两个数据集上的表现呈正相关, 意味着模型在仿真数据集上的评估结果与真实情况下的表现具有一致性, 这对于感知模型的评估以及提升模型鲁棒性和安全性相关研究有重要意义. |
| 关键词: 自动驾驶 深度学习 仿真测试 场景描述语言 子图匹配 |
| DOI：10.13328/j.cnki.jos.006877 |
| 分类号: |
| 基金项目:科技创新2030 —“新一代人工智能”重大项目(2018AAA0100500); 国家自然科学基金(62272434); 安徽省重点研究与开发计划标准化专项(202104h04020039) |
|
| Scenario Description Language of Autonomous Driving Embedded with Road Network Graph Model |
|
GONG Lei1, SUN Xin-Yu2, ZHANG Yu1,2, ZHANG Yan-Yong1, JI Jian-Min1, HUA Bei1
|
|
1.School of Computer Science and Technology, University of Science and Technology of China, Hefei 230027, China;2.Institute of Advanced Technology, University of Science and Technology of China, Hefei 230094, China
|
| Abstract: |
| The development of deep learning technology has driven the rapid progress of autonomous driving. While the accuracy of perception models based on deep learning is gradually improved, hidden dangers related to robustness and reliability still exist. Therefore, tests should be conducted thoroughly under various scenes to ensure acceptable security levels. Scene-based simulation testing is crucial in autonomous driving technology. One key challenge is to describe and generate diversified simulation testing scenes. Scenario description languages can describe autonomous driving scenes and instantiate the scenes in virtual environments to obtain simulation data. However, most existing scenario description languages cannot provide high-level abstractions and descriptions of the road structure of the scene. This study presents a road network property graph to represent the abstracted entities and their relations within a road network. It also introduces SceneRoad, a language specifically designed to provide concise and expressive descriptions of the road structure in a scene. SceneRoad can build a road network feature query graph based on the described road structure features of a scene. Thus, the problem of searching the road structures in the road network is abstracted as a subgraph matching one on the property graph, which can be solved by the VF2 algorithm. Additionally, SceneRoad is incorporated as an extension into the Scenic scenario description language. With this extended language, a diverse set of static scenes are employed to build a simulation dataset. Statistical analysis of the dataset indicates the wide variety of scenes that have been generated. The results of training and testing various perception models on both real and simulated datasets show that the model’s performance on the two datasets is positively correlated, which shows that the model’s evaluation on the simulated dataset aligns with its performance in real scenes. This is significant for perception model evaluation and research on improving the model’s robustness and safety. |
| Key words: autonomous driving deep learning simulation testing scenario description language subgraph matching |
