用行车轨迹自动修复自动驾驶数据集中的交通灯状态缺失问题。
Improving Traffic Signal Data Quality for the Waymo Open Motion Dataset
- 结合车辆轨迹与交通规则，自动推断交通灯状态。
- 修复后缺失率从71.7%降至0%，红灯闯越率由15.7%降为2.9%。
- 适用于复杂路口，适合自动驾驶与交通研究者使用。
自动驾驶相关数据集在人工智能、自动驾驶和交通工程等领域具有重要价值，但常存在交通灯状态缺失或不准确的问题，影响模型性能与数据可靠性。本文提出一种全自动方法，利用车辆轨迹数据与交通领域知识，对Waymo Open Motion Dataset（WOMD）中的交通灯信息进行有效补全与修正。该方法具备强鲁棒性与灵活性，可处理真实场景中多样的交叉口几何结构与信号配置。在全部53万条驾驶场景中，针对超过36万条涉及交通灯的场景进行了全面验证。原始数据中71.7%的交通灯状态缺失或未知，均被本方法成功补全。在无真值情况下，通过车辆闯红灯率评估精度：原始数据估计闯越率为15.7%，修复后降至2.9%，显著提升数据质量。该工作大幅改善了自动驾驶数据集的可用性，推动相关研究发展。代码与改进后的数据已开源：https://github.com/michigan-traffic-lab/WOMD-Traffic-Signal-Data-Improvement。
原文摘要 · Abstract (English)
Datasets pertaining to autonomous vehicles (AVs) hold significant promise for a range of research fields, including artificial intelligence (AI), autonomous driving, and transportation engineering. Nonetheless, these datasets often encounter challenges related to the states of traffic signals, such as missing or inaccurate data. Such issues can compromise the reliability of the datasets and adversely affect the performance of models developed using them. This research introduces a fully automated approach designed to tackle these issues by utilizing available vehicle trajectory data alongside knowledge from the transportation domain to effectively impute and rectify traffic signal information within the Waymo Open Motion Dataset (WOMD). The proposed method is robust and flexible, capable of handling diverse intersection geometries and traffic signal configurations in real-world scenarios. Comprehensive validations have been conducted on the entire WOMD, focusing on over 360,000 relevant scenarios involving traffic signals, out of a total of 530,000 real-world driving scenarios. In the original dataset, 71.7% of traffic signal states are either missing or unknown, all of which were successfully imputed by our proposed method. Furthermore, in the absence of ground-truth signal states, the accuracy of our approach is evaluated based on the rate of red-light violations among vehicle trajectories. Results show that our method reduces the estimated red-light running rate from 15.7% in the original data to 2.9%, thereby demonstrating its efficacy in rectifying data inaccuracies. This paper significantly enhances the quality of AV datasets, contributing to the wider AI and AV research communities and benefiting various downstream applications. The code and improved traffic signal data are open-sourced at https://github.com/michigan-traffic-lab/WOMD-Traffic-Signal-Data-Improvement
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。