面向自动驾驶离线学习的多场景驾驶数据集构建
Multi-Scenario Driving Dataset for Offline Learning in Autonomous Driving
谢天乐 1张月静 1王越1
作者信息
- 1. 辽宁石油化工大学信息与控制工程学院，抚顺，113001
- 折叠
摘要
为满足自动驾驶离线学习与仿真实验对高质量、多场景、可复现轨迹数据的需求，本文围绕CARLA仿真平台下的城市场景数据集构建开展研究。首先，选取Town10HD地图中的直线定速巡航、行人横穿马路和红绿灯路口通行三类典型驾驶场景，分别明确任务目标、失败条件和采集边界。其次，通过固定路线筛选、候选route预览、场景可重复化设计和规则专家控制器设计，生成包含成功、临界和失败三类质量层次的离线驾驶轨迹。针对行人横穿场景，本文在固定直线路线基础上参数化生成横穿事件，并记录让行、停车、安全距离和碰撞等行为信息；针对红绿灯路口场景，构建停止线推断、交通灯关联和红/黄/绿灯状态机，使数据覆盖红灯等待与绿灯通行两类典型路口行为。最后，对不同来源的episode级数据进行统一字段提取、标签归一化、索引组织和训练/验证/测试集划分，形成面向后续离线策略学习使用的标准化多场景离线驾驶数据集。结果表明，所构建的数据集在三类场景下均具有明确的任务语义、稳定的路线基础和可追踪的数据结构，能够为自动驾驶场景建模、轨迹分析以及后续算法训练与对比实验提供可靠的数据支撑。
Abstract
To meet the demand for high-quality, multi-scenario, and reproducible trajectory data for autonomous driving offline learning and simulation experiments, this paper focuses on the construction of an urban scene dataset under the CARLA simulation platform. Firstly, three typical driving scenarios, namely straight-line constant speed cruise, pedestrian crossing, and traffic light intersection passing, are selected from the Town10HD map. The task objectives, failure conditions, and collection boundaries are clearly defined for each scenario. Secondly, through fixed route screening, candidate route preview, scene reproducibility design, and rule-based expert controller design, offline driving trajectories with three quality levels: successful, critical, and failed, are generated. For the pedestrian crossing scenario, this paper parameterizes the crossing event based on a fixed straight route, and records behavior information such as yielding, stopping, safety distance, and collision. For the traffic light intersection scenario, a stop line inference, traffic light association, and red/yellow/green light state machine are constructed to cover two typical intersection behaviors: red light waiting and green light passing. Finally, episode-level data from different sources undergoes unified field extraction, label normalization, index organization, and training/validation/test set division to form a standardized multi-scenario offline driving dataset for subsequent offline policy learning. The results show that the constructed dataset has clear task semantics, stable route foundation, and traceable data structure in all three scenarios, providing reliable data support for autonomous driving scene modeling, trajectory analysis, and subsequent algorithm training and comparative experiments.关键词
自动驾驶/CARLA/驾驶场景/离线数据集/专家轨迹/数据标准化Key words
Autonomous driving/CARLA/Driving scenarios/Offline dataset/Expert trajectories/Data standardization引用本文复制引用
谢天乐,张月静,王越.面向自动驾驶离线学习的多场景驾驶数据集构建[EB/OL].(2026-10-20)[2026-10-21].http://www.paper.edu.cn/releasepaper/content/202610-14.学科分类
公路运输工程
