自动驾驶车辆换道意图识别研究现状  

Current status of lane change intention recognition for autonomous vehicles

作　　者：方华珍 刘立[1] 顾青[1] 肖小凤 孟宇[1] FANG Huazhen;LIU Li;GU Qing;XIAO Xiaofeng;MENG Yu(School of Mechanical Engineering,University of Science and Technology Beijing,Beijing 100083,China)

机构地区：[1]北京科技大学机械工程学院,北京100083

出　　处：《工程科学学报》2024年第10期1845-1855,共11页Chinese Journal of Engineering

基　　金：国家自然科学基金青年科学基金资助项目(52202505);国家重点研发计划资助项目(2019YFC0605300)。

摘　　要：近年来基于数据驱动的自动驾驶车辆换道意图识别研究取得了显著进展,学者们发布了大量的研究成果.针对该领域面临的一些共性的技术挑战,如换道过程的认定、换道标签的缺失以及数据类别不均衡等问题,从不同的数据驱动方法进行分类,主要包括基于传统机器学习、基于深度学习和基于集成学习的换道意图识别方法,对近年来这些方法的研究成果进行了回顾和总结.关于换道行为的认定,存在两种主流方案,即车辆穿越车道线和未穿越车道线.对于未穿越车道线的车辆,主要应用于驾驶者换道意图的早期识别方法;而当车辆穿越过车道线时,则通常被用于完整的换道过程的识别.在换道意图标注的研究中,研究者们针对固定时间窗口和航向角阈值对标注精度的影响进行了深入探讨.为了找到最优参数,如最佳的固定时间窗口和航向角阈值,研究者们采用了网格搜索进行寻优.虽然这种方法在固定的驾驶场景中表现良好,但在不同的驾驶场景中,如何实现参数的自适应调节仍然是一个挑战.针对换道数据类别不均衡的问题,研究者采用两种策略:一是调整数据采样方法,利用欠采样和过采样技术平衡各类别样本数量;二是采用对不均衡数据适应性强的分类模型,如集成学习算法或代价敏感学习,以维持较好的分类性能.In recent years,with the rapid development of big data and artificial intelligence technology,data-driven automatic driving vehicle lane change intention recognition has become an active research area in the transportation field.Numerous studies have reported innovative and practical research results.However,this field still presents common technical challenges,such as accurately identifying the lane change process,handling missing lane change labels,and addressing imbalanced data categories.These issues remain the focal points of current research.This paper aims to classify and organize various data-driven methods,mainly focusing on lane change intention recognition methods based on traditional machine learning,deep learning,and ensemble learning.In the academic community,two primary approaches exist for identifying lane change behavior.The first approach mainly focuses on the vehicle not crossing the lane line,which is suitable for early recognition of the driver’s intention to change lanes.The second approach focuses on the actual crossing of lane markings by vehicles,which is often considered the complete lane change process.In academic research on lane change intention annotation,the selection of fixed time windows and heading angle thresholds plays a crucial role in the accuracy of annotation.These parameters affect the accurate recognition of lane change behavior and are directly related to the stability and reliability of autonomous driving and intelligent transportation system performance.Therefore,researchers have conducted in-depth investigations on the impact of these two parameters on annotation accuracy.To identify the optimal fixed time window and heading angle threshold,researchers have used the grid search optimization algorithm.This method performs well in fixed driving scenarios by traversing all possible parameter combinations and selecting the optimal parameters according to preset evaluation criteria.However,in practical applications,driving scenarios often exhibit diversity and complexity

关 键 词：交通工程 自动驾驶 车辆 数据驱动 换道意图识别 

分 类 号：U495[交通运输工程—交通运输规划与管理]