汽车工程 ›› 2024, Vol. 46 ›› Issue (11): 1993-2004.doi: 10.19562/j.chinasae.qcgc.2024.11.006

• • 上一篇    下一篇

交叉口车辆行为感知在线半监督混合方法

张海伦,王广玮,孟庆文,许庆,王建强(),李克强   

  1. 清华大学车辆与运载学院,智能绿色车辆与交通全国重点实验室,北京 100084
  • 收稿日期:2024-04-13 修回日期:2024-05-26 出版日期:2024-11-25 发布日期:2024-11-22
  • 通讯作者: 王建强 E-mail:wjqlws@tsinghua.edu.cn
  • 基金资助:
    北京市自然科学基金(3244031);国家资助博士后研究人员计划(GZB20230355);国家自然科学基金(52131201)

An Online Semi-supervised Hybrid Approach for Vehicle Behavior Perception at Intersections

Hailun Zhang,Guangwei Wang,Qingwen Meng,Qing Xu,Jianqiang Wang(),Keqiang Li   

  1. School of Vehicle and Mobility,Tsinghua University,State Key Laboratory of Intelligent Green Vehicle and Mobility,Beijing 100084
  • Received:2024-04-13 Revised:2024-05-26 Online:2024-11-25 Published:2024-11-22
  • Contact: Jianqiang Wang E-mail:wjqlws@tsinghua.edu.cn

摘要:

自动驾驶感知系统须对目标车辆运动进行感知,以制定合理交互决策。针对行为感知在时间上的滞后性和数据中可能存在的波动和异常值导致感知准确率差的问题,本文提出一种在线半监督混合方法。首先,采用自回归积分移动平均和在线梯度下降优化器设计基于数据驱动的车辆运动状态在线预测算法。然后,构建基于微簇的初始模型,并以K近邻为基分类器建立集成学习策略,设计错误驱动代表性学习和指数衰减策略实现对初始模型的迭代更新。最后,基于驾驶模拟平台采集了验证所提算法有效性的实验数据。结果表明,所提出的方法对于车辆行为波动具有快速适应性,在线预测算法可准确预测车辆运动趋势,行为感知算法对于不同预测时间下的车辆行为均有较强适应能力。

关键词: 自动驾驶, 行为预测, 自回归积分移动平均, 集成学习, 半监督学习

Abstract:

The autonomous driving perception system must perceive the movement of the target vehicle to make reasonable interactive decisions. For the time lag in behavior perception, as well as the problem that possible fluctuations and outliers in the data lead to poor perception accuracy, an online semi-supervised hybrid approach is proposed in this paper. Firstly, a data-driven online prediction algorithm for vehicle motion state is designed using autoregressive integral moving average and online gradient descent optimizer. Then, an initial model based on micro-clusters is constructed, and an ensemble learning strategy is established using K nearest neighbor as the base classifier. Error-driven representative learning and exponential decay strategies are designed to achieve iterative updates of the initial model. Finally, experimental data to verify the effectiveness of the proposed algorithm is collected based on the driving simulation platform. The results show that the proposed method has rapid adaptability to vehicle behavior fluctuations. The online prediction algorithm can accurately predict vehicle motion trends, and the behavior perception algorithm has strong adaptability to vehicle behavior at different prediction times.

Key words: autonomous driving, behavior prediction, autoregressive integral moving average, ensemble learning, semi-supervised learning