欢迎访问《汽车安全与节能学报》，
汽车安全与节能学报

汽车安全与节能学报 ›› 2024, Vol. 15 ›› Issue (5): 763-773.DOI: 10.3969/j.issn.1674-8484.2024.05.014

• 智能驾驶与智慧交通 • 上一篇    下一篇

基于LSTM-多头混合注意力的可解释换道意图预测

高凯1,2(), 刘健1, 刘林鸿1, 刘欣宇1, 张金来1,*(), 杜荣华1,2   

  1. 1.长沙理工大学 汽车与机械工程学院，长沙 410000，中国
    2.长沙理工大学 智能道路与车路协同湖南省重点实验室，长沙 410000，中国
  • 收稿日期:2024-03-07 修回日期:2024-05-01 出版日期:2024-10-31 发布日期:2024-11-07
  • 通讯作者: 张金来，讲师。E-mail：jinlai.zhang@csust.edu.cn。
  • 作者简介:高凯（1985—），男（汉），河北，副教授。E-mail：kai_g@csust.edu.cn。
  • 基金资助:
    湖南省自然科学基金项目(2024JJ5023);国家自然科学基金青年项目(62403076)

Explainable lane change intention prediction based on LSTM-multi-head mixed attention

GAO Kai1,2(), LIU Jian1, LIU Linhong1, LIU Xinyu1, ZHANG Jinlai1,*(), DU Ronghua1,2   

  1. 1. Changsha University of Science and Technology, School of Automotive and Mechanical Engineering, Changsha 410000, China
    2. Changsha University of Science and Technology, Hunan Key Laboratory of Intelligent Road and Vehicle Road Collaboration, Changsha 410000, China
  • Received:2024-03-07 Revised:2024-05-01 Online:2024-10-31 Published:2024-11-07

RichHTML

7

PDF

323

摘要：

为了使自动驾驶汽车准确地预测其周围车辆的换道意图，提出了一种基于长短期记忆神经网络（LSTM）-多头混合注意力的可解释换道意图预测模型。该模型可以充分提取目标车辆与其周围车辆之间的时空交互关系，并且提出了一种基于最大熵的Shapley加性解释方法（SHAP）来解释各个特征在特定时间步对模型输出的影响程度，在HighD数据集上进行了实验。并通过SHAP值的可视化，直观解释了换道预测模型在特定时刻的目标车辆的换道行为。 结果表明：该换道预测模型在换道前3 s的综合准确率，分别比LSTM、卷积神经网络（CNN）、多头注意力高出4.03%、9.51%、5.16%，这证明了模型在长时域预测的有效性；错误预测样本归因于模型缺陷或样本稀疏。该换道预测模型可为用户进行模型优化提供指导。

关键词: 自动驾驶汽车, 换道意图预测, 注意力机制, 长短期记忆神经网络（LSTM）, Shapley加性解释方法（SHAP）

Abstract:

An interpretable lane change intention prediction model was proposed to enable the autonomous vehicle to accurately predict the lane change intention of the vehicles around them. This model based on the Long Short-Term Memory (LSTM) and the multi-head mixed attention, which can fully extract the spatiotemporal interaction between the target vehicle and its surrounding vehicles. A Shapley additive interpretation method (SHAP) based on maximum entropy was proposed to explain the degree of influence of each feature on the model output at a specific time step, and experiments on the HighD dataset were carried out. The results show that the comprehensive accuracy of the proposed model is 4.03%, 9.51%, and 5.16% higher than that of the LSTM, the Convolutional Neural Network (CNN), and the multi-head attention, respectively, before lane change, which fully proves the validity of the model in the long time horizon. And the wrong prediction samples can be attributed to model defects or sparse samples on the other hand, guiding users to optimize the model.

Key words: autonomous driving vehicles, lane change intention prediction, attention mechanism, long short-term memory (LSTM), Shapley additive explanations (SHAP)

中图分类号: