计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (2): 344-352.DOI: 10.3778/j.issn.1002-8331.2210-0058

• 工程与应用 • 上一篇    

多模型融合的换道意图识别研究

方艺洁,廖祝华,黄浩楷,李彦君   

  1. 1.湖南科技大学 计算机科学与工程学院,湖南 湘潭 411100
    2.湖南科技大学 元宇宙创新研究院,湖南 湘潭 411100
    3.浪潮云信息技术股份公司,北京 100085
  • 出版日期:2024-01-15 发布日期:2024-01-15

Lane Change Intention Recognition Based on Multi-Model Fusion

FANG Yijie, LIAO Zhuhua, HUANG Haokai, LI Yanjun   

  1. 1.College of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan, Hunan 411100, China
    2.Metaverse Innovation Research Institute, Hunan University of Science and Technology, Xiangtan, Hunan 411100, China
    3.Cloud Inspur Information Technology Co., Ltd., Beijing 100085, China
  • Online:2024-01-15 Published:2024-01-15

摘要: 快速准确识别出周围车辆的换道意图对高级自动驾驶辅助系统的决策支持和安全预防具有重要意义。针对现有方法未能充分考虑车辆之间的交互作用以及轨迹数据的前后依赖性问题,提出一种基于多模型融合的换道意图识别框架。该换道意图识别框架主要包括输入处理与换道意图识别两部分。输入处理部分对车辆轨迹数据进行清洗、贴标、切片以及one-hot编码。换道意图识别部分则具体提出BiLSTM-F(BiLSTM-fusion)模型,在该模型中将注意力机制(attention mechanism)引入双向长短期记忆网络(BiLSTM),对输入处理部分的输出信息进行权重划分,使模型能将注意力更加集中于对换道意图影响较大的特征信息上,最后引入条件随机场(conditional random field),充分学习输入数据的前后依赖性并快速找出全局最优的换道意图。实验中使用公开数据集NGSIM进行训练并评估,验证结果表明该模型的准确率最高能达到97.19%,并且可在车辆到达换道点前2?s识别车辆的换道意图,准确率为94.16%。与基线换道意图识别模型相比,所提出模型的准确率、损失、F1值和稳定性均优于基线模型。

关键词: 智能交通, 换道意图识别, 前后依赖性, 注意力机制, 条件随机场

Abstract: Fast and accurate identification of lane-changing intentions of surrounding vehicles is of great significance for decision support and safety prevention of advanced autonomous driving assistance systems. Aiming at the problems that the existing methods fail to fully consider the interaction between vehicles and the front-to-back dependence of trajectory data, this paper proposes a lane-changing intention recognition framework based on multi-model fusion. The framework mainly includes input processing and lane-changing intention recognition. The input processing part cleans, labels, slices and one-hot codes the vehicle track data. The BiLSTM-F(BiLSTM-fusion) model is specifically proposed for lane change intention recognition. In this model, the attention mechanism is introduced into BiLSTM, and the weight of the output information in the input processing part is divided. Finally, conditional random field is introduced to fully learn the dependency of input data and quickly find the global optimal lane-changing intention. In the experiment, NGSIM is used for training and evaluation. The validation results show that the model can achieve the highest accuracy of 97.19%, and can identify the vehicle’s lane changing intention 2 s before the vehicle arrives at the lane-changing point, with an accuracy of 94.16%. Compared with the baseline lane-change intention recognition model, the accuracy, loss, F1 value and stability of the proposed model are better than those of the baseline model.

Key words: intelligent transportation, lane change intention recognition, anterior and posterior dependence, attention mechanism, conditional random field