华南理工大学学报(自然科学版) ›› 2024, Vol. 52 ›› Issue (6): 34-44.doi: 10.12141/j.issn.1000-565X.230258

• 绿色智慧交通 • 上一篇    下一篇

轨迹数据驱动的车辆换道意图识别模型

苑仁腾1,2 (),  王晨竹1 ,  项乔君1†() ,  郑欧2 ,  丁圣轩2   

  1. 1.东南大学 交通学院,江苏 南京 211189
    2.中佛罗里达大学 土木、环境和建筑工程系,美国 奥兰多 32816-2450
  • 收稿日期:2023-04-21 出版日期:2024-06-25 发布日期:2023-10-19
  • 通信作者: 项乔君(1964—),男,博士,教授,主要从事交通设计和道路交通安全研究。 E-mail:xqj@seu.edu.cn
  • 作者简介:苑仁腾(1994—),男,博士生,主要从事驾驶行为及交通安全研究。E-mail:rtengyuan123@126.com
  • 基金资助:
    国家自然科学基金资助项目(52372323);江苏省研究生科研与实践创新计划项目(KYCX22_0270);山西省交通道路设计数字化技术创新中心资助项目(202104010911019)

Trajectory Data-Driven Model for Vehicle Lane Change Intention Recognition

YUAN Renteng1,2 () , WANG Chenzhu1 ,  XIANG Qiaojun1 () ,  ZHENG Ou2 ,  DING Shengxuan2   

  1. 1.School of Transportation,Southeast University,Nanjing 211189,Jiangsu,China
    2.College of Engineering and Computer Science,University of Central Florida,Orlando 32816-2450,America
  • Received:2023-04-21 Online:2024-06-25 Published:2023-10-19
  • Contact: 项乔君(1964—),男,博士,教授,主要从事交通设计和道路交通安全研究。 E-mail:xqj@seu.edu.cn
  • About author:苑仁腾(1994—),男,博士生,主要从事驾驶行为及交通安全研究。E-mail:rtengyuan123@126.com
  • Supported by:
    the National Natural Science Foundation of China(52372323);the Postgraduate Research & Practice Innovation Program of Jiangsu Province(KYCX22_0270)

摘要:

为及时识别、预测车辆的换道行为,综合考虑目标车辆及周边车辆的时空交互关系,结合时间卷积网络(Temporal Convolutional Network,TCN)的时序处理能力和长短期记忆(Long Short Term Memory,LSTM)神经网络的门控记忆机制,构建了基于TCN-LSTM网络的车辆换道意图识别模型。首先,将目标车辆的驾驶意图分为直行、向左换道和向右换道3种类型,从CitySim车辆轨迹数据集中提取出目标车辆及对应同车道、左侧车道、右侧车道的相邻前车和相邻后车的轨迹数据,并利用中值滤波算法获得车辆运行状态指标。其次,针对统计学理论和机器学习方法面临的识别精度不高、训练时间长、参数更新慢等问题,提出利用膨胀卷积技术提取时间序列的时序特征,采用门控记忆单元捕捉时序特征的长期依赖关系,并以目标车辆及周边相邻车辆的速度、加速度、航向角、航向角变化率和相对位置信息等54个车辆状态指标为输入变量,以车辆的换道意图为输出变量,构建了一个基于TCN-LSTM网络的车辆换道意图识别模型。最后,对比分析了不同输入时间步长下TCN、支持向量机(Support Vector Machines,SVM)、LSTM和TCN-LSTM模型的识别精度。结果表明:输入时间序列长度为150帧时,TCN-LSTM模型的识别精度达到最高值96.67%;从整体分类精度来看,相比LSTM、TCN和SVM模型,TCN-LSTM模型的换道意图分类准确率分别提升了1.34、0.84和2.46个百分点,展现出了更高的分类性能。

关键词: 交通工程, 换道意图, 时间卷积网络, 长短期记忆神经网络, 车辆轨迹, CitySim数据集

Abstract:

In order to accurately recognize and estimate the lane-changing intentions of vehicles, a vehicle lane change intention recognition model based on TCN-LSTM network is proposed, which combines the temporal processing capability of TCN (Temporal Convolutional Network) with the gate memory mechanism of LSTM (Long Short Term Memory Network). In the investigation, firstly, the driving intentions of the target vehicle are divided into three types, namely going straight, changing lanes to the left, and changing lanes to the right. The running state indicators of the target vehicle and its surrounding neighboring vehicles (including the adjacent front and rear vehicles in the same lane, left lane and right lane) are extracted from the Citysim vehicle trajectory dataset using the median filtering algorithm. Secondly, to overcome the low recognition accuracy, long training time and slow parameter updating existing in statistical theories and traditional machine learning methods, the dilated convolution technique is used to extract the temporal features of time series, and the gate memory units are used to capture the long-term dependency relationships of temporal features. With 54 indicators, including the speed, acceleration, heading angle, heading angle change rate, and relative position information of the target vehicle and surrounding neighboring vehicles, as input parameters, and with the lane change intention of the vehicle as the output indicator, a vehicle lane -change intention recognition model based on the TCN-LSTM network is constructed. Finally, the recognition accuracy of TCN, SVM (Support Vector Machines), LSTM, and TCN-LSTM models under different input time steps are comparatively analyzed. The results show that, when the input time series length is 150 frames, the recognition accuracy of the TCN-LSTM model reaches a maximum of 96.67%; and that, in terms of overall classification accuracy, as compared with LSTM, TCN and SVM models, the TCN-LSTM model improves the classification correctness of lane change intention by 1.34, 0.84 and 2.46 percentage points, respectively, which demonstrates better classification performance.

Key words: transportation engineering, lane change intention, temporal convolutional network, long short term memory neural network, vehicle trajectory, CitySim dataset

中图分类号: