%0 Journal Article
%T 基于CNN-Transformer-GRU-Att的车辆换道意图识别<br>Vehicle Lane-Changing Intent Recognition Based on CNN-Transformer-GRU-Att
%A 韩天立
%A 徐晴晴
%A 胡林治
%J Operations Research and Fuzziology
%P 806-822
%@ 2163-1530
%D 2024
%I Hans Publishing
%R 10.12677/ORF.2024.141075
%X 在当前自动驾驶车辆与传统车辆并存的复杂交通环境中，快速准确地识别车辆换道意图能够帮助自动驾驶系统做出更加安全舒适的操作决策。本文首先采用扩展卡尔曼滤波方法对车辆轨迹数据进行筛选并平滑处理，再基于航向角的变化对车辆驾驶行为分类并标注驾驶意图。然后，为了充分考虑车辆之间交互作用、高效提取换道过程的时序连续特征以及捕捉车辆行驶轨迹序列中局部和长期依赖性，本文构建了一种基于CNN-Transformer-GRU-Att的车辆换道意图识别模型，将目标车辆和周围车辆的行驶数据信息作为输入，实验结果表明，本文所提模型对车辆换道意图的准确率为91.38%，推理耗时为10.08 s，多种评价指标显著优于其他模型。此外消融实验证明引入的Transformer模块、GRU层和注意力机制能够分别提高3.19%，5.07%和1.08%的准确率。最后分析模型输入车辆历史行驶轨迹序列的不同长度下的意图识别结果，模型可在车辆换道前2 s内能以89%以上的准确率识别换道意图。<br />
In the current complex traffic environment where self-driving vehicles and traditional vehicles coexist, fast and accurate identification of vehicle lane changing intention can help the self-driving system make safer and more comfortable operation decisions. Firstly, the vehicle trajectory data are first filtered and smoothed using the extended Kalman filter method. Secondly, vehicle driving behavior was classified based on changes in heading angle and labeled with driving intentions. Thirdly, a vehicle lane change intention recognition model based on CNN-Transformer-GRU-Att was constructed to fully consider inter-vehicle interactions, efficiently extract time-continuous features of the lane-changing process, and capture local and long-term dependencies in the sequence of vehicle trajectories. Taking as input the information on the traveling data of the target vehicle and the surrounding vehicles, it was proved that the model proposed in this paper has an accuracy of 91.38% for the intention of vehicles to change lanes, with an inference time of 10.08 s. A variety of evaluation indexes are significantly better than those of other models. Fourthly, the ablation experiments demonstrated that the introduced Transformer module, GRU layer and attention mechanism can help the model to improve the accuracy by 3.19%, 5.07% and 1.08% respectively. Finally, the results of intent recognition under different lengths of the historical vehicle trajectory sequences inputted into the model demonstrated the model can recognize the intent to change lanes within 2 seconds before the vehicle changes lanes with an accuracy of more than 89%.
%K 换道意图识别，深度学习，Transformer模型，门控神经单元网络，注意力机制，智能交通<br>Lane Changing Intent Recognition
%K Deep Learning
%K Transformer Model
%K Gated Neural Unit Network
%K Attention Mechanism
%K Intelligent Transportation
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=82081