摘要
精准的车辆轨迹预测模型可以为自动驾驶车辆提供其周围车辆的准确运动状态信息,进而判断本车与周围车辆短期内是否有发生冲突的可能性。本文提出一种基于时域卷积网络与注意力机制(Temporal Convolutional Networks with Attention mechanism,TCN-Attention)的车辆换道轨迹预测模型。该模型以时域卷积网络作为当前输入的特征提取器,利用时间与空间注意力机制使模型在不同时间和空间位置之间建立动态关联,更准确地捕捉车辆之间的动态时空相关性,实现准确预测车辆换道轨迹。与传统单一车辆轨迹特征输入不同,本文通过对输入特征进行多维扩充与融合,进一步提高了轨迹预测准确率。此外,本文提出一种换道执行起止时刻定义方法更准确地确定数据集中的换道起止时刻。实验表明,本文所提模型能以高准确率预测变换车道轨迹,在整体效果上优于其他深度学习模型,与ConvLSTM (Convolution Long Short-Term Memory)相比,TCN-Attention的平均绝对误差(Mean Absolute Error,E_(MAE))降低了69.8%,均方根误差(Root Mean Square Error,E_(RMSE))降低了49.15%,平均绝对百分比误差(Mean Absolute Percentage Error,E_(MAPE))降低了14.24%。
An accurate vehicle trajectory prediction model can provide self-driving vehicles with precise information about the motion states of surrounding vehicles in mixed traffic flow environments,allowing it to assess the possibility of conflicts with neighboring vehicles in the short term.This paper proposes a vehicle lane-changing trajectory prediction model based on Temporal Convolutional Networks with Attention Mechanism(TCN-Attention)to improve the accuracy of vehicle lane-changing trajectory prediction.This model uses Temporal Convolutional Networks as the current input's feature extractor and utilizes a temporal and spatial attention mechanism to establish dynamic correlations between different time steps and spatial positions.Specifically,the combination of temporal and spatial attention mechanisms helps the model extract essential semantic features in both the temporal and spatial dimensions before and after lane-changing,enabling it to more accurately capture the dynamic spatiotemporal relationships between vehicles.This enables precise predictions of lane-changing trajectories on highways.Different from the traditional only using a trajectory features as input,our method achieves the multi-dimensional expansion and fusion of the input features,and further improves the accuracy of the trajectory prediction.In addition,this paper proposes a new method to define the start and end time of lane-changing in the dataset more accurately.Experiments show that the proposed model can predict the trajectory of the lane-changing with high accuracy,and the overall effect is better than other deep learning models.Compared with the Convolution Long Short-Term Memory(ConvLSTM),the Mean Absolute Error(E_(MAE))of TCN-Attention is reduced by 69.8%,the Root Mean Square Error(E_(RMSE))is reduced by 49.15%and the Mean Absolute Percentage Error(E_(MAPE))is reduced by 14.24%.
作者
杨达
刘家威
郑斌
孙峰
YANG Da;LIU Jiawei;ZHENG Bin;SUN Feng(School of Transportation and Logistics,Southwest Jiaotong University,Chengdu 610031,China;Department of Civil and Environmental Engineering,University ofWisconsin,Madison 53706,Wisconsin,USA)
出处
《交通运输系统工程与信息》
EI
CSCD
北大核心
2024年第2期114-126,共13页
Journal of Transportation Systems Engineering and Information Technology
基金
国家自然科学基金(52172333)
四川省自然科学基金(24NSFSC1109)。
