交通运输系统工程与信息

交通运输系统工程与信息 ›› 2025, Vol. 25 ›› Issue (2): 128-137.DOI: 10.16097/j.cnki.1009-6744.2025.02.012

• 智能交通系统与信息技术 • 上一篇    下一篇

基于概率融合的车辆意图识别与轨迹预测方法

辛嵩，刘晗，王可*，曲奕润，宋新宇   

  1. 山东科技大学，交通学院，山东青岛266000
  • 收稿日期:2024-11-26 修回日期:2024-12-24 接受日期:2025-01-18 出版日期:2025-04-25 发布日期:2025-04-20
  • 作者简介:辛嵩(1968—)，男，山东烟台人，教授，博士。
  • 基金资助:
    国家自然科学基金(52102417)。

Vehicle Intention Recognition and Trajectory Prediction Based on Probabilistic Fusion

XIN Song,LIU Han,WANG Ke*,QUYirun,SONG Xinyu   

  1. College of Transportation, Shandong University of Science and Technology, Qingdao 266000, Shandong, China
  • Received:2024-11-26 Revised:2024-12-24 Accepted:2025-01-18 Online:2025-04-25 Published:2025-04-20
  • Supported by:
    National Natural Science Foundation of China (52102417)。

PDF

摘要： 传统的轨迹预测机动模型需要识别车辆所在场景和驾驶意图等因素，选择相应的轨迹预测模型。然而，该框架在场景复杂或驾驶意图不明确时，其识别误差可能会传递至轨迹预测模型，影响预测精度。因此，本文提出一种概率融合的方法，并引入概率校准技术，通过意图识别预测概率融合多种车辆轨迹预测模型的结果，弥补机动模型框架中意图识别误差对轨迹预测精度的影响。首先，以目标车辆和周围车辆的历史轨迹为基础，提取输入输出模块；其次，采用XGBoost(eXtremeGradientBoosting)算法进行意图识别和意图概率预测，为每种识别出的意图分别构建三层门控循环单元(GatedRecurrentUnit,GRU)轨迹预测模型，并使用普拉特缩放和保序回归两种方法进行概率校准，基于概率校准和轨迹预测的结果实现概率融合；最后，使用CQSkyEyeX数据集训练和验证模型。实验结果表明：基于普拉特缩放和概率融合的组合模型在各项评估指标上均优于其他模型；当使用4s的时间窗口和2s的预测步长时，该模型的平均绝对误差(Mean Absolute Error, MAE)为0.87 m；与未使用概率融合的模型(1.32 m)相比，误差降低了34%；与仅采用加权融合的模型(1.18m)相比，误差降低了26%。

关键词: 智能交通, 概率融合, 门控循环单元, 车辆轨迹预测, 换道意图识别

Abstract: Traditional maneuver-based trajectory prediction models need to identify factors such as the scenario where the vehicle is located and the driving intention in order to select the corresponding trajectory prediction model. However, the recognition error of this framework may be propagated to the trajectory prediction model when the scenario is complex or the driving intention is unclear, thus affecting the prediction accuracy. Therefore, this study proposes a probabilistic fusion approach and introduces a probabilistic calibration technique to fuse the results of multiple vehicle trajectory prediction models through the intention recognition prediction probabilities to mitigate the impact of intention recognition errors in the maneuver-based model framework on the trajectory prediction accuracy. First, the input and output modules are extracted based on the historical trajectories of the target vehicle and the surrounding vehicles. Second, the XGBoost (Extreme Gradient Boosting) algorithm is used for intention recognition and intention probability prediction, and a three-layer Gated Recurrent Unit (GRU) is constructed separately for each recognized intention. The trajectory prediction model is constructed for each identified intention, probability calibration is performed using both Platt scaling and isotonic regression, and probabilistic fusion based on results of probabilistic calibration and trajectory prediction. Finally, the model is trained and validated using the CQSkyEyeX dataset. The experimental results show that the combined model based on Platt scaling and probabilistic fusion outperforms the other models in all evaluation metrics. When using a time window of 4 s and a prediction step of 2 s, the model has a Mean Absolute Error (MAE) of 0.87 m. Compared with the model without probabilistic fusion (1.32 m), the error is reduced by 34%. Compared to the model using only weighted fusion (1.18 m), the error is reduced by 26%.

Key words: intelligent transportation, probability fusion, gated recurrent unit, vehicle trajectory prediction, lane change intention recognition

中图分类号: