| 徐东伟*,刘靥宛则**,马潇畅**,张彪***,陈滨****.基于联合预测框架的鲁棒车辆轨迹预测[J].高技术通讯(中文),2025,(7):799~812 |
| 基于联合预测框架的鲁棒车辆轨迹预测 |
| Robust vehicle trajectory prediction based on united prediction framework |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 07. 012 |
| 中文关键词: 智能交通;车辆轨迹预测;鲁棒性;知识蒸馏;剪枝 |
| 英文关键词: intelligent transportation; vehicle trajectory prediction; robustness; knowledge distillation; pruning |
| 基金项目: |
| 作者 | 单位 | | 徐东伟* | (* 浙江工业大学网络安全研究院 杭州 311121)
(** 浙江工业大学信息工程学院 杭州 311121)
(*** 中国物品编码中心 北京 100011)
(**** 嘉兴市智慧交通重点实验室 嘉兴 314000) | | 刘靥宛则** | | | 马潇畅** | | | 张彪*** | | | 陈滨**** | |
|
| 摘要点击次数: 77 |
| 全文下载次数: 176 |
| 中文摘要: |
| 针对现有车辆轨迹预测方法在面对恶劣噪声轨迹时呈现出的脆弱性问题,本文提出了联合时空注意力网络(united spatio-temporal attention net,USTAN)实现鲁棒的车辆轨迹预测。 首先,一种联合预测框架被提出:USTAN 包括基于时空注意力的车辆轨迹预测模型(spatio-temporal attention net,STAN)与其学生模型,联合预测结果由 2 个模型的输出决定;其次,一种鲁棒压缩策略被用于提取 STAN 的学生模型;随后,基于极值理论的检测算法被应用于恶劣噪声轨迹的识别;最后,研究采用真实道路数据集 NGSIM US-101 和I-80 验证方法的可行性并评估方法的有效性。 实验结果表明,该方法在使用恶劣噪声轨迹进行预测时,鲁棒性较好,平均位移误差仅上升 32. 04% ,且能有效识别所使用的恶劣
噪声轨迹。 |
| 英文摘要: |
| To address the vulnerability of existing trajectory prediction methods when handling noisy trajectoriy data,the paper proposed the united spatio-temporal attention network (USTAN) to realize robust vehicle trajectory prediction. First,a united prediction framework is proposed: the USTAN consists of a spatio-temporal attention-based vehicle trajectory prediction model (STAN) and its student model,the united prediction result is determined by the outputs of both models. Second,a robust compression strategy is used to extract a student model from the STAN.Subsequently,a detection algorithm based on extreme value theory is applied to the identification of harsh noise trajectories. Finally,a real roadway dataset,NGSIM US-101 and I-80,is used in the study to validate and evaluate the feasibility of the method. The results of the multiple comparative analysis experiments show that the method has good robustness when using harsh noise trajectories for prediction,with the average displacement error rising by only 32. 04% ,and effectively identifies the presence of the harsh noise trajectories used. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
