基于动态注意目标导向的车辆轨迹预测方法

doi:10.19562/j.chinasae.qcgc.2025.10.002

摘要/Abstract

摘要：

在复杂的交通场景中，可靠且有效地预测周边车辆的轨迹对于自动驾驶汽车的安全运行至关重要。然而，现有的预测方法往往面临高计算开销的挑战，难以在不牺牲精度的情况下实现实时高效的轨迹预测。为此，本文提出了一种结合动态注意力和目标引导的创新方法DAGG，该方法可以精确捕捉动态场景的变化并捕获终点目标。针对连续预测中存在的冗余编码问题和推理延迟，本文构建了一个本地时空参照框架，用于解耦场景实例之间的内在特征和相对信息。进一步设计了高效且紧凑的三重因子化注意力融合模块，用于聚合本地上下文特征，从而捕获场景中丰富的时空背景信息。为了实现多模态预测并更好地利用场景编码，本文将场景融合特征注入地图信息，并采用多模态运动预测解码模块来指导目标选择，以捕获高质量的预测目标，同时降低基于目标轨迹生成的计算成本。在公开数据集Argoverse上的验证结果表示，本文所提的方法最小平均位移误差为0.84 m，最小最终位移误差为1.26 m，显著优于主流基线模型，展现了其在复杂动态场景中的优异预测能力。

关键词: 智能汽车, 轨迹预测, 深度学习, 注意力机制

Abstract:

In complex traffic scenarios， reliably and effectively predicting the trajectories of surrounding vehicles is crucial for the safe operation of autonomous vehicles. However， existing prediction methods often face challenges related to high computational overhead， making it difficult to achieve real-time and efficient trajectory prediction without sacrificing accuracy. Therefore， an innovative method called Dynamic Attention and Goal Guidance （DAGG） combining dynamic attention and goal guidance is proposed， which accurately captures the dynamics of changing scenes and identifies endpoint goals. To reduce redundant encoding and reasoning delay in continuous prediction， a local spatiotemporal reference framework is constructed that decouples intrinsic features from relative information between scene instances. Furthermore， an efficient and compact triple-factor attention fusion module is designed to aggregate local context features， capturing rich spatiotemporal background information. To achieve multimodal prediction and better utilize scene encoding， scene fusion features are injected into map information and a multimodal motion prediction decoding module is adopted to guide goal selection， capturing high-quality predicted goals while reducing the computational cost of goal-based trajectory generation. The validation results on the publicly available Argoverse dataset demonstrate that the proposed method achieves a minimum average displacement error （minADE） of 0.84 m and a minimum final displacement error （minFDE） of 1.26 m， significantly outperforming mainstream baseline models， which highlights its superior predictive capability in complex and dynamic scenarios.

Key words: autonomous driving, trajectory prediction, deep learning, attention mechanism

韩悦,蔡英凤,陈龙,孙晓强,王海,刘泽,饶中钰. 基于动态注意目标导向的车辆轨迹预测方法[J]. 汽车工程, 2025, 47(10): 1861-1871.

Yue Han,Yingfeng Cai,Long Chen,Xiaoqiang Sun,Hai Wang,Ze Liu,Zhongyu Rao. Vehicle Trajectory Prediction Method Based on Dynamic Attention and Goal-Guided Mechanism[J]. Automotive Engineering, 2025, 47(10): 1861-1871.

图/表 8

图1

图2

表1

表2

表3

场景注意力融合块的消融实验（验证集）"

因子化注意力特征融合块				指标
车辆间交互注意	目标-地图注意	历史时间注意	$N a t t n$ 次数	minADE	minFDE	MR
√	√			1.13	1.47	0.23
√		√		0.91	1.26	0.15
√	√	√	1	0.80	1.23	0.12
√	√	√	2	0.78	1.19	0.11
√（无门控）	√	√	2	0.81	1.23	0.13

表3

表4

表5

图3

参考文献 28

[1]	CHAI Y， SAPP B， BANSAL M， et al. Multipath： multiple probabilistic anchor trajectory hypotheses for behavior prediction［J］. arXiv preprint arXiv：， 2019.
[2]	HONG J， SAPP B， PHILBIN J. Rules of the road： predicting driving behavior with a convolutional model of semantic interactions［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2019： 8454-8462.
[3]	TAN M， LE Q. Efficientnet： rethinking model scaling for convolutional neural networks［C］. International Conference on Machine Learning. PMLR， 2019： 6105-6114.
[4]	GAO J， SUN C， ZHAO H， et al. VectorNet： encoding HD maps and agent dynamics from vectorized representation［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2020： 11525-11533.
[5]	LIANG M， YANG B， HU R， et al. Learning lane graph representations for motion forecasting［C］. Computer Vision-ECCV 2020： 16th European Conference， Glasgow， UK， August 23-28， 2020， Proceedings， Part II 16. Springer， 2020： 541-556.
[6]	YE M， CAO T， CHEN Q. TPCN： temporal point cloud networks for motion forecasting［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2021： 11318-11327.
[7]	DEO N， TRIVEDI M M. Convolutional social pooling for vehicle trajectory prediction［M/OL］. arXiv， 2018［2023-03-16］.
[8]	VARADARAJAN B， HEFNY A， SRIVASTAVA A， et al. Multipath++： efficient information fusion and trajectory aggregation for behavior prediction［C］. 2022 International Conference on Robotics and Automation （ICRA）. IEEE， 2022： 7814-7821.
[9]	MOHAMED A， QIAN K， ELHOSEINY M， et al. Social-STGCNN： a social spatio-temporal graph convolutional neural network for human trajectory prediction［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2020： 14424-14432.
[10]	LI L L， YANG B， LIANG M， et al. End-to-end contextual perception and prediction with interaction transformer［C］. 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems （IROS）. IEEE， 2020： 5784-5791.
[11]	MERCAT J， GILLES T， EL ZOGHBY N， et al. Multi-head attention for multi-modal joint vehicle motion forecasting［C］. 2020 IEEE International Conference on Robotics and Automation （ICRA）. IEEE， 2020： 9638-9644.
[12]	YU C， MA X， REN J， et al. Spatio-temporal graph transformer networks for pedestrian trajectory prediction［C］. Computer Vision-ECCV 2020： 16th European Conference， Glasgow， UK， August 23-28， 2020， Proceedings， Part XII 16. Springer， 2020： 507-523.
[13]	NGIAM J， CAINE B， VASUDEVAN V， et al. Scene transformer： a unified architecture for predicting multiple agent trajectories［J］. arXiv preprint arXiv：， 2021.
[14]	HUANG Z， MO X， LV C. Multi-modal motion prediction with transformer-based neural network for autonomous driving［C］. 2022 International Conference on Robotics and Automation （ICRA）. IEEE， 2022： 2605-2611.
[15]	VASWANI A， SHAZEER N， PARMAR N， et al. Attention is all you need［J］. Advances in Neural Information Processing Systems， 2017， 30.
[16]	NAYAKANTI N， AL-RFOU R， ZHOU A， et al. Wayformer： motion forecasting via simple & efficient attention networks［C］. 2023 IEEE International Conference on Robotics and Automation （ICRA）. IEEE， 2023： 2980-2987.
[17]	ZHOU Z， YE L， WANG J， et al. HiVT： hierarchical vector transformer for multi-agent motion prediction［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2022： 8823-8833.
[18]	LIU Y， ZHANG J， FANG L， et al. Multimodal motion prediction with stacked transformers［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2021： 7577-7586.
[19]	WANG X， SU T， DA F， et al. ProphNet： efficient agent-centric motion forecasting with anchor-informed proposals［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2023： 21995-22003.
[20]	GUPTA A， JOHNSON J， FEI-FEI L， et al. Social GAN： socially acceptable trajectories with generative adversarial networks［C］. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition， 2018： 2255-2264.
[21]	SALZMANN T， IVANOVIC B， CHAKRAVARTY P， et al. Trajectron++： dynamically-feasible trajectory forecasting with heterogeneous data［C］. Computer Vision-ECCV 2020： 16th European Conference， Glasgow， UK， August 23-28， 2020， Proceedings， Part XVIII 16. Springer， 2020： 683-700.
[22]	ZHAO H， GAO J， LAN T， et al. TNT： target-driven trajectory prediction［C］. Conference on Robot Learning. PMLR， 2021： 895-904.
[23]	GU J， SUN C， ZHAO H. DenseTNT： end-to-end trajectory prediction from dense goal sets［C］. Proceedings of the IEEE/CVF International Conference on Computer Vision， 2021： 15303-15312.
[24]	PHAN-MINH T， GRIGORE E C， BOULTON F A， et al. CoverNet： multimodal behavior prediction using trajectory sets［C］.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2020： 14074-14083.
[25]	ZENG W， LIANG M， LIAO R， et al. LanerCNN： distributed representations for graph-centric motion forecasting［C］. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems （IROS）. IEEE， 2021： 532-539.
[26]	GILLES T， SABATINI S， TSISHKOU D， et al. HOME： heatmap output for future motion estimation［C］. 2021 IEEE International Intelligent Transportation Systems Conference （ITSC）. IEEE， 2021： 500-507.
[27]	CHANG M F， LAMBERT J， SANGKLOY P， et al. Argoverse： 3D tracking and forecasting with rich maps［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2019： 8748-8757.
[28]	GILLES T， SABATINI S， TSISHKOU D， et al. GOHOME： graph-oriented heatmap output for future motion estimation［C］. 2022 International Conference on Robotics and Automation （ICRA）. IEEE， 2022： 9107-9114.

Argoverse运动预测排行
方法	K=1			K=6
方法	minADE	minFDE	MR	minADE	minFDE	MR
Argoverse Baseline	2.96	6.81	0.81	2.34	5.44	0.69
Argoverse Baseline （NN）	3.45	7.88	0.87	1.71	3.29	0.54
LaneGCN	1.71	3.78	0.59	0.87	1.36	0.16
mmTransform	1.77	4.00	0.62	0.84	1.34	0.15
DenseTNT	1.68	3.63	0.58	0.88	1.28	0.13
HOME+GOHOME	1.70	3.68	0.57	0.89	1.29	0.08
DAGG（our）	1.68	3.57	0.55	0.84	1.26	0.13

方法	参数	推理速度
方法	参数	批量大小	时间/ms
LaneGCN	3 701K	32	173
DenseTNT	1 103K	32	531
mmTransformer	2 607K
HOME+GOME	5 100K
DAGG（our）	1 208K	32	126

方法	minADE₆	minFDE₆	MR
DAGG+NMS	0.81	1.22	0.10
DAGG+our	0.76	1.16	0.11

采样密度	minADE₆	minFDE₆
3 m	0.90	1.43
2 m	0.83	1.30
1.5 m	0.80	1.22
1 m	0.78	1.19
0.5 m	0.78	1.19