基于时空维度分解与异构解码的轻量化轨迹预测模型

林清华; 王美芳; 李辉; 葛同澳; 颜军

doi:10.12263/DZXB.20260073

您当前的位置：

首页 >

文章列表页 >

基于时空维度分解与异构解码的轻量化轨迹预测模型

学术论文 | 更新时间：2026-07-30

- 基于时空维度分解与异构解码的轻量化轨迹预测模型
- A Lightweight Trajectory Prediction Model with a Spatio-Temporal Factorized Encoder and a Heterogeneous Decoder
- 电子学报 2026年54卷第4期页码：1748-1760
- 作者机构：
  
  1.青岛科技大学数据科学学院，山东青岛 266100
  2.武汉理工大学计算机与人工智能学院，湖北武汉 430070
- 作者简介：
  
  林清华男，1998年2月出生于山东省潍坊市。现为青岛科技大学数据科学学院硕士研究生。主要研究方向为多目标跟踪、轨迹预测。 E-mail: f023111001@mails.qust.edu.cn
  王美芳女，2001年2月出生于山东省东营市。现为青岛科技大学数据科学学院硕士研究生。主要研究方向为多目标跟踪、轨迹预测。 E-mail: wangmeifang@mails.qust.edu.cn
  李辉男，1984年3月出生于河南省平顶山市。现为青岛科技大学数据科学学院副教授、硕士生导师。主要研究方向为计算机视觉、3D目标检测及跟踪。 E-mail: lihui@qust.edu.cn
  葛同澳男，1999年6月出生于山东省菏泽市。2024年毕业于青岛科技大学数据科学学院。现为武汉理工大学博士研究生。主要研究方向为自动驾驶、目标检测。 E-mail: getongao@whut.edu.cn
  颜军男，1962年12月出生于山东省高密市。现为青岛科技大学数据科学学院院长、俄罗斯工程院外籍院士、俄罗斯自然科学院外籍院士。主要研究方向为自动驾驶、遥感图像处理。 E-mail: yanjun@qust.edu.cn
- 基金信息：
  
  国家重点研发计划(2023YFF0612100);山东省自然科学基金面上项目(ZR2024MF023);山东省重点研发计划(2024TSGC0170)
- DOI：10.12263/DZXB.20260073
  中图分类号： TP391;
- 收稿：2026-02-27，
  
  录用：2026-04-08，
  
  网络首发：2026-05-22，
  
  纸质出版：2026-04-25
- 稿件说明：
移动端阅览
林清华, 王美芳, 李辉, 等. 基于时空维度分解与异构解码的轻量化轨迹预测模型[J]. 电子学报, 2026, 54(04): 1748-1760.

LIN Qinghua, WANG Meifang, LI Hui, et al. A Lightweight Trajectory Prediction Model with a Spatio-Temporal Factorized Encoder and a Heterogeneous Decoder[J]. Acta Electronica Sinica, 2026, 54(04): 1748-1760.
林清华, 王美芳, 李辉, 等. 基于时空维度分解与异构解码的轻量化轨迹预测模型[J]. 电子学报, 2026, 54(04): 1748-1760. DOI：10.12263/DZXB.20260073

LIN Qinghua, WANG Meifang, LI Hui, et al. A Lightweight Trajectory Prediction Model with a Spatio-Temporal Factorized Encoder and a Heterogeneous Decoder[J]. Acta Electronica Sinica, 2026, 54(04): 1748-1760. DOI：10.12263/DZXB.20260073

摘要

高精度与高效率的轨迹预测模型是自动驾驶系统实现安全导航的关键。现有方法多采用图神经网络（Graph Neural Networks，GNNs）或注意力机制联合建模智能体与高精地图之间的时空依赖关系，虽然在预测精度上取得了显著进展，但其耦合式的时空建模方式导致计算复杂度过高，严重限制了模型在车载嵌入式系统等实际平台上的部署能力。此外，目前主流方法在解码阶段普遍采用单一的参数化解码范式处理场景中所有智能体，忽视了不同智能体在预测精度与多模态性方面的差异化需求，造成计算资源的低效分配。为此，本文提出一种基于时空维度分解与异构解码的轻量化轨迹预测模型（lightweight trajectory prediction Model based on spatio-temporal dimension decomposition and heterogeneous decoding，MAM-M），该模型由时空维度分解编码器与异构解码器两大核心模块构成。在编码阶段，针对传统Mamba模型在长序列建模中存在的单向信息衰减问题，设计了时序双流Mamba网络，通过正向与反向双流并行处理历史轨迹，使每个时间步同时包含完整的过去与未来信息，有效缓解了长程遗忘问题，并利用线性复杂度技术实现了时间维度上的高效编码。随后，采用自注意力机制在空间维度

上捕捉智能体与高精地图之间的全局交互关系，从而实现“先时间后空间”的解耦编码策略，显著减少了时空维度上的冗余计算。在解码阶段，提出一种异构解码器以适配不同智能体的预测需求：对运动复杂、行为不确定性高的目标智能体，采用基于可学习查询的模式查询解码器，通过交叉注意力机制主动检索场景上下文信息，精确捕捉其未来轨迹的多模态分布；对数量较多且运动相对规则的非目标智能体，引入神经-贝塞尔解码器，通过神经网络预测贝塞尔曲线的控制点并结合数学表达式生成轨迹，在保证精度的同时显著降低推理延迟（latency），实现了高效的端到端训练与推理。在Argoverse与Argoverse 2数据集上的实验结果表明：相较于主流轻量化模型，本文模型以仅2.7 × 10

的参数量和18.6 ms的latency，取得了更高的预测精度，在精度、参数量与推理速度三者之间实现了最佳平衡，展现出优异的实际部署潜力。

Abstract

Accurate and efficient trajectory prediction is pivotal for safe navigation in autonomous driving systems. Existing methods typically employ graph neural networks (GNNs) or attention mechanisms to jointly model the spatio-temporal dependencies between agents and high-definition maps. Despite notable gains in accuracy

the coupled spatio-temporal modeling incurs excessive computational complexity

severely limiting deployment on embedded vehicular platforms. Moreover

prevailing decoders apply a uniform parameterized paradigm to all agents

overlooking the distinct requirements of different agents in terms of prediction accuracy and multimodality

which leads to inefficient resource allocation. To address these challenges

we propose a lightweight trajectory prediction model based on spatio-temporal dimension decomposition and heterogeneous decoding (MAM-M). In the encoding stage

we design a temporal dual-stream Mamba network that processes historical trajectories through parallel forward and reverse streams

enabling each time step to capture both past and future context. This effectively mitigates the long-range forgetting problem while maintaining linear computational complexity along the temporal dimension. A self-attention mechanism is then employed along the spatial dimension to capture global interactions between agents and map elements

realizing a “temporal-first

spatial-second” decoupled encoding strategy that significantly reduces redundant computations. In the decoding stage

we propose a heterogeneous decoder tailored to differen

t agents: for the target agent with complex motion and high behavioral uncertainty

a mode-query decoder with learnable queries retrieves scene context via cross-attention to accurately capture multimodal trajectory distributions; for the more numerous non-target agents with relatively regular motion

a neural-Bézier decoder predicts Bézier control points through a neural network and generates trajectories via mathematical formulation

substantially reducing inference latency while preserving accuracy and enabling efficient end-to-end training. Experiments on the Argoverse and Argoverse 2 datasets demonstrate that our model

with only 2.7×10

parameters and 18.6 ms inference latency

achieves superior prediction accuracy compared to mainstream lightweight models

attaining an optimal balance among accuracy

model size

and inference speed

and exhibiting excellent potential for practical deployment.

关键词

Keywords

references

Tang Xiaolong , Kan Meina , Shan Shiguang , et al . HPNet: Dynamic trajectory forecasting with historical prediction attention [C ] // IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway : IEEE , 2024 : 15261 - 15270 . DOI: 10.1109/CVPR52733.2024.01445 http://dx.doi.org/10.1109/CVPR52733.2024.01445

孔玮 , 刘云 , 李辉 , 等 . 基于全局自适应有向图的行人轨迹预测 [J ] . 电子学报 , 2022 , 50 ( 8 ): 1905 - 1916 . DOI: 10.12263/DZXB.20211613 http://dx.doi.org/10.12263/DZXB.20211613

Kong Wei , Liu Yun , Li Hui , et al . Pedestrian trajectory prediction based on global adaptive directed graph [J ] . Acta Electronica Sinica , 2022 , 50 ( 8 ): 1905 - 1916 . (in Chinese) . DOI: 10.12263/DZXB.20211613 http://dx.doi.org/10.12263/DZXB.20211613

Nayakanti N , Al-Rfou R , Zhou A , et al . Wayformer: Motion forecasting via simple & efficient attention networks [C ] // IEEE International Conference on Robotics and Automation . Piscataway : IEEE , 2023 : 2980 - 2987 . DOI: 10.1109/icra48891.2023.10160609 http://dx.doi.org/10.1109/icra48891.2023.10160609

Jia Xiaosong , Wu Penghao , Chen Li , et al . HDGT: Heterogeneous driving graph transformer for multi-agent trajectory prediction via scene encoding [J ] . IEEE Transactions on Pattern Analysis and Machine Intelligence , 2023 , 45 ( 11 ): 13860 - 13875 . DOI: 10.1109/TPAMI.2023.3298301 http://dx.doi.org/10.1109/TPAMI.2023.3298301

Zhou Zikang , Wang Jianping , Li Y H , et al . Query-centric trajectory prediction [C ] // IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway : IEEE , 2023 : 17863 - 17873 . DOI: 10.1109/cvpr52729.2023.01713 http://dx.doi.org/10.1109/cvpr52729.2023.01713

Huang Yizhou , Cheng Yihua , Wang Kezhi . Trajectory Mamba: Efficient attention-mamba forecasting model based on selective SSM [C ] // IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway : IEEE , 2025 : 12058 - 12067 . DOI: 10.1109/CVPR52734.2025.01126 http://dx.doi.org/10.1109/CVPR52734.2025.01126

Gilles T , Sabatini S , Tsishkou D , et al . THOMAS: Trajectory heatmap output with learned multi-agent sampling [C/OL ] // Proceedings of the 10th International Conference on Learning Representations , 2022 : 1 - 18 . https://openreview.net/forum id=QDdJhACYrlX https://openreview.net/forumid=QDdJhACYrlX . DOI: 10.1109/icra46639.2022.9812253 http://dx.doi.org/10.1109/icra46639.2022.9812253

Gu A , Dao T . Mamba: Linear-time sequence modeling with selective state spaces [PP/OL ] . V1.arXiv ( 2023-12-01 )[ 2025-11-01 ] . https://arxiv.org/abs/2312.00752 https://arxiv.org/abs/2312.00752 .

Dao T , Gu A . Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality [C ] // Proceedings of the 41st International Conference on Machine Learning . Vienna : PMLR , 2024 : 10041 - 10071 . DOI: 10.52202/079017-3520 http://dx.doi.org/10.52202/079017-3520

Li Wenbing , Zhou Hang , Yu Junqing , et al . Coupled Mamba: Enhanced multimodal fusion with coupled state space model [PP/OL ] . V1.arXiv ( 2024-05-29 )[ 2025-11-01 ] . https://arxiv.org/abs/2405.18014 https://arxiv.org/abs/2405.18014 . DOI: 10.52202/079017-1910 http://dx.doi.org/10.52202/079017-1910

Waleffe R , Byeon W , Riach D , et al . An empirical study of Mamba-based language models [EB/OL ] . ( 2024-06-12 )[ 2025-11-01 ] . https://arxiv.org/abs/2406.07887 https://arxiv.org/abs/2406.07887 .

Liang Ming , Yang Bin , Hu Rui , et al . Learning lane graph representations for motion forecasting [C ] // Proceedings of the 16th European Conference on Computer Vision . Berlin : Springer , 2020 : 541 - 556 . DOI: 10.1007/978-3-030-58536-5_32 http://dx.doi.org/10.1007/978-3-030-58536-5_32

Cheng Jie , Mei Xiaodong , Liu Ming . Forecast-MAE: Self-supervised pre-training for motion forecasting with masked autoencoders [C ] // IEEE/CVF International Conference on Computer Vision . Piscataway : IEEE , 2023 : 8645 - 8655 . DOI: 10.1109/iccv51070.2023.00797 http://dx.doi.org/10.1109/iccv51070.2023.00797

Zhou Zikang , Ye Luyao , Wang Jianping , et al . HiVT: Hierarchical vector transformer for multi-agent motion prediction [C ] // IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway : IEEE , 2022 : 8813 - 8823 . DOI: 10.1109/cvpr52688.2022.00862 http://dx.doi.org/10.1109/cvpr52688.2022.00862

Zhang Zhejun , Liniger A , Sakaridis C , et al . Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding [C ] // Proceedings of the 37th International Conference on Neural Information Processing Systems . New York : Curran Associates Inc. , 2023 : 2507 . DOI: 10.52202/075280-2507 http://dx.doi.org/10.52202/075280-2507

Shi Shaoshuai , Jiang Li , Dai Dengxin , et al . Motion transformer with global intention localization and local movement refinement [C ] // Proceedings of the 36th International Conference on Neural Information Processing Systems . New York : Curran Associates Inc. , 2022 : 473 . DOI: 10.52202/068431-0473 http://dx.doi.org/10.52202/068431-0473

Choi J W , Curry R , Elkaim G . Path planning based on Bézier curve for autonomous ground vehicles [C ] // Advances in Electrical and Electronics Engineering - IAENG Special Edition of the World Congress on Engineering and Computer Science . Piscataway : IEEE , 2008 : 158 - 166 . DOI: 10.1109/wcecs.2008.27 http://dx.doi.org/10.1109/wcecs.2008.27

Wilson B , Qi W , Agarwal T , et al . Argoverse 2: Next generation datasets for self-driving perception and forecasting [C ] // Proceedings of the 35th Conference on Neural Information Processing Systems (NeurIPS 2021) Track on Datasets and Benchmarks . New York : Curran Associates Inc. , 2021 .

Chang Mingfang , Lambert J , Sangkloy P , et al . Argoverse: 3D tracking and forecasting with rich maps [C ] // IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway : IEEE , 2019 : 8740 - 8749 . DOI: 10.1109/cvpr.2019.00895 http://dx.doi.org/10.1109/cvpr.2019.00895

Ngiam J , Caine B , Vasudevan V , et al . Scene transformer: A unified architecture for predicting future trajectories of multiple agents [C/OL ] // Proceedings of the 10th International Conference on Learning Representations , 2022 . https://openreview.net/pdf id=Wm3EA5OlHsG https://openreview.net/pdfid=Wm3EA5OlHsG .

Chen Weihuang , Wang Fangfang , Sun Hongbin . S2TNet : Spatio-temporal transformer networks for trajectory prediction in autonomous driving [C ] // Proceedings of the 13th Asian Conference on Machine Learning . PMLR , 2021: 454 - 469 .

Liu Yicheng , Zhang Jinghuai , Fang Liangji , et al . Multimodal motion prediction with stacked transformers [C ] // IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway : IEEE , 2021 : 7573 - 7582 . DOI: 10.48550/arXiv.2103.11624 http://dx.doi.org/10.48550/arXiv.2103.11624

Cui Henggang , Radosavljevic V , Chou F C , et al . Multimodal trajectory predictions for autonomous driving using deep convolutional networks [C ] // IEEE International Conference on Robotics and Automation . Piscataway : IEEE , 2019 : 2090 - 2096 . DOI: 10.1109/icra.2019.8793868 http://dx.doi.org/10.1109/icra.2019.8793868

Varadarajan B , Hefny A , Srivastava A , et al . MultiPath++: Efficient information fusion and trajectory aggregation for behavior prediction [C ] // IEEE International Conference on Robotics and Automation . Piscataway : IEEE , 2022 : 7814 - 7821 . DOI: 10.1109/icra46639.2022.9812107 http://dx.doi.org/10.1109/icra46639.2022.9812107

Zhang Lu , Li Peiliang , Liu Sikang , et al . SIMPL: A simple and efficient multi-agent motion prediction baseline for autonomous driving [J ] . IEEE Robotics and Automation Letters , 2024 , 9 ( 4 ): 3767 - 3774 . DOI: 10.1109/lra.2024.3370039 http://dx.doi.org/10.1109/lra.2024.3370039

Aydemir G , Akan A K , Güney F . ADAPT: Efficient multi-agent trajectory prediction with adaptation [C ] // IEEE/CVF International Conference on Computer Vision . Piscataway : IEEE , 2023 : 8261 - 8271 . DOI: 10.48550/arXiv.2307.14187 http://dx.doi.org/10.48550/arXiv.2307.14187

Prutsch A , Bischof H , Possegger H . Efficient motion prediction: A lightweight & accurate trajectory prediction model with fast training and inference speed [C ] // IEEE/RSJ International Conference on Intelligent Robots and Systems . Piscataway : IEEE , 2024 : 9411 - 9417 . DOI: 10.1109/IROS58592.2024.10802425 http://dx.doi.org/10.1109/IROS58592.2024.10802425

Wagner R , Tas Ö S , Klemp M , et al . JointMotion: Joint self-supervision for joint motion prediction [C ] // Proceedings of the 8th Conference on Robot Learning . PMLR , 2025 : 3395 - 3406 .

Peng Ling , Liu Xiaohui , Lian Heng . Functional adaptive Huber linear regression [PP/OL ] . V1.arXiv ( 2024-09-17 )[ 2025-11-01 ] . https://arxiv.org/abs/2409.11053 https://arxiv.org/abs/2409.11053 .

桑海峰 , 陈旺兴 , 王海峰 , 等 . 基于多模式时空交互的行人轨迹预测模型 [J ] . 电子学报 , 2022 , 50 ( 11 ): 2806 - 2812 . DOI: 10.12263/DZXB.20210752 http://dx.doi.org/10.12263/DZXB.20210752

Sang Haifeng , Chen Wangxing , Wang Haifeng , et al . Pedestrian trajectory prediction model based on multi-model space-time interaction [J ] . Acta Electronica Sinica , 2022 , 50 ( 11 ): 2806 - 2812 . (in Chinese) . DOI: 10.12263/DZXB.20210752 http://dx.doi.org/10.12263/DZXB.20210752

Wang Zhepeng , Chen Feng , Lertniphonphan K , et al . Technical report for argoverse challenges on unified sensor-based detection, tracking, and forecasting [PP/OL ] . V1.arXiv ( 2023-11-27 )[ 2025-11-01 ] . https://arxiv.org/abs/2311.15615 https://arxiv.org/abs/2311.15615 .

Wang Mingkun , Zhu Xinge , Yu Changqian , et al . GANet: Goal area network for motion forecasting [C ] // IEEE International Conference on Robotics and Automation . Piscataway : IEEE , 2023 : 1609 - 1615 . DOI: 10.1109/icra48891.2023.10160468 http://dx.doi.org/10.1109/icra48891.2023.10160468

Cui A , Casas S , Wong K , et al . GoRela: Go relative for viewpoint-invariant motion forecasting [C ] // IEEE International Conference on Robotics and Automation . Piscataway : IEEE , 2023 : 7801 - 7807 . DOI: 10.1109/icra48891.2023.10160984 http://dx.doi.org/10.1109/icra48891.2023.10160984

Messaoud K , Cord M , Alahi A . Towards generalizable trajectory prediction using dual-level representation learning and adaptive prompting [C ] // IEEE/CVF Conference on Computer Vision and Pattern Recognition . Piscataway : IEEE , 2025 : 27564 - 27574 . DOI: 10.1109/CVPR52734.2025.02567 http://dx.doi.org/10.1109/CVPR52734.2025.02567

Cho K , Van Merriënboer B , Gulcehre C , et al . Learning phrase representations using RNN encoder-decoder for statistical machine translation [C ] // Proceedings of the Conference on Empirical Methods in Natural Language Processing . Doha : Association for Computational Linguistics , 2014 : 1724 - 1734 . DOI: 10.3115/v1/d14-1179 http://dx.doi.org/10.3115/v1/d14-1179

Vaswani A , Shazeer N , Parmar N , et al . Attention is all you need [C ] // Proceedings of the 31st International Conference on Neural Information Processing Systems . New York : Curran Associates Inc. , 2017 : 6000 - 6010 .

Graves A , Schmidhuber J . Framewise phoneme classification with bidirectional LSTM networks [C ] // IEEE International Joint Conference on Neural Networks . Piscataway : IEEE , 2005 , 4 : 2047 - 2052 . DOI: 10.1109/IJCNN.2005.1556215 http://dx.doi.org/10.1109/IJCNN.2005.1556215

王中天 , 吴一全 . 基于视觉与深度学习的无人机自主着陆场景感知方法研究进展 [J ] . 电子学报 , 2025 , 53 ( 11 ): 4171 - 4198 .

Wang Zhongtian , Wu Yiquan . Research progress of UAV autonomous landing scene perception methods based on vision and deep learning [J ] . Acta Electronica Sinica , 2025 , 53 ( 11 ): 4171 - 4198 . (in Chinese)

浏览量

下载量

CSCD

文章被引用时，请邮件提醒。

提交

工具集

关联资源

高精度高速度低算力的AUV水下导引光学定位方法

基于图卷积与自适应Transformer的行人轨迹预测

基于多尺度特征自适应调制的单图像超分辨率网络

基于坐标重要性池化和解耦类别对齐蒸馏的图像分类算法

基于特斯拉线圈的脉冲式小型化低频天线设计