Aug 2026· Frontiers in Artificial Intelligence· Vol 9, pp.
1856630
· 0 citations· 34 references
Medicine
TL;DR
A Federated Spatio-Temporal Synchronous Dynamic Graph Neural Network (Federated STSDGNN) framework for privacy-preserving and adaptive trajectory prediction, which achieves approximately 23.5% lower RMSE compared to the centralized STSDGNN baseline in the conducted experiments, and has the potential to support future privacy-preserving V2X (Vehicle-to-Everything) applications.
Abstract
Accurate vehicle trajectory prediction plays a vital role in autonomous driving and intelligent transport systems. Deep learning models like LSTM, CNN, GNN, etc., have shown remarkable performance but often operate in a centralized setting, aggregating raw trajectory data at the server. Furthermore, the majority of models focus on either spatial or temporal features alone, but overlook the information that can be obtained by combining spatio-temporal features. This leads to major privacy concerns, issues with centralized data, and scalability problems. To overcome these challenges, we introduce a Federated Spatio-Temporal Synchronous Dynamic Graph Neural Network (Federated STSDGNN) framework for privacy-preserving and adaptive trajectory prediction. Using the highD dataset, trajectories are segmented into spatiotemporal sequences and represented as dynamic interaction graphs. Each client (vehicle or roadside unit) locally trains an STSDGNN consisting of a pre-processing module, a spatial-temporal synchronization module (GCN/GAT with GRU) and a prediction module (CNN with MLP). Clients send only model updates, which are aggregated by the server using a federated learning algorithm. This design improves privacy, achieves approximately 23.5% lower RMSE compared to the centralized STSDGNN baseline in the conducted experiments, and has the potential to support future privacy-preserving V2X (Vehicle-to-Everything) applications.
Short-term traffic forecasting plays an important role in intelligent transportation systems, as applications such as route guidance, adaptive traffic signal control, emergency response, congestion mitigation, and logistics planning depend on accurate estimates of future traffic states. Recent spatio-temporal graph neural network models, particularly Diffusion Convolutional Recurrent Neural Network (DCRNN) and Spatio-Temporal Graph Convolutional Network (STGCN), have improved traffic prediction by representing road sensors as graph nodes and jointly learning spatial and temporal relationships. However, strong performance on clean benchmark datasets does not always guarantee reliability in real-world conditions, where sensor reading may be missing, noisy, or unavailable due to hardware faults or communication issues. This paper provides a robust focused comparative evaluation of seven traffic forecasting approaches: Persistence, Historical Average, ARIMA, Random Forest, LSTM, STGCN, and DCRNN. Experiments are conducted on the METR-LA and PEMS-BAY speed datasets using 5-minute data intervals, a 12-step historical input sequence, and a 12-step forecasting horizon. Under clean-data conditions, DCRNN achieves the best overall MAE on both datasets, with 3.548 mph on METR-LA and 1.905 mph on PEMS-BAY. However, when 40% random missing input corruption is introduced, STGCN shows greater robustness than DCRNN; for METR-LA, STGCN’s MAE increases by 27.0%, while DCRNN’s MAE increases by 34.5%. Under complete sensor-failure conditions, the model ranking changes further, with the LSTM baseline showing greater stability than both graph-based models on METR-LA. Graph ablation analysis also indicates that temporal modeling accounts for most of the forecast improvement, while the evaluated graph topology provides only limited additional benefits. These findings suggest that traffic forecasting models should be assessed not only by clean-data accuracy but also by their robustness under degraded sensing conditions before deployment in real intelligent transportation systems.
Shreya N. Desai, Kasim Ishaque Ghanchi, Ali Mehdi Mirza et al.· International journal of res...· 0 citations
Experiments on controlled synthetic data and the Q-Traffic real-world dataset demonstrate that the proposed framework improves predictive accuracy, cross-client stability, and robustness under heterogeneous federated traffic scenarios.
Zhi-Cheng Wang, Tao Zhang, Yi-Meng Zhu et al.· Applied Sciences· 0 citations
Trajectory prediction is a critical component of autonomous driving systems, enabling vehicles to anticipate future motion behaviors and perform safe decision-making in dynamic traffic environments. While recent trajectory forecasting methods achieve state-of-the-art prediction accuracy, many operate as black-box systems and are evaluated primarily on high-end computing platforms, limiting their interpretability and practical deployment feasibility in resource-constrained autonomous driving systems. To address these limitations, this work proposes an interpretable and edge-deployable spatio-temporal trajectory prediction framework for autonomous driving. The proposed architecture integrates a Temporal Convolutional Network with Multi-Head Self-Attention (TCN–MHSA) in ActorNet for selective temporal modeling, a Lane Graph Attention Network (LaneGAT) for structured spatial reasoning, and a multi-stage FusionNet for actor–lane interaction. To improve model interpretability, a comprehensive Explainable AI (XAI) evaluation framework is introduced, including temporal sensitivity analysis, interaction-aware perturbation studies, spatial influence analysis, and gradient-based feature attribution methods. These analyses provide insights into how the model captures temporal motion dependencies, neighboring vehicle interactions, and environmental context during trajectory prediction. To improve the robustness of the interpretability analysis, temporal sensitivity was additionally evaluated over 100 validation scenes, demonstrating that recent observations consistently exert the greatest influence on trajectory prediction, while neighboring interaction effects gradually diminish with increasing spatial separation. Furthermore, practical real-world deployment feasibility is investigated on the NVIDIA Jetson Xavier NX platform using edge-aware optimization strategies, including mixed-precision inference and graph-complexity reduction techniques for efficient resource-constrained inference, achieving 125.74 ms latency at 12.86 W. Additional edge deployment comparisons with HiVT and SIMPL approaches under identical hardware conditions demonstrate that the proposed framework provides a more favorable balance between computational efficiency and embedded deployment performance. Experimental evaluation on the Argoverse 1 dataset demonstrates a minimum Average Displacement Error (minADE) of 0.90 m, a minimum Final Displacement Error (minFDE) of 1.50 m, a Miss Rate (MR) of 0.19, and DAC = 0.95, while establishing an accuracy–deployability operating point under embedded hardware constraints with low power consumption and practical inference throughput.
R. Megalingam, Naveen Prasaad Selvarajan, Pritty Vijay· Italian National Conference...· 0 citations
With the continuous improvement in intelligent transportation and data perception levels, determining how to achieve high-precision and generalizable traffic flow prediction based on historical traffic data has become an important issue in intelligent highway management. The Transformer model, with its strong temporal modeling capabilities, has gradually become a research hotspot in time series prediction. However, its original structure has certain limitations in modeling spatial dependencies, making it difficult to fully exploit the topological relationships of the traffic network, and it has weak adaptability to external environmental changes. To address these issues, this paper proposes an improved Transformer prediction model that integrates a graph convolutional network (GCN) and a self-attention mechanism. The model captures spatial topological information through the GCN module, introduces temporal self-attention mechanism and temporal encoding to enhance the modeling of temporal features, and combines weather factors to achieve perception modeling of external disturbances. In the experimental design, considering the uneven distribution of perception resources in reality, the model input only uses the historical traffic data of some nodes, and different node coverage rates are set to test the performance of the model under sparse input conditions. The results show that the model can maintain good accuracy and stability under multiple coverage rates, verifying the effectiveness and application prospects of the structural improvement.
Jin Zhang, Fengmin Tan, Wei Bai et al.· Italian National Conference...· 0 citations
Accurate prediction of future vehicle trajectories is essential for ensuring safety and reliable decision-making in autonomous driving systems. However, existing deep learning-based approaches exhibit several limitations. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) struggle to effectively model long-term temporal dependencies and complex agent interactions, while Transformer-based architectures often suffer from high computational complexity and limited efficiency. To overcome these challenges, this paper proposes an efficient Mamba-based feature extraction framework for jointly encoding vehicle trajectories and map information. By leveraging state-space modeling and a selective scanning mechanism, the proposed approach effectively captures longrange dependencies and enhances the representation of complex traffic behaviors. Specifically, raw scene data are first normalized and embedded into a unified feature space. A Mamba Encoder is then employed to extract high-level features from historical vehicle trajectories and map elements. Subsequently, Vehicle-Vehicle and Vehicle-Map interaction modules are introduced to explicitly model dynamic interactions among traffic participants and between vehicles and the surrounding map. The resulting high-dimensional features are further fused using an additional Mamba Encoder, while a Global Interaction Module is designed to capture scenelevel dependencies. Finally, a Gated Recurrent Unit (GRU) decoder generates multi-modal future trajectory predictions. Experimental results on the Argoverse 1 dataset demonstrate that the proposed method achieves superior performance in terms of minADE, minFDE, and minMR, while maintaining high computational efficiency.OPEN ACCESS Received: 28/01/2026 Accepted: 16/04/2026
J. Li, L. Wang, J. Pei· Revista Internacional de Mét...· 0 citations
Urban traffic congestion poses a persistent challenge to sustainable mobility, demanding intelligent forecasting systems capable of capturing both the complex spatial layout of road networks and the dynamic evolution of traffic patterns over time. Traditional statistical and deep learning methods often fail to jointly model these interdependent spatiotemporal characteristics, particularly when road topology is incomplete, outdated, or oversimplified. To address this gap, we propose a novel spatiotemporal graph neural network architecture specifically designed for short-term traffic flow prediction. Our framework introduces an adaptive graph learning module that dynamically infers meaningful connectivity relationships among traffic sensors—not relying on fixed geographic or distance-based assumptions—but instead leveraging real-time traffic correlations and node-level embeddings. This learned structure is then integrated into a hierarchical encoder that combines spectral graph convolution with dilated causal temporal convolutions, enabling effective modeling of both localized spatial interactions and multi-scale temporal dependencies across varying prediction horizons. A lightweight, shared-weight decoder supports flexible multi-step forecasting without sacrificing accuracy or efficiency. We rigorously evaluate our approach on two large-scale, real-world traffic datasets—PeMSD4 and PeMSD8—covering diverse urban environments and sensor densities. Experimental results demonstrate consistent and statistically significant improvements over state-of-the-art baselines, achieving more than a twelve percent reduction in mean absolute error for fifteen-minute forecasts. Ablation studies confirm the critical role of adaptive graph construction and hierarchical temporal modeling, while visualization of learned graphs reveals interpretable, physically plausible connectivity patterns aligned with actual road infrastructure and traffic behavior. The system shows strong deployment potential, balancing predictive performance with practical considerations such as inference speed and compatibility with edge-computing platforms used in modern traffic management centers.
Zhengxu Luan, Huan Wang, Miaobowen Wang et al.· Computers and artificial int...· 0 citations