An Interpretable and Edge Deployable Spatio-Temporal Trajectory Prediction for Autonomous Driving

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.