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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Trajectory prediction is crucial for autonomous driving safety and decision-making.
• Existing methods often lack interpretability and are not optimized for edge deployment.
• Resource-constrained systems require efficient and understandable prediction models.

Purpose of the Study:

• To develop an interpretable and edge-deployable spatio-temporal trajectory prediction framework for autonomous driving.
• To enhance model transparency through a comprehensive Explainable AI (XAI) evaluation.
• To demonstrate practical feasibility on embedded hardware with optimized performance.

Main Methods:

• Integrated Temporal Convolutional Network with Multi-Head Self-Attention (TCN-MHSA) and Lane Graph Attention Network (LaneGAT).
• Employed a multi-stage FusionNet for actor-lane interaction modeling.
• Utilized XAI techniques including temporal sensitivity and spatial influence analysis.
• Implemented edge-aware optimization strategies for deployment on NVIDIA Jetson Xavier NX.

Main Results:

• Achieved state-of-the-art prediction accuracy (minADE: 0.90m, minFDE: 1.50m, MR: 0.19).
• Demonstrated practical deployment feasibility with 125.74 ms latency at 12.86 W on edge hardware.
• XAI analysis revealed recent observations have the greatest influence on predictions.
• Outperformed existing approaches in computational efficiency and embedded performance.

Conclusions:

• The proposed framework offers a robust solution for interpretable and edge-deployable trajectory prediction in autonomous driving.
• It successfully balances prediction accuracy with the demands of resource-constrained embedded systems.
• The XAI framework provides valuable insights into model behavior, enhancing trust and reliability.