TPK: Trustworthy Trajectory Prediction Integrating Prior Knowledge For Interpretability and Kinematic Feasibility

Researchers developed TPK, a trajectory prediction system for autonomous vehicles that integrates prior knowledge to make predictions more trustworthy and physically feasible. The system incorporates interaction and kinematic models for vehicles, pedestrians, and cyclists, improving interpretability while ensuring predictions adhere to physics.

• →TPK system addresses trustworthiness issues in current deep learning trajectory prediction models by incorporating prior knowledge.
• →The approach uses class-specific interaction layers to capture behavioral differences between vehicles, pedestrians, and cyclists.
• →DG-SFM interaction importance score improves interpretability by revealing correlations between incorrect predictions and interaction prior divergence.
• →Kinematic models eliminate physically infeasible trajectories despite causing slight accuracy decreases.
• →Testing on Argoverse 2 dataset shows improved interaction interpretability and physics-adherent predictions.