Understanding Failure Modes in Pedestrian Trajectory Prediction via Interpretable Risk Analysis

This paper investigates failure modes in pedestrian trajectory prediction models, with a focus on understanding when and why predictions become unreliable. I propose a model-agnostic and interpretable risk analysis framework that identifies input conditions associated with high prediction error. Applied to a Social Force Model evaluated on the ETH pedestrian dataset across two real-world scenes, the framework combines input-space sensitivity analysis with interpretable decision tree classifiers to produce human-readable failure rules. Results show that initial orientation and position are the dominant factors determining prediction reliability, and that failure patterns differ across scenes — suggesting that scene-specific safety assessment is necessary. This work is an independent research project.