Evidential pretraining improves prediction reliability for vulnerable road users
Observability-Aligned Evidential Masked Pretraining for Reliable VRU Motion Prediction
Machine Learning
Summary
Predicting how vulnerable road users like pedestrians will move is tricky because sensors often miss information and conditions can change. The authors introduce a new training method that helps models understand when they have less information and be more confident when they know more. Their method improves prediction accuracy and uncertainty estimation, especially when tested in new environments. This means the model can be more reliable in real-world conditions where visibility changes.
What this means in practice
- •For autonomous vehicle engineers: Improve pedestrian and cyclist trajectory prediction models to handle missing sensor data and environmental changes with better uncertainty estimates.
- •For smart city planners: Deploy predictive systems that reliably anticipate vulnerable road user behavior even under changing weather and visibility conditions to enhance traffic safety.
Authors
Shang Gao, Franz Götz-Hahn, Konrad Doll, Bernhard Sick
Abstract
Vulnerable road user (VRU) prediction must operate under partial observability and distribution shift, so reliable uncertainty matters alongside accuracy. Masked pretraining is attractive in this setting because it naturally exposes the model to missing content during self-supervision, but existing time-series masked autoencoding methods are typically deterministic and do not constrain how confidence should change with observability. We introduce eCoMAE, an evidential masked pretraining method for multivariate time series that explicitly aligns predictive evidence with what is observed. eCoMAE combines block-wise time-step masking with a Normal-Inverse-Gamma decoder and three observability-aware mechanisms: uncertainty-weighted mean consistency on positions visible in both a masked view and a reference pass, evidence suppression on masked positions, and representation consistency across stochastic masks. On IMPTC benchmark for VRU intention prediction under leave-one-environment-out weather shift, eCoMAE improves the strongest baseline from 0.7971 to 0.8068 in OOD accuracy and from 0.0718 to 0.0648 in OOD AURC. Additional evaluation on pedestrian trajectory forecasting across IMPTC, Waymo, nuScenes, and inD further yields the best fine-tuning MSE on all four datasets. These results show that observability-aligned evidential pretraining improves reliability under shift and remains effective beyond the primary benchmark.