PACE: Planner-Aligned Calibrated Ellipses for Probabilistic Pedestrian Crossing Trajectory Prediction

Pedestrian crossing is one of the most safety-critical scenarios for autonomous vehicles. Although recent trajectory prediction methods can provide reasonably accurate and diverse future trajectories, motion planning tends to rely on stepwise probabilistic coverage of pedestrian positions under a specified risk level. This paper formulates pedestrian crossing trajectory prediction as a probabilistic coverage problem and proposes the planner-aligned calibrated ellipse (PACE) model. PACE represents each prediction step as a Gaussian distribution and maps it in closed form to a confidence ellipse. The model fuses the semantic map, temporal occupancy maps of nearby road users, and trajectory heatmaps within a target-centric local region, and extracts scene features through a lightweight multi-resolution backbone. It then incorporates social features into both the Gaussian mean and covariance through social attention and multi-scale FiLM modulation. Stepwise quantile calibration further aligns nominal confidence levels with empirical coverage. Two metrics, confidence ellipse coverage and confidence ellipse compactness, are introduced to jointly quantify the trade-off between safety and efficiency. Experiments show that PACE surpasses strong baselines in ADE/FDE. The model also outputs calibrated 90% and 95% confidence ellipses and maintains real-time inference with adjustable conservatism. Therefore, the framework offers a planner-compatible and calibrated safety-boundary representation for autonomous driving.