Pedestrian Crossing Prediction for Automated Vehicles: A Geometry-aware Method Using Surface Normal

Abstract

Accurate prediction of pedestrian crossing intention is critical for the safety of automated vehicles. Existing approaches mainly rely on visual appearance and scene context derived from raw images and semantic segmentation. We propose a geometry-aware framework that incorporates surface normal maps encoding pixel-level 3D surface orientation. To integrate heterogeneous information, we extend a multi-stream GRU–attention architecture with surface-normal features. The visual branch combines surface normals, local context, and segmentation overlays, while the non-visual branch includes body pose, bounding boxes, and ego-vehicle speed. Extensive ablation studies on the PIE dataset demonstrate that surface normal maps provide consistent performance gains. When combined with contextual features, they improve F1 score and AUC by 3.85% and 2.35%, respectively, and yield more stable predictions up to two seconds before crossing events. These results highlight the importance of geometry-informed features as an effective complement to conventional visual inputs.