Future-State Guided Learning for Early Pedestrian Crossing Prediction

Reliable early pedestrian crossing prediction is essential for safe automated driving. Existing methods usually predict crossing or noncrossing directly from a short early observation. Pedestrian behaviour several seconds before crossing onset often provides limited evidence of the intended action, while a later observation segment of the same pedestrian’s track provides clearer and more discriminative information. In this paper, we use this later segment only during training to provide more informative supervision. We introduce Future-State Guided Learning (FSGL), a two-stage approach. First, a future-state branch learns a compact representation of the pedestrian’s crossing-related state from the later segment, after which its parameters are frozen. Second, a prediction model learns to estimate this future representation from the early observation and use it to predict crossing intention. During inference, FSGL uses only the early observation and does not require the later observation or the future-state branch. On the Pedestrian Intention Estimation (PIE) benchmark, FSGL improves a matched early-only baseline by 6.8% in relative AUC and 11.8% in relative F1 score, while reducing expected calibration error by 59.2%, from 0.071 to 0.029. Ablation studies show complementary benefits from the frozen future-state branch, sample-level alignment, and multimodal context. These results demonstrate that short-term future pedestrian behaviour provides effective supervision for earlier crossing prediction.