ULPTP-ID: A Unified Learning Framework Under Imbalanced Data Distributions for Pedestrian Trajectory Prediction

For pedestrian trajectory prediction, the inherent long-tail distribution of pedestrian behaviors in real-world traffic environments inevitably leads to imbalanced data distribution in collected datasets, causing models to be dominated by abundant head samples (e.g., straight walking) during training and thus underperformed on rare but challenging long-tail cases (e.g., sudden crossings or unusual maneuvers). Existing approaches often tackle this issue by separately augmenting tail samples or decoupling head and tail training, which increases complexity and limits practical applicability. To overcome these limitations, we propose ULPTP-ID, a unified learning framework that achieves consistently superior trajectory prediction performance for pedestrians across both head and tail samples. First, to provide samples with explicit categorical labels, we propose a classification method based on prediction difficulty estimation, where the resulting sample-wise categorization serves as the basis for a multi-granularity contrastive learning strategy to enhance feature representations. Afterwards, a context-conditioned adaptive decoder allows the model to adjust trajectory generation based on the learned contextual features. Experiments on the ETH-UCY and SDD benchmarks demonstrate that ULPTP-ID achieves strong overall accuracy and significantly enhances prediction performance in long-tail scenarios.