Proactive Prediction of Child Pedestrian Trajectories Using Trajectory Unified Transformer (TUTR) Models

Abstract

Enhancing child pedestrian safety is crucial due to their unpredictable behaviors, which pose significant risks. Existing post-accident measures, such as traffic regulations and infrastructure improvements, face limitations in real-time applicability and cost-efficiency. This study proposes a deep learning approach using the Trajectory Unified Transformer (TUTR) model to predict the future trajectories of child pedestrians, enabling proactive safety interventions. The AI-HUB dataset, featuring hazardous child behaviors like sudden road appearances and abrupt direction changes, was used for training. Methodologies include video data preprocessing, trajectory extraction, and data augmentation to improve model performance. Evaluation metrics such as Average Displacement Error (ADE) and Final Displacement Error (FDE) demonstrate high prediction accuracy, with a 0.24 difference between ADE and FDE values. This research highlights the potential of integrating TUTR models into autonomous vehicles and infrastructure, providing real-time safety measures and reducing child pedestrian accidents.