Pedestrian Trajectory Prediction via Transformers: a Comparison between Federated and Centralized Training

Intelligent Transport Systems (ITS) rely increasingly on accurate perception and prediction capabilities to support safety-critical applications in urban environments. Among these tasks, pedestrian trajectory prediction plays a fundamental role, enabling vehicles to anticipate vulnerable road user behavior and mitigate potential collision risks. Recent advances based on Deep Learning (DL), and in particular Transformer architectures, have demonstrated strong predictive performance but typically rely on centralized training paradigms that require large-scale data aggregation, raising concerns related to scalability and excessive usage of communication resources in vehicular networks. This thesis investigates a Federated Learning framework for pedestrian trajectory prediction in ITS, with the objective of evaluating whether distributed training can achieve performance comparable to centralized learning. A Transformer-based prediction model is adapted and extended to operate in a federated setting, where multiple automated vehicles collaboratively train a shared model for pedestrian trajectory prediction without exchanging raw sensor data. Two model configurations with different settings are examined, together with multiple update strategies and local training–to–communication ratios, allowing a systematic analysis of the trade-offs between model accuracy and communication efficiency. The proposed framework is evaluated through extensive simulations using heterogeneous urban trajectory datasets, divided per geographic areas. Performance is assessed using standard trajectory prediction metrics, including the Mean Average Displacement (MAD) and Final Average Displacement (FAD), errors via a newly proposed metric called Miss Rate, devised to evaluate the reliability of trajectory prediction models under safety-critical constraints. Communication-level behavior is further analyzed through a novel communication-oriented metric called Occupancy Time, which quantifies the duration for which the wireless channel is occupied for model transmission.