A Lightweight Federated Prediction Approach for Urban VRU Movement Understanding in Autonomous Driving

Abstract

The growing use of autonomous vehicles in modern city transport systems shows that there is an urgent need for accurate short-term prediction of how pedestrians and cyclists will move, especially in mixed and crowded environments where movement intention, social interaction, and road layout keep changing, and this forms the main background and motivation of the study. Existing centralized learning techniques do not scale well because they face privacy rules, data ownership issues, and heavy communication requirements, which finally result in weak performance across different domains. To solve these difficulties, this research puts forward a new federated trajectory prediction approach that mixes onboard perception, lightweight tracking of detected objects, and a Social-LSTM prediction model that is improved using the FedProx algorithm, which becomes the main method contribution. The system first uses YOLO detection to find vulnerable road users, then uses SORT tracking to keep motion continuity, and trains the Social-LSTM locally while sharing only gradient updates for safe global aggregation without sharing raw sensing data. Experiments on ETH, UCY, SDD, and NuScenes datasets show reduced domain drift, better stability in different scenes, and improved ADE and FDE scores over centralized models, showing the results achieved. The concluding part states that this federated spatio-temporal learning system offers a scalable, privacy-safe, and ready-to-deploy solution for trajectory prediction, giving a new and practical step toward safer autonomous driving decisions.