Federated Edge Intelligence for Real-Time Obstacle Detection in Vehicular Networks

Abstract

Obstacle detection is the perception task where delay is measured against the laws of motion, and cloud centred inference cannot meet that budget without also spending bandwidth and exposing sensor data. This chapter develops Federated Edge Intelligence, a three tier design uniting Federated Learning with Multi-Access Edge Computing so raw data never leaves the vehicle, aggregation happens at the roadside, and cloud coordination stays loose. It names four open problems: convergence under heterogeneity, communication under mobility, end to end co-design, and the joint pursuit of privacy and robustness. It then proposes HARP-FEI, which answers all four through a proximal local objective, mobility aware client weighting, coordinate wise robust aggregation and quantised update transmission, with a full algorithm, a system model and a complexity analysis. Against centralised, FedAvg, FedProx and edge only baselines, HARP-FEI is projected to approach centralised accuracy while cutting per round upload, converging in fewer rounds and holding accuracy under attack and tightened privacy.