Decentralized Federated Learning for Road User Classification in Enhanced V2X Networks

Federated Learning (FL) techniques are emerging in the automotive context to support connected automated driving services. Yet, when applied to vehicular use cases, conventional centralized FL policies show some drawbacks in terms of latency and scalability. This paper focuses on decentralized FL solutions, which attempt to overcome such limitations, by introducing a distributed computing architecture: vehicles exchange the parameters of a shared Machine Learning (ML) model via V2V links, without the need of a central orchestrator. Sharing all ML parameters, however, might not be feasible when minimal V2X bandwidth usage is required or the model is highly complex (e.g., extremely deep networks) as in advanced scenarios for high levels of automation. We thus propose a modular decentralized FL solution and we discuss its application to road user classification in a cooperative vehicular sensing use case. The proposed FL solution performs the point cloud processing of Lidar sensor inputs using a PointNet compliant architecture. It enables the exchange of a subset of the model parameters, namely selected ML model layers, optimized for communication efficiency, convergence and accuracy. We use real sensor data extracted from a publicly available dataset to validate the method, focusing on non-uniform scenarios where sensor data are highly unbalanced across the connected vehicles. For all cases, FL is shown to outperform the ego-sensing approach with minimal bandwidth usage.