Trajectory prediction training scheme in vehicular ad-hoc networks based on federated learning

Abstract

Vehicle trajectory prediction (TP) plays a crucial role in autonomous driving systems and is the core elements to improve traffic conditions and reduce the risk of accidents. However, establishing accurate TP models in real-time still presents numerous challenges. Firstly, relying on a central server for real-time model updates not only poses the risk of privacy leakage, but also increases the resource load with frequent data interactions. In addition, considering the changes in driving habits and traffic environment during vehicle traveling, the use of static TP models can lead to underfitting. Therefore, this paper proposes a TP Training Scheme in Vehicular Ad-hoc Networks Based on Federated Learning (FL-VANETs). In FL-VANETs, a Vehicle Relevance-Oriented Collaborative Vehicle Node Selection Algorithm (VR-CVNS) is designed to ensure that the reasonable construction of a decentralized Ad-hoc networks and enable serverless computing, thereby optimizing computational and communication efficiency. Additionally, through the FL framework, vehicle computing tasks are categorized, ensuring privacy security in the VANETs and enabling dynamic training of the TP model during vehicle movement, thereby improving the model’s predictive accuracy. The effectiveness and improvement of the method are verified through experiments and simulations.