Autonomous Braking Algorithm for Rear-End Collision via Communication-Efficient Federated Learning

Abstract

Realizing driving safety is the fundamental goal pursued by artificial intelligence (AI)-enabled autonomous driving. However, due to the limited capacity of vehicles and the limited scenarios involved, the reliability and environmental adaptability of the current single-vehicle intelligence still need to be improved. In addition, driving knowledge only exists locally in each connected autonomous vehicle (CAV) and cannot be effectively reused or shared with other CAVs. To solve the above problems, this paper proposes to use federated learning (FL) to realize the collaboration of CAVs without revealing local data, thereby improving the accuracy of CAVs decision-making and driving safety. First, for the typical rear-end collision scenario, we propose a local decision-making model that comprehensively considers multiple influencing factors to fit the actual traffic environment. Then, we design a federated learning process for knowledge sharing. In particular, we propose a model similarity method to select high-quality local models for upload, thereby reducing communication overhead while improving the accuracy of the global model. Extensive simulations validate the performance of our proposal in reducing communication overhead and improving decision accuracy.