Research on smart connected vehicles has recently targeted the integration of vehicle-to-everything (V2X) networks with Machine Learning (ML) tools and distributed decision making. Among these convergent paradigms, Federated Learning (FL) allows the vehicles to train a deep ML model collaboratively, by exchanging model parameters (i.e., neural network weights and biases), rather than raw sensor data, via V2X links. Early FL approaches resorted to a server-client architecture, where a Parameter Server (PS) acts as edge device to orchestrate the learning process. Novel FL tools, on the other hand, target fog architectures where the model parameters are mutually shared by vehicles and synchronized in a distributed manner via consensus algorithms. These tools do not rely on the PS, but take advantage of low-latency V2X links. In line with this recent research direction, in this paper we investigate distributed FL methods for augmenting the capability of road user/object classification based on Lidar data. More specifically, we propose a new modular, decentralized approach to FL, referred to as consensus-driven FL (C-FL), suitable for PointNet compliant deep ML architectures and Lidar point cloud processing for road actor classification. The C-FL process is evaluated by simulating a realistic V2X network, based on the Collective Perception Service (CPS), for mutual sharing of the PointNet model parameters. The performance validation considers the impact of the degree of connectivity of the vehicular network, the benefits of continual learning over heterogeneous training data, convergence time and loss/accuracy tradeoffs. Experimental results indicate that C-FL complies with the extended sensors use cases for high levels of driving automation, it provides a low-latency training service, compared with existing distributed ML approaches, and it outperforms ego learning with minimal bandwidth usage.

最近，智能互联汽车的研究目标是将车对万物 (V2X) 网络与机器学习 (ML) 工具和分布式决策进行集成。在这些收敛范式中，联邦学习 (FL) 允许车辆通过 V2X 链接交换模型参数（即神经网络权重和偏差）而不是原始传感器数据来协作训练深度 ML 模型。早期的 FL 方法采用服务器-客户端架构，其中参数服务器 (PS) 作为边缘设备来协调学习过程。另一方面，新型 FL 工具针对雾架构，其中模型参数由车辆相互共享并通过共识算法以分布式方式同步。这些工具不依赖于 PS，而是利用低延迟 V2X 链接。根据最近的研究方向，在本文中，我们研究了基于激光雷达数据增强道路用户/对象分类能力的分布式 FL 方法。更具体地说，我们提出了一种新的模块化、分散的 FL 方法，称为共识驱动的 FL (C-FL)，适用于符合 PointNet 的深度 ML 架构和用于道路行为者分类的激光雷达点云处理。C-FL 过程是通过模拟真实的 V2X 网络来评估的，该网络基于集体感知服务 (CPS)，用于相互共享 PointNet 模型参数。性能验证考虑了车辆网络连接程度的影响、持续学习对异构训练数据的好处、收敛时间和损失/准确性权衡。