Distributed Dynamic Map Fusion For Intelligent Networked Vehicle Systems Via Federated Learning

The intelligent networked vehicle systems(INVSs)is an emerging vehicle-edgecloud system that can achieve cooperative perception,map management,planning and maneuvering tasks via vehicle-to-everything communication.Among all the tasks,distributed map management aims to enlarge sensing ranges and improve sensing accuracies.The research on dynamic map fusion has to deal with the following technical challenges:uncertainties of sensing data at each individual vehicle due to sensor limitation and environmental complexity;uncertainties of mobile object feature models for construction of dynamic maps;lack of labels of new data samples collected at vehicles.This thesis aims to develop a dynamic map fusion strategy to make full use of the sensor data of different autonomous vehicles and eliminate the uncertainties of perception;design a model parameter updating algorithm,under the premise of ensuring data security,complete the data labeling and model training at the same time to improve the models.In this work,we propose a distributed dynamic map fusion framework via federated learning.A three-stage dynamic map fusion algorithm at object level,a federated learning algorithm for point clouds,and an ensemble knowledge distillation algorithm are developed respectively.The proposed dynamic map fusion framework can deal with map fusion and model training as a unified problem.The developed ensemble knowledge distillation algorithm can transfer knowledge from the fusion results to the feature models.We have used the CARLA autonomous driving simulation platform and designed evaluation benchmarks to verify the proposed distributed dynamic map fusion framework.The experimental results show that the proposed dynamic map fusion framework can effectively reduce dynamic map errors caused by sensing uncertainties;the point clouds based federated learning algorithm can effectively improve the quality of object feature models;and the ensemble knowledge distillation algorithm can effectively solve the problem of missing labels for online learning.