Edge-assisted Federated Learning for Autonomous Vehicle Trajectory Prediction

The evolution of cars has always been driven by the aim of making transport safer for passengers and pedestrians, but also the need to improve the livability of the cities themselves and therefore the driving comfort of all drivers. This has led to the availability of vehicles equipped with a wide set of sensors for the monitoring of internal and external signals. In the near future, the trend is to inevitably increase the number of data collected from the vehicle itself, necessary to provide services as useful and effective as possible. With the born of applications based on the use of neural networks, there is a huge need of a large amount of data and computational power to make their use profitable. The most efficient way to train a neural network dedicated to a vehicular application is to use a server located at the edge of the network, because it can exchange information with vehicles through the mobile network, using low-latency communication. The main problems of this approach are the high computational power required by the training of the neural network and the high amount of data exchanged between vehicles and the server. In order to overcome these problems, a new approach has been envisioned, allowing to exploit a connected network of vehicles sharing their computational power to support the server in the training of a neural network. The work of this thesis focused on the development of a Federated Learning framework to test how a distributed learning approach can help to train a Deep Neural Network (DNN) for trajectory prediction in an urban environment. We leveraged urban mobility traces to select clients to participate in the Federated Learning (FL) process. The proposed framework makes use of three different components: SUMO [1], ms-van3t [2] and Flower framework [3]. The first two components have the role of generating an urban mobility trace file starting from real traffic data. The trace file, describing the behaviour of connected cars under the coverage of a mobile network, results from a simulation of a realistic urban mobility traffic environment made by vehicles communicating with seven LTE eNBs. The third component is a federated learning framework that was properly modified and tuned to leverage the output of the mobility trace file to select at each federated learning round a predefined number of clients reflecting a realistic behavior. Our simulations show that choosing FL clients based on real mobility traces outperforms, in terms of overall simulation time required to reach a certain accuracy threshold, a simulation with a synthetic scenario where a predefined and less dynamic selection policy for clients is used.