polito.it
Politecnico di Torino (logo)

Data-Driven Road Hazard Detection for Automated Driving

Giuseppe Di Giacomo

Data-Driven Road Hazard Detection for Automated Driving.

Rel. Carla Fabiana Chiasserini. Politecnico di Torino, Corso di laurea magistrale in Ict For Smart Societies (Ict Per La Società Del Futuro), 2021

[img]
Preview
PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (2MB) | Preview
Abstract:

Recently, Machine Learning and Deep Learning have become the state-of-theart techniques in different domains, such as computer vision, natural language processing and speech recognition. One of the main drawbacks is the necessity of a huge amount of data to be trained correctly. So far, the traditional way consists in gathering samples at only one central infrastructure, which trains the model. However, if data are recorded by different devices, this process presents two main disadvantages: first, transmitting all of them requires a great consumption of network resources and, second, it can expose sensitive information. In such a context, a new procedure has emerged: Federated Learning. Basically, it leverages the computational power of the agents collecting data, which, instead of uploading them to the central server in charge of the model training, keep samples locally and use them for the learning process. Iteratively, the server only aggregates all the received models. ML and DL are being increasingly studied also in the transportation field, as they can be used for ADAS and autonomous driving. Other implementations are related to the intelligent transportation system, as these techniques may be applied, for instance, for travel time estimation, to decrease fuel consumption and for traffic optimization. Thanks to the many contexts in which it may be used, FL has gained interest also in vehicular networks. In this work a novel method, named Hybrid Federated Learning, is introduced: with respect to the standard FL, users are gathered in groups. For each cluster, the server receives a model that has been sequentially trained by devices of the ensemble. Compared to the vanilla FL algorithm, the presented approach shows better performances in terms of global iterations and number of transmissions. Also, the communication load on the central server is alleviated.

Relatori: Carla Fabiana Chiasserini
Anno accademico: 2020/21
Tipo di pubblicazione: Elettronica
Numero di pagine: 63
Soggetti:
Corso di laurea: Corso di laurea magistrale in Ict For Smart Societies (Ict Per La Società Del Futuro)
Classe di laurea: Nuovo ordinamento > Laurea magistrale > LM-27 - INGEGNERIA DELLE TELECOMUNICAZIONI
Ente in cotutela: TELECOM ParisTech - EURECOM (FRANCIA)
Aziende collaboratrici: Eurecom
URI: http://webthesis.biblio.polito.it/id/eprint/17844
Modifica (riservato agli operatori) Modifica (riservato agli operatori)