Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Federated Learning for Assisting the Visually-Impaired Using Augmented Reality

Vlaeminck, Robbe LU and Verheijen, Joran LU (2025) EITM01 20251
Department of Electrical and Information Technology
Abstract
This thesis investigates the creation and assessment of a Federated Learning (FL) system which
improves real-time object recognition for visually-impaired applications through embedded AI and
edge computing. The research divides into two distinct sections. The initial phase of the FL
validation used a CNN trained on the MNIST dataset to allow controlled testing of model updates
and communication between clients and server. This is for example, relevant for text-to-speech
conversion to allow the visually-impaired to be able to read. The Jetson Nano devices operated
as local clients which exchanged information with a central server, a computer, to train the model
jointly while maintaining data confidentiality.
The second phase... (More)
This thesis investigates the creation and assessment of a Federated Learning (FL) system which
improves real-time object recognition for visually-impaired applications through embedded AI and
edge computing. The research divides into two distinct sections. The initial phase of the FL
validation used a CNN trained on the MNIST dataset to allow controlled testing of model updates
and communication between clients and server. This is for example, relevant for text-to-speech
conversion to allow the visually-impaired to be able to read. The Jetson Nano devices operated
as local clients which exchanged information with a central server, a computer, to train the model
jointly while maintaining data confidentiality.
The second phase evaluated obstacle detection aimed at enabling visually-impaired individuals
to independently navigate unfamiliar outdoor environments, assessing both performance and energy
efficiency across various FL configurations.A custom lightweight CNN received specific training to
match the hardware constraints of the Jetson Nano. The model served as the basis for multiple
tests which evaluated centralized and stand-alone and FL systems based on accuracy and energy
usage and communication requirements.
The research shows that FL provides a suitable solution for edge AI applications, e.g., AR/VR
glasses, because it maintains privacy protection while achieving high accuracy and energy efficiency.
The research demonstrates how lightweight models work together with decentralized training methods on Jetson Nano devices without needing centralized infrastructure or extensive resources. (Less)
Please use this url to cite or link to this publication:
author
Vlaeminck, Robbe LU and Verheijen, Joran LU
supervisor
organization
alternative title
Federated Learning for Assisting the Visually-Impaired Using Augmented Reality
course
EITM01 20251
year
type
H1 - Master's Degree (One Year)
subject
keywords
Federated Learning Jetson Nano Visually-Impaired application MNIST Outdoor object detection
report number
LU/LTH-EIT 2025-1070
language
English
id
9201445
date added to LUP
2025-06-18 13:47:04
date last changed
2025-06-18 13:47:04
@misc{9201445,
  abstract     = {{This thesis investigates the creation and assessment of a Federated Learning (FL) system which
improves real-time object recognition for visually-impaired applications through embedded AI and
edge computing. The research divides into two distinct sections. The initial phase of the FL
validation used a CNN trained on the MNIST dataset to allow controlled testing of model updates
and communication between clients and server. This is for example, relevant for text-to-speech
conversion to allow the visually-impaired to be able to read. The Jetson Nano devices operated
as local clients which exchanged information with a central server, a computer, to train the model
jointly while maintaining data confidentiality.
The second phase evaluated obstacle detection aimed at enabling visually-impaired individuals
to independently navigate unfamiliar outdoor environments, assessing both performance and energy
efficiency across various FL configurations.A custom lightweight CNN received specific training to
match the hardware constraints of the Jetson Nano. The model served as the basis for multiple
tests which evaluated centralized and stand-alone and FL systems based on accuracy and energy
usage and communication requirements.
The research shows that FL provides a suitable solution for edge AI applications, e.g., AR/VR
glasses, because it maintains privacy protection while achieving high accuracy and energy efficiency.
The research demonstrates how lightweight models work together with decentralized training methods on Jetson Nano devices without needing centralized infrastructure or extensive resources.}},
  author       = {{Vlaeminck, Robbe and Verheijen, Joran}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Federated Learning for Assisting the Visually-Impaired Using Augmented Reality}},
  year         = {{2025}},
}