Indoor Localisation with Limited Resources for Elderly Health Monitoring at Home

: Techniques for Limited Training Data and Model Optimisation for Resource-Constrained Low-Power Devices

Abstract

Indoor localisation enables user tracking within homes, offering numerous benefits for health
applications, particularly for elderly individuals who often require additional support. Such
systems can monitor mobility patterns to detect early signs of declining health, track
recovery progress after surgery, and identify prolonged inactivity to ensure timely intervention.
This health monitoring enhances safety, promotes independence, and supports proactive care,
ultimately improving the quality of life for elderly users.

To be practical, efficient, and scalable for home use, indoor localisation systems must be
non-intrusive, seamlessly integrate into daily life, and require minimal effort during deployment.
Data collection during system setup should be quick and simple, while energy efficiency is
essential to reduce the need for frequent charging, which can be challenging for elderly users.
However, existing systems often rely on tedious data collection and annotation processes, which
are particularly difficult for individuals with mobility limitations, such as elderly users, resulting
in limited training data. Additionally, these systems are often centralized, which limits their
scalability for deployment across various homes and raises privacy concerns. These limitations
create barriers to practical adoption.

This thesis addresses these challenges through the development of Received Signal Strength
Indicator (RSSI) augmentation methods and training data-sharing frameworks that use machine
learning techniques to maximise the utility of limited training data, minimizing elderly user
effort during setup. The proposed augmentation pipeline significantly improves localisation
performance, increasing the macro F1 score by up to 23% when using as little as four seconds
of data per room. When data is transferred from other sources, the system achieves up to 81%
macro F1, demonstrating strong generalisation under limited supervision.

Furthermore, this thesis explores transitioning indoor localisation models from centralised
servers to low-power microcontrollers. These devices, which consume significantly less power
than other edge devices, offer improved scalability, enhanced privacy, and better energy efficiency.
To achieve this, the thesis optimises indoor localisation models to create lightweight and efficient
models that meet the constraints of low-power microcontrollers, achieving a model size under 32
KB while maintaining up to 85% macro F1 score. The proposed methods were validated using
real-world datasets collected in typical UK homes, showcasing their effectiveness in addressing
real-world challenges.

With a focus on low-cost design and seamless deployment, the indoor localisation system
supports continuous health monitoring, enabling elderly individuals to maintain independence
while improving their quality of life and reducing the burden on caregivers. Overall, this thesis
provides a foundation for scalable, energy-efficient, and easily deployable indoor localisation
systems that can be further advanced by future research.

Date of Award	17 Jun 2025
Original language	English
Awarding Institution	University of Bristol
Sponsors	Royal Thai Government
Supervisor	Ryan McConville (Supervisor) & Ian J Craddock (Supervisor)