Enhancing Wi-Fi localization of IoT devices in multistorey buildings

Abstract

The Internet of Things (IoT) has proliferated in various sectors, such as cities, healthcare, manufacturing, construction, and others, enriching our daily interaction with the physical world. Incorporating location sensing into the IoT ecosystem advances location and contextbased services. This has already been observed in recently progressed satellite-based outdoor positioning systems, resulting in navigation, tracking, and advertising applications. However, the occlusion of satellite signals indoors has forced researchers to investigate alternate wireless technologies and techniques for indoor localization. Wi-Fi fingerprinting has become a de facto method in multistorey buildings primarily due to its use of pre-built infrastructure, scalability, accuracy, and cost. Enhancing this method can extend its application to more resourceconstrained IoT devices.
This work reviews the relevant literature on indoor Wi-Fi fingerprinting localization approaches to identify how they can be improved and utilized for the localization of such devices in multifloor settings. Additionally, it proposes a framework for real-time localization and experiments on public datasets. Several preprocessing techniques are compared to identify appropriate combinations. Preprocessed fingerprints with these combinations are trained on several Convolutional Neural Network (CNN) architectures, and the trained models are compressed using Tensorflow Lite. After running inference on compressed models, the average localization time and floor-level localization accuracy are compared. Results on a public dataset show that models smaller than 500 KB in size can achieve above 90% floor-level accuracy with an average localization time of less than 5 ms. This confirms the applicability of the method on computationally constrained IoT devices.
Localizing additional devices with pre-existing fingerprinting methods optimizes their use, making it an affordable solution for various applications like asset tracking, robot navigation, geo-fencing, and occupancy sensing. Moreover, this work's framework, techniques, and architectures contribute to a deeper understanding of using deep learning to implement Wi-Fi fingerprinting-based indoor localization.

Date of Award	2024
Original language	English
Awarding Institution	Western Sydney University
Supervisor	Seyed Shahrestani (Supervisor) & Chun Ruan (Supervisor)