DumbLoc: dumb indoor localization framework using WiFi fingerprinting

• Main Authors: Narasimman, Srivathsan Chakaravarthi, Alphones, Arokiaswami
• Other Authors: School of Electrical and Electronic Engineering
• Format: Article
• Language: English
• Published: 2024
• Subjects: Engineering; Dimensionality reduction; Estimation; Wi-Fi; Fingerprinting; Floor prediction; Indoor localization; Machine learning; Indoor positioning; Representation learning; Training time; Transfer learning
• Online Access: https://hdl.handle.net/10356/176574
• Institution: Nanyang Technological University

Extant indoor positioning techniques based on received signal strength (RSS) fingerprinting have achieved accurate positioning results. They leverage existing open-source datasets for training and comparing their model performance. However, the models trained on one building cannot be used for another since these models learn the relationship of a specific set of RSSs to the building and floor locations necessitating the expensive, time-consuming process of fingerprinting. Even when we consider the individual datasets producing these excellent results, a lot of painstaking optimization is required which precludes a lot of people trying to implement indoor positioning quickly. Most of the input RSS vector is empty with redundant information and the static class labels used for buildings and floors make the models unusable on other buildings. We propose a machine learning-based framework that uses RSS values from the strongest access point (AP) signals and normalized output labels to combat this issue. The framework was used on the open-source UJI dataset using less than 5% of the 520 APs to achieve 94.15% and 8.45 m floor prediction accuracy and mean positioning error respectively without any optimization. This technique was reused on 10 other public datasets and achieved an average floor estimation accuracy of 91.93% when trained with new data and 88.68% without any new data compared to 87.1% of the closest competitor.