AI assisted indoor localization

Abstract

Navigational systems are vital due to their prominence in many sectors such as humanitarian, construction, healthcare etc. With a growing number of new infrastructures due to urbanization, there is a need for new technology to overcome the inefficiencies that outdoor navigational systems, such as the Global Positioning System (GPS), face when applied indoors. The solution to this is Indoor Localization (IL). Many methodologies for IL have been experimented and resulted in the Wi-Fi fingerprinting approach being relied on the most. However, the issue faced with Wi-Fi fingerprinting pertains to the tedious collection of large amounts of fingerprint data, which requires a lot of manpower. The fingerprint data to collect is also sometimes unavailable. Machine learning models have been created to tackle the difficult data collection. However, the accuracy of these models can be greatly improved. Furthermore, they are computation-intensive and time-consuming.

In this report, the obstacles that Wi-Fi fingerprinting and traditional machine learning methods face will be overcome by relying on deep learning approaches. Different deep learning techniques and models such as deep neural network (DNN), convolutional neural network (CNN), as well as Transfer Learning (TL) are proposed, implemented, and evaluated to ensure satisfactory location prediction results are obtained. The deep learning models used are trained and evaluated on publicly available IL datasets such as the UJI Indoor dataset, as well as data collected from a building complex (BC). The eventual selected model comprising of an autoencoder and CNN, augmented with TL, is shown to be effective in different domains and for public deployment.