extendedGAN+ : improved data augmentation method for indoor localisation

Abstract

Indoor Positioning System (IPS) contains the same function as a Global Positioning System (GPS), which is to give the location of the device. IPS is built for indoor usage, while GPS is for outdoors. IPS can be useful for indoor tours, or even locating another device within the building. This project used Received Signal Strength Indicator (RSSI) emitted from the wireless Access Points (AP) to develop the IPS as many existing infrastructures, such as shopping malls, already have existing APs. An IPS consists of 2 parts: a phone application for the user, and the indoor localisation model to predict the location via RSSI. The use of APs for indoor localisation meant that no additional devices were needed for installation. The development of the indoor localisation model requires the use of deep learning. Deep learning imitates the way humans gain knowledge to make a prediction. It gains knowledge from the data we feed. A good localisation model requires a huge diversity of good data. This means that a huge amount of time and effort is required for data collection. To minimise the physical effort needed for data collection, this report proposes the use of extendedGAN+, which contains artificially generated data, also known as data augmentation, to improve the performance of deep learning. The proposed method was compared with some existing data augmentation methods. The extendedGAN+ leverage on the data aggregation method using Dirichlet and Wasserstein Generative Adversarial Network with Gradient Penalty (WGANGP) to generate augmented dataset. The result uses the public dataset UJIndoorLoc, and the selfcollected data at a building complex. The extendedGAN+ performs better than the other augmentation methods, giving an improved performance of the localisation model by a maximum of 0.32 m. The results can be further improved with more tests on different datasets, such as by increasing the number of augmented data and optimising the localisation model.