Mitigating class imbalance in indoor localization through autoencoder driven data augmentation

Abstract

Indoor Localization Systems (ILS) are crucial for applications such as asset tracking, emergency response, and indoor navigation. Despite significant advancements in the field, challenges related to accuracy and class imbalance in localization datasets remain. This thesis explores the application of generative models, specifically autoencoders, to address these challenges. By focusing on the JUIndoorLoc and UJIIndoorLoc datasets, we aim to create synthetic data that can balance the datasets and improve localization accuracy. The methodology involves tuning the autoencoder models according to the specific characteristics of each dataset. For JUIndoorLoc, collected from the 4th floor of a university building, the autoencoder is trained on data with 51 labels and 172 access points (APs). The UJIIndoorLoc dataset, from the IPIN2015 competition, presents a more complex scenario with 520 APs. Here, the autoencoder is trained on selected labels with specific sample instances to generate balanced synthetic data. Our results demonstrate that using autoencoders for data augmentation leads to significant improvements in localization accuracy. The training and validation loss metrics show that the autoencoder models effectively learn from the training data and generalize well to new data. This is crucial for enhancing the performance of ILS in real-world scenarios where data imbalance and limited labelled instances are common.