Data augmentation models for improved indoor positioning accuracy using RSS Fingerprinting

Abstract

Fingerprint-based indoor localization is a widely used technique for estimating the position of a device in environments where GPS is unavailable, such as inside buildings. This method maps the measured fingerprints, typically Wi-Fi signal strengths, against a database maintained by the localization service provider. However, a key challenge in using RSS fingerprinting is the requirement for large amounts of data to ensure accurate positioning, which is often time-consuming and costly to collect. To address this limitation, this research explores the application of data augmentation techniques to generate additional training data, improving the performance and accuracy of indoor positioning systems. By applying various data augmentation methods, this work aims to overcome the constraints of data deficit and improve localization accuracy, particularly in complex, real-world environments where data collection is limited. The study provides a comparative analysis of different augmentation models applied to RSS fingerprint data to identify the most effective techniques for improving indoor localization. The study demonstrates that Linear Interpolation achieved substantial accuracy improvements (up to 32.16%) for structured datasets, while Generative Adversarial Networks (GANs) provided competitive performance (14.90%) and excelled in sparse scenarios. Additionally, findings highlight the importance of balancing the amount of augmented data and spatial coverage to avoid diminishing returns. These results emphasize the practical benefits of data augmentation techniques in reducing data collection costs and improving localization performance, highlighting their potential for application in diverse environments.