DeepRSSI: Generative Model for Fingerprint-Based Localization

Abstract

In this paper, we present an innovative methodology for generating virtual received signal strength indicator (RSSI) fingerprint maps to improve indoor localization systems and wireless communication systems using RSSI. Focusing on the challenge of extensive labor and time required in traditional data collection, we propose a generative model that combines customized attention mechanism with a conditional variational autoencoder (cVAE), leveraging datasets compiled from direct measurements of RSSI values from different access points in a real-world indoor environment. Our model uniquely synthesizes high-quality virtual RSSI maps, significantly reducing the need for extensive physical data collection while enhancing the accuracy and efficiency of indoor positioning systems. By integrating measured data with innovative data generation techniques, our approach offers a novel solution to indoor localization challenges. In addition, this model can augment high-quality synthetic data for indoor wireless signals to expand the volume of available data. We quantitatively demonstrate the effectiveness of our model, showing an average improvement of over 40% in Euclidean distance errors across several machine learning algorithms compared to existing methods. Our experiments validate that the virtual RSSI fingerprint map yields accurate position estimates, with performance enhancements observed in algorithms that confirm the utility in real-world scenarios. The contribution of our research improves indoor localization systems by improving indoor positioning accuracy and addresses the limitations of traditional fingerprinting methods, setting the stage for future innovations in wireless communication.