WIFIND: Enhancing Wi-Fi Fingerprint Indoor Localization with a Spatially Conditioned Diffusion Model-Based Data Augmentation

Fingerprint-based indoor positioning systems require exhaustive site-surveying processes to establish the radio maps of wireless signals, such as Wi-Fi, at the target sites for localization, which are extremely time-consuming. Unfortunately, a rough, coarse-grained site-surveying run gathers inadequate radio fingerprints to produce lower-resolution radio maps. To address such a time-versus-accuracy dilemma, this paper proposes WIFIND, a novel approach employing diffusion models to generate additional details to reinforce the collected fingerprints. WIFIND employs a diffusion model trained on real-world fingerprint data to generate synthetic Channel State Information (CSI) that mimics actual Wi-Fi signals. Using a Spatially Conditioned Diffusion Model (SCDM), SCDM generates synthetic CSI data based on the given position, ensuring the data accurately reflects the spatial variations of Wi-Fi signals within the environment. This results in a richer and more diverse dataset, enhancing the model's ability to localize devices precisely. Extensive experiments demonstrate that WIFIND not only improves upon the baseline models trained with actual data but also outperforms existing state-of-the-art methods for CSI data augmentation. Preliminary results on the Ultra Dense Indoor MaMIMO dataset show that WIFIND achieves improvements ranging from 47.56% to 78.53% compared to the baseline, respectively.