Wi-Fi RSSI Fingerprint-Based Indoor Localization: Evaluating Regression and Generative Data Augmentation Under High Spatial Sparsity

Abstract

This thesis explores the claim that WiFi Received Signal Strength Indicator (RSSI) fingerprint-based indoor localization actually requires generative data augmentation, or whether the benefits it claims tend to offset poor baseline models. UTSIndoorLoc dataset (522 samples, 206 features of wireless access points, and 198 reference locations) was experimented on. A three-step assessment scheme was developed: (i) a comparison of four down stream regressors (kNN, SVR, XGBoost, and DNN) in full-feature and selected-feature conditions, (ii) a controlled sparsity stress test, with 10% - 90% location drop out, with farthest-point thinning, and (iii) a generative rescue benchmark, consisting of six models. The findings indicate that optimization of the baseline is important. At full density, XGBoost achieved the best performance with a mean Euclidean error (MEE) of 1.66m, clearly outperforming the competing regressors without any synthetic data. XGBoost was the most competitive until 50 percent omission, with kNN being more competitive at extreme sparsity. Generative augmentation was of little use in dense regimes, but was beneficial as the spatial coverage diminished. The largest rescue effect was achieved between GAN at 90% omission lowering MEE of 9.79m to 8.90m. Not every generative model, however, was predominant in all sparsity regimes, and augmentation did not in any case recover dense-regime accuracy. This evidence demonstrates that generative augmentation must be considered a tailored rescue plan to severe data sparsity as opposed to a universal element in WiFi localization pipelines.