Offset-aware transformer for continuous Wi-Fi fingerprinting localization

Abstract

Wi-Fi fingerprint-based indoor localization is widely deployed due to its compatibility with existing wireless infrastructure. However, conventional fingerprinting is designed as a classification task over discretized reference points (RPs), making it difficult to infer continuous positions between RP grids. This structural limitation degrades both localization accuracy and spatial resolution, and retraining is required whenever new RPs are added, reducing scalability. In this paper, we propose Offset-Aware Transformer Localization (OTL), a new indoor localization framework that preserves the conventional RP grid-based structure while estimating continuous positions through offset regression. OTL employs a Transformer-based cross-attention mechanism, where measured RSS values are used as queries and radio map embeddings are used as keys and values, enabling the model to attend to spatially relevant RPs. The final position is predicted by regressing a continuous offset from the classified RP coordinate. Experiments conducted on the AI-Hub indoor localization dataset, collected from underground shopping malls and metro stations in Seoul, demonstrate that OTL improves localization accuracy compared to classification-based methods, achieving 38.02% accuracy and an average positioning error of 2.88 m. These results confirm that OTL effectively overcomes the inherent discretization constraints of fingerprinting and enables scalable localization without retraining when new RPs are added.