Indoor localization algorithm based on geometric deep learning

Popular machine learning based fingerprint localization methods often struggle to effectively capture non-Euclidean characteristics present in fingerprint data, while geometric deep learning can effectively process such data. In this paper, we propose a geometric fingerprinting based graph neural network indoor localization algorithm (GFGNN), which is models access points (APs) and reference points (RPs) using received signal strength (RSS) fingerprint. This approach maximizes the utilization of the unstructured nature of fingerprint data to enhance indoor localization accuracy and stability in dynamic environments. The algorithm establishes the fingerprint data as a graph feature representation, we first employ a graph convolutional network at the AP level to aggregate RSS values containing spatial relationships. Subsequently, graph isomorphism networks are employed at the RP level to further extract and update the aggregated fingerprint features. Finally, a multi-layer Perceptron is utilized to regressively predict the localization of the target to be located. We evaluate the proposed GFGNN on a self-built dataset, and the localization accuracy remains within 0.43 meters at the 80th percentile of the cumulative distribution function, with stable localization performance even in dynamic scenarios.