An Efficient Radio Map Construction Method Using Region-Aware Adaptive Ensemble Regression for Indoor Localization

Abstract

The increasing adoption of indoor localization has increased demand for location-based services (LBS) to enhance daily efficiency. Fingerprinting is a widely used method due to its high accuracy and resilience to multipath fading, but it requires dense sampling and extensive data collection, making large-scale site surveys labor-intensive and often impractical. To alleviate these challenges, interpolation and regression-based methods have been explored for their scalability and reduced manual effort. However, these approaches often struggle to model the complex, multimodal nature of real-world Received Signal Strength Indicator (RSSI) distributions. To address this limitation, we propose a novel Region-Aware Adaptive Ensemble Regressor (RAER) for generating synthetic RSSI values. RAER adapts to the statistical characteristics of the training data to efficiently produce high-quality virtual RSSI maps with low computational cost. The model integrates two components: 1) a clustering algorithm that partitions the radio map based on fingerprint signal similarity, and 2) an adaptive ensemble regression framework that combines multiple base regressors using weighted averaging, where weights are inversely proportional to each regressor’s mean absolute error (MAE). By prioritizing regressors with lower localization error, RAER improves the realism of synthetic RSSI data and enhances radio map accuracy. Experimental evaluations conducted in a multi-story university building demonstrate that RAER can reconstruct radio maps, which reduces the need for site surveys by up to 25%. Furthermore, it improves localization accuracy by 5.06%, outperforming existing methods and offering a scalable and practical solution for fingerprint-based indoor positioning systems.