Adaptive Sampling for Fingerprinting Localization

Accurate fingerprint localization requires extensive numbers of measurements to build a comprehensive data set. We consider adaptive sampling to reduce the cost of measurement campaigns. Existing adaptive sampling algorithms target scalar or low-dimensional features and do not generalize to high-dimensional Channel State Information (CSI) features. To address this challenge, we create a radio environment map using Kriging interpolation. Sample points with high variances are identified in the 2D coverage area. The fingerprints of these points are obtained and used to adaptively update the data set. Simulation results demonstrate the effectiveness of proposed adaptive sampling strategy in improving localization accuracy while significantly reducing data collection effort. Compared to uniformly random sampling, the proposed adaptive sampling method enhances localization accuracy by 35% in root mean squared error when using weighted K nearest neighbor regression. An equivalent localization performance is achieved with 70% fewer training points compared to sampling uniformly at random.