Toward Regression-Based Estimation of Localization Errors in Fingerprinting-Based Localization

Location information is a valuable source of context that can be utilized by end-user applications and wireless networks to optimize their performance and usability. When used, location information should ideally be considered jointly with the estimate of its accuracy. Most of the current approaches for estimating the accuracy rely on performing a static performance benchmark of a localization solution in a deployment environment, which fails to capture the dynamic nature of the environment. We address this problem for fingerprinting-based localization by grounding the estimation of localization errors on the low-level features, i.e. the RSS values from APs used in fingerprinting. We use these low-level features measured at different locations in an environment, as well as their respective localization errors, to train different regression models, allowing us to predict the localization errors at new locations, given new observed values of the low-level features at these locations. Our evaluation results show substantially better performance of the proposed regression-based estimation of localization errors compared to static performance benchmarks.