Signal Distribution Oriented Mean Functions in GPR based Fingerprint Augment

Abstract

WiFi fingerprinting-based localization is limited by heavy site survey in the offline phase, for which fingerprint augment based on gaussian process regression (GPR) is an effective solution. This paper proposes const-linear (CL) and quadratic polynomial (QP) mean function for GPR based on the cognition of signal distribution in indoor positioning scene. The proposed mean function is designed after comprehensively discussing the complexity of the mean function and the fitting accuracy of the received signal strength (RSS) distribution. Furthermore, we compare the performance of fingerprint database acquired by GPR with different mean functions in two typical positioning scenes and the results show that GPR with the proposed mean functions produce up to 53.3% lower fingerprint augment errors and 18.73% lower localization errors than the basic GPR. Our work confirms that when the selected prior mean is closer to the RSS distribution in the positioning scene, the GPR estimation result is more accurate, which is significative for GPR based fingerprint augment.