Kernel-based online prediction algorithms for indoor localization in Internet of Things

Abstract

Precise indoor position estimations of the targets play an increasingly important role in the enhancement of IoT-oriented applications. As an effective indoor localization technology, the fingerprint method has been extensively studied due to the widespread establishment of the wireless communication infrastructure. However, the fingerprint mismatch, which always leads to degradation of the positioning accuracy, commonly arises due to the diversity of the indoor environment and the existence of the abrupt noise in real-world applications. In this paper, a kernel adaptive filtering (KAF) assisted indoor position sensing framework based on the reproducing kernel Hilbert space (RKHS) model is exhibited, which is different from the conventional fingerprint-based positioning methods in two senses: first, in order to model a real-world environment, the linearity assumption in the traditional positioning approaches is replaced by the nonlinear counterpart; and second, a reproducing kernel function is employed to combat the impulsive noise and thus guaranteeing the robust positioning accuracy. Meanwhile, a generalized path-loss model is proposed to further enhance the accuracy performance of the KAF-based algorithms. The accuracy performance comparison between the proposed algorithms and widely-used machine learning methods such as KNN, naive Bayesian and the concept of transfer learning are carried out for different access points and different measurement areas. Experimental results exhibit that the proposed KAF framework-based algorithms can increase the positioning accuracy at least by 7% compared with the conventional fingerprint algorithms in representative indoor scenarios.