LC-KDE: A novel scheme for Wi-Fi localization

Signal strength difference (SSD) is widely utilized as the feature for Wi-Fi fingerprint localization to tackle the heterogeneity between training device and target device, but the correlation between SSDs is largely ignored. In this paper, a novel scheme named LC-KDE is proposed. It utilizes local Fisher discriminant analysis (LFDA) to transform the original SSDs into weakly correlated features, with which the SSD selection process is eliminated and the complexity of model training is reduced. Moreover, a compressed version of multivariate kernel density estimation (mKDE) is adopted to avoid over-fitting when training the probability density function (PDF) of the features. To validate the practical performance of LC-KDE, an experiment system is developed using commercial devices and measurements are conducted in an indoor setting. The results indicate LC-KDE provides superior performance compared to existing methods.