Enhanced Weighted K-nearest Neighbor Positioning

Abstract

We consider fingerprinting-based localization in highly cluttered multipath environments with non-line-of-sight conditions, typical of indoor scenarios. Channel state information (CSI) from multiple Base Stations (BSs) is used to construct a fingerprint. We investigate the physical geometry of the k nearest neighbors found by feature distances, as well as possible enhancements to boost achievable positioning accuracy. We observe that the performance of Weighted K-Nearest Neighbor (WKNN) regression depends on the relation between the true position and its k nearest feature neighbors. Better accuracy is achieved when the true position is inside the convex hull of the k nearest neighbors, otherwise localization performance degrades. Consequently, we devise a neighborhood selection algorithm to increase the possibility of a point being inside the convex hull of the k nearest feature neighbors. WKNN localization is also affected by the weighting function used. To further improve performance, we consider a general framework to find the optimum weighting function, utilizing Laguerre polynomials. We benchmark performance against WKNN with exponential weight and deep neural network based localization. Simulation results show that the optimum weighting function with neighbor selection outperforms the benchmark algorithms.