Evaluation and Application of Modified Fixed-Radius Near Neighbours Searches

Abstract

Indoor positioning systems (IPS) are applications that provide estimated positions of objects in enclosed environments. Among various positioning techniques for IPS, Wi-Fi fingerprinting has emerged as one of the most widely used approaches. This method relies on the collection of Wi-Fi signal measurements known as fingerprints at reference locations to create a database, which is then used to estimate the position of a target based on similarity matching. This matching task has been adopted widely with the use of the k-Nearest Neighbours (kNN) model but its accuracy can be limited due to certain environmental conditions. To deliver greater positioning accuracy, researchers have come up with numerous variants of kNN that aim to enhance neighbour selection, weighting, and adaptability. Fixed-Radius Near Neighbours (FRNN) is a modified version of kNN that utilizes a different neighbour-selecting technique with radius. While FRNN has not been deeply studied in the context of IPS, this thesis explores further whether it can be a potential candidate for indoor positioning. This thesis also proposes two new versions of FRNN: Adaptive Radius Near Neighbours (ARNN) and Weighted Adaptive Near Neighbours (WARNN). They are examined alongside FRNN under different hyperparameters against 13 kNN models on 22 Wi-fi datasets in this study. The results show that whereas FRNN is not highly comparable to the baseline models, ARNN performs moderately better than FRNN but not quite competitive. However, WARNN displays the most promising performance that outperforms all baseline models on 22 datasets, especially when combined with the Cityblock distance and a customized weighting function, showing that it is a strong candidate for indoor positioning.