Machine Learning Based Indoor Localization Using a Representative k-Nearest-Neighbor Classifier on a Low-Cost IoT-Hardware

In order to make Internet of Things (IoT) applications easily available and cheap, simple sensors and devices have to be offered. To make this possible, our vision is to use simple hardware for measurements and to put more effort in the signal processing and data analysis to the cloud. In this paper, we present a machine learning algorithm and a simple technical implementation on a hardware platform for the localization of a low accuracy microphone via room impulse response. We give a proof-of-concept via a field test by localization of multiple positions of the IoT device. The field test shows that the recorded signals from the same source are unique at any position in a room due to unique reflections. In contrast to other methods, there is no need for high accuracy microphone arrays, however, at the expanse of multiple measurements and training samples. Our representative k-nearest-neighbor algorithm (RKNN) classifies a recording using a k-nearest-neighbor method (KNN) after choosing representatives for the KNN classifier, which reduces computing time and memory of the KNN classifier.