LocSwayamwar: Finding a Suitable ML Algorithm for Wi-Fi Fingerprinting Based Indoor Positioning System

Abstract

Indoor localization has been a challenging problem for over a decennium. Wi-Fi Fingerprinting based solutions stand out in comparison with Angle-Of-Arrival (AOA), Time-Of-Arrival (TOA), Time-Difference-Of-Arrival (TDOA) approaches as they inherently incorporate radio propagation models in fingerprints (FP) which provide more realistic information than radio signal propagation models as well as do not need extra hardware. Diverse Location Based Services (LBS) heavily rely on the performance of localization algorithms used for pattern matching with the collected FP database. This work investigates the performance of several machine learning algorithms as a multiclass classifier for room-level indoor localization including K*, k-NN, Random Forest, FURIA, Multi-Layer Perceptron, and J48. We report results of top five algorithms along with five algorithms selected from various algorithmic categories obtaining an accuracy greater than 95%. Data was generated by collecting 14,080 fingerprints from 20 Access Points at 180 reference points in 1209 m2 area of Software Engineering Center, University of Engineering and Technology (UET), Lahore to construct real-world FP dataset. The results obtained indicate that the best performance is achieved by K* followed by k-NN, Random Forest, FURIA, Multilayer Perceptron, J48 with accuracies 99.52, 99.06, 98.76, 97.26, 97.05, and 95.91% respectively.