Analysis of WiFi localization techniques for kidnapped robot problem

Abstract

© 2022 IEEE.This work proposes an analysis of the earliest indoor localization techniques based on recurrent neural networks (RNN) like Gated Recurrent Unit (GRU) and Long-Short Term Memory (LSTM), including k-Nearest Neighbors (KNN) machine learning, to process WiFi received signal strength data (RSS) for the kidnapped robot problem (KRP). The proposed solutions uses processed data generated in a Webots simulation of the iRobot Create robot, with the RSS signals simulated based on fingerprinting data from a real indoor area with 6 dedicated access points as reference. The efficiency of each system is evaluated using cumulative distribution function for several access point combinations, noise and vanishing levels for a model trained with the base test parameters from the reference material, with all 6 access points (APs) activated, ldBm Gaussian noise, 10% masking level and using 10 time steps of data as history inputs. The results show that RNN systems can achieve mean localization accuracy between $0.44\mathrm{m}\pm 0.39\mathrm{m}$ for LSTM and $0.50\mathrm{m}\pm 0.38\mathrm{m}$ for GRU and the KNN proposal reaching $0.68\mathrm{m}\pm 0.73\mathrm{m}$, proving the capability of those systems to recover from a KRP event keeping similar results obtained without any event.