Comparação entre técnicas de classificação aplicadas em sinais de wi-fi para localização de células nas instalações da UTFPR-CM

Abstract

Internal location systems and their applicability are on the rise, however, mitigating errors and interference are fundamental points for a good result in addition to the cost of applying it, either due to the acquisition of technology or the resources available to make the execution of the localization. In this work, four experiments based on machine learning are carried out to solve the problem of indoor localization within the UTFPR-CM. The main objective is to identify the best way to carry out the mapping of cells (rooms) and in addition to comparing the accuracies obtained by the classical K-NN and SVM classifiers, to then define the most efficient model for the internal location. The experiments were carried out by collecting 1, 5, 6 or more mapping points per cell, and later used in the K-NN and SVM classifiers to obtain the accuracy. To build predictive models, the GridSearch method is used through the Sklearn library. The most efficient method was the scenario in which points were collected every 1.5m, ranging from 14 to 36 points per cell, and using the SVM classifier with linear Kernel, C equal to 1 and Gamma equal to 0.002. Obtaining an accuracy of 82%. This result is promising and opens the way for further research in addition to significant improvements making the solution more efficient.