Indoor localization using wifi signals

Over the past years, there has been an increasing demand on indoor localization. This has led to diverse solutions: from triangulation algorithms, to the implementation of machine learning technologies. The main difficulty about indoor localization is the precision. In fact, outdoor localization coped the accuracy problem by using the GPS. However, as for indoor location goes, GPS cannot be used because of its low accuracy. Through different implementations, these new technologies and methods have been trying to cope with the problem of accuracy and location prediction in indoor localization. One of the solutions that has been developed, is the use Convolutional Neural Networks (CNNs) by analyzing the Wifi signals of the access points (APs) to predict a location (X,Y). Some datasets have been made public in order to help with the research, which has helped to do a fair comparison between different approaches. The advantages of these datasets is that they are big enough to train the models, and a fair accuracy can be achieved thanks to them. However, when it comes to private datasets, the accuracy of the prediction is limited by the size of them: a private dataset can have 1000 rows and the public ones can go up to nearly 20000 rows. The scope of the thesis was to build a model that could predict the localization in indoor areas, using small (private) datasets but with the the help of the public available datasets. The personal contribution to this work was selecting the appropiate method to deal with the problem of indoor localization. After choosing Convolutional Neural Network as the solution, the right needed to be taken to develop the model so the prediction of the localization could be accurate enough to be used. The results regarding the private dataset were considered as good enough, a prediction error of 3.91m inside a facility of 99.5 x 67m. A comparison can be made with the public dataset UJI-Lib. The facility's dimensions are: 308.4 m2 and 448 wifi Access Points (AP) as the facility has nearly 10 times more APs and is 20 times smaller compared to the private environment (6667 m2 and 58 APs), but the prediction error is only half (1.82m for UJI-Lib and 3.91m for Office-D). That is why the error prediction for private dataset has been considered good enough.