Developing location systems that use mobile device sensors has been a topic of interest to industry and the academy. In this paper, we describe an experiment that was performed for evaluating the feasibility to create a mobile indoor localization model at room level based on data from participatory sensing. In order to achieve it, seven subjects that had a smartphone with magnetometer collected magnetic field information in a building composed by five rooms with different dimensions. The information collected was used to train three machine learning algorithms: k Nearest Neighbors (kNN), decision trees (J48) and Naïve Bayes. The performance of the algorithms was measured through the accuracy and the kappa statistics. Our results show that it is possible to create a mobile indoor localization model at room level using data from participatory sensing. The model with the highest performance was obtained with the kNN algorithm with a value of k = 3, since this offer an accuracy of 97.12% with a concordance level (kappa) of 0.9639 to estimate the localization of individual inside a room at the indoor environment. While the model with the lowest performance was Naïve Bayes, it offers an accuracy of 50.79% with a concordance level of 0.3834.
i think that deep learning could be more interesting.
You are right, we use the most common algorithms machine learning algorithms reported in literature on indoor location systems. This only to prove our hypothesis of using crowdsourcing as fingerprinting training phase. We also believe that deep learning could be interesting approach to develop an indoor localization model. As future researchwork, we will consider to use deep learning models such as CNNs to develop indoor localization systems.
Regards,
Juan Pablo
this because the due author only is using x,y,z values of magnetic field
In the paper we comment the we use of the intensity of the magnetic field in its components x, y e z. However, a stage of extracting signal features of the magnetic field signal was not performed.
Perhaps it might be interesting to analyze magnetic field signatures considering the crowdsourcing approach. I invite you to read an article where I worked with a colleague in the extraction of magnetic field signal features (https://www.mdpi.com/1424-8220/14/6/11001/htm).
regards,
In this article we assume that fingerprint has 100% accuracy.
So our results are compared against the accuracy of fingerprinting.
You can see that this accuracy is very close to fingerprinting, even though in our proposal not all exhaustive samples of the interior environment were taken.
