Novas estratégias para o projeto de sistemas de posicionamento 3D em ambientes fechados baseados em aprendizado de máquina

Abstract

Indoor positioning systems (IPS) have attracted much attention in recent years, and this is motivated mainly by a large number of potential applications. However, it remains challenging to maximize the precision of this type of system, especially for three-dimensional (3D) estimates. In this research, this problem is discussed in a broad way. In addition, three solutions based on Bayesian inference are proposed. Among these solutions, we highlight the IPS-MAS system, which was developed from a multiagent system composed of a Bayesian network and a deep neural network. Additionally, this proposed system was designed to combine the multilateration and fingerprint methods in order to reduce the acquisition region of the received signal strength vectors. Additionally, the relationship between the quality of the received signal and the noise level, which is influenced by the increase in the number of access points and the number of people moving within the environment, is considered by the system. The proposed systems presented better performance when compared to the others, resulting in mean positioning errors of 0.90 m, 1.80 m, 1.82 m, for the IPS-MAS, k means-NB and k means-NB, respectively (scenario where the combination between the multilateration and fingerprint method was considered only for the IPS-MAS algorithm) and 0.90 m, 1.12 m, 1.19 m for the algorithms IPS-MAS, k means-NB and kNN-Bayes, respectively (scenario where the combination between the multilateration and fingerprint method was considered for the three solutions).