Application of Machine Learning for Angle of Arrival Estimation in Indoor Environment

Abstract

Indoor localization plays a significant role in enabling Location Based Services (LBS). The global indoor LBS market has seen tremendous growth in recent years. Indoor LBS refers to solutions to track the location of a target in an indoor environment. Global Navigation Satellite Systems (GNSS) can be used for outdoor localization with high accuracy. However, in indoor environments, GNSS signals are weak and inaccessible. Due to the limitations of GNSS, the problem of indoor localization has gained a great deal of attention. In real life, the indoor environment is complex where radio signals are reflected by obstacles. This phenomenon leads to multipath signal reflections. In addition, a Line of Sight (LoS) between the transmitter and receiver cannot always be guaranteed, which is required for accurate indoor localization. Therefore, Non-Line of Sight (NLoS) signals are often used for localization. Both multipath phenomenon and absence of a LoS signal severely degrade the localization accuracy, hence it makes the indoor localization very challenging. Wi-Fi-based localization is one of the key technologies for enabling LBS. Some of the advantages of Wi-Fi-based localization include rapid deployment and cost efficiency, as most indoor environments such as homes and offices have Wi-Fi networks installed. Many studies have been performed, in both industry and academia, on various localization techniques such as Angle of Arrival (AoA), Time of Arrival (ToA), or Time Difference of Arrival (TDoA). However, an accurate and cost-effective solution for indoor localization using Wi-Fi still does not exist due to the aforementioned challenges. Two main contributions were achieved during the entire MSc studies. First, an indoor localization system has been proposed and simulated where AoA measurements are incorporated into a Particle Filter (PF) for location estimation. In addition, acceleration data obtained from Inertial Measurement Unit (IMU) sensors is fused with AoA measurements using the PF. The effects of multipath and NLoS signals are reduced by exploiting the variability of the multipath in the PF in order to improve the indoor localization accuracy. Second, to further improve the indoor localization accuracy and reduce computational complexity, a Machine Learning (ML) framework for AoA estimation is adopted. The Convolutional Neural Network (CNN) model including a dedicated preprocessing step is proposed to estimate the AoA in a multipath environment. The deep learning model is data driven and uses the preprocessed data in the form of eigenvectors matrix to estimate the AoA in a multipath environment. The performance of the proposed approach is validated by simulations. In terms of accuracy and computational complexity, the proposed CNN model achieves a significantly better performance than the traditional MUltiple SIgnal Classification (MUSIC) algorithm.

Date of Award	Dec 2021
Original language	American English