Position Dependent Anchor Selection for Scalable Hybrid RF-Acoustic Indoor Positioning

Accurate 3D indoor positioning is a sought-after feature in many applications. To achieve this, we perform a comparative study of proposed anchor selection methods applied to a hybrid RF-acoustic indoor positioning system. The positioning accuracy is evaluated for two 3D positioning mechanisms. In the first positioning approach, the proposed anchor selection methods are tested in a classical system which uses the selected noisy range estimates which are converted into a 3D position using a least squares (LS) estimator. This mechanism requires the exact knowledge of the anchor locations, which introduces undesired anchor position calibration, identification and commissioning steps when deploying such a system. In order to overcome this problem, a second positioning mechanism, by means of circular convolutional neural network (CCNN), is proposed. This model does not require the knowledge of the anchor locations, as it is implicitly learned during training. The proposed anchor selection methods are evaluated both for the classical LS estimator and the circular convolutional neural network (CCNN) model in both a shoebox and L-shaped room. The results indicate a significant improvement in accuracy when using anchor selection and/or the machine learning (ML) based models, primarily in rooms where Non-Line-of-Sight (NLoS) situations are encountered. For the classical system, it is more effective to use fewer, but more pertinent anchors, whereas using all anchors is optimal for ML methods. Nevertheless, ML methods outperform traditional positioning algorithms, and anchor selection addresses scalability.