Bluetooth Direction Finding using Recurrent Neural Network

• Published in: 2021 International Conference on Indoor Positioning and Indoor Navigation (IPIN) pp. 1 - 7
• Main Authors: Babakhani, Pedram, Merk, Timon, Mahlig, Matthias, Sarris, Ioannis, Kalogiros, Dimitris, Karlsson, Peter
• Format: Conference Proceeding
• Language: English
• Published: IEEE 29-11-2021
• Subjects: Angle of Arrival; Bluetooth; Deep Learning; Direction Finding; Embedded systems; Estimation; Gated Recurrent Unit; Indoor Positioning; Recurrent Neural Network; Recurrent neural networks; Scattering; Signal processing algorithms; Time series analysis
• ISSN: 2471-917X
• DOI: 10.1109/IPIN51156.2021.9662611

Multipath propagation in an indoor environment has a detrimental impact on the performance of Angle of Arrival (AoA) estimation methods due to the existence of obstacles introducing reflections and scattering. This paper proposes a new architecture for AoA estimation, utilizing a robust and fast signal processing algorithm and a small Recurrent Neural Network (RNN) to improve performance by considering AoA estimation as a time series problem. The proposed method uses the Spatial Power Spectrum (SPS) of the well-established Propagator Direct Data Acquisition (PDDA) algorithm as an input feature for a Gated Recurrent Unit (GRU), which enhances the accuracy of PDDA by learning dependencies of spatial power spectrum features through previous time steps. Experimental results on a simulated rectangular indoor environment, with four different obstacle sets, show significant performance benefits (PDDA MAE =7.0° vs GRU MAE=3.7°) of GRU. Furthermore, the proposed method outperforms PDDA in a real indoor environment measurement (PDDA MAE = 12.2° vs GRU MAE = 7.1°). Additionally, the proposed method is sufficiently small in size (830 kB) to be employed on a wide range of embedded systems.