Simulation of machine learning inferences in real-time operating system to improve direction finding in an embedded environment

Bluetooth Low Energy (BLE) offers an inexpensive low-power technology for precise indoor positioning. However, multipath propagation and signal scattering deteriorate localization performances. Machine learning methods were previously demonstrated to significantly improve BLE direction finding estimates over classical signal processing methods in several studies. Performances as well as memory and computing requirements of on-device inference approaches were however not evaluated with respect to embedded hardware. We developed a computational framework to validate storage size, inference time and performances of different previously proposed machine learning and deep learning architectures within a ray-tracing indoor positioning dataset. We highlight that gradient-boosted decision tree methods outperformed deep learning approaches, met embedded hardware requirements, and generalized well when indoor environment changes were introduced. Additionally, we compared static multi-channel features with BLE protocol compliant single random channel features and found a minor single channel performance deterioration, but still significant performance benefit over signal processing methods. Our study therefore provides a validated framework to assess machine learning performances for indoor positioning in embedded hardware.