Pose recognition system using bluetooth and machine learning techniques

Abstract

Due to the need to understand the physical condition of workers in hazardous environments, where the use of PPE is mandatory, this work proposes a pose recognition system capable of identifying states such as standing, sitting, lying down, and walking from wearable device signals. This study proposes a human pose recognition system based on Angle of Arrival (AoA) data obtained from Bluetooth Low Energy (BLE) 5.1 devices. Three neural network architectures were evaluated, namely Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and hybrid models, with the aim of identifying four physical states: standing, sitting, lying down, and walking. The data were collected in a real indoor environment using smart Personal Protective Equipment (PPE) equipped with BLE transmitters placed on a helmet, on the upper and lower regions of a shirt, and on a boot. The preprocessing pipeline included filtering, interpolation, feature extraction, and application of the Fast Fourier Transform (FFT). The models were evaluated in two scenarios: (i) known positions (used during training) and (ii) unknown positions (never seen before). The best overall performance in unknown positions was achieved by the Attention-based BiLSTM–CNN model, reaching an accuracy of A = 94.89% and a macro-average F1 score of F1 = 95.59%, demonstrating strong generalization capability. In the scenario with known positions, the best performance was obtained by the Deep CNN, with A = 97.27% and F1 = 97.28%, closely followed by the Attention-based BiLSTM–CNN, which also showed high classification accuracy across all posture classes. The analysis confirmed that the use of four wearable devices resulted in the best performance, and that sensor combinations significantly impacted accuracy. The study also makes available a public dataset to support reproducibility and benchmarking.