Upper Body Posture Recognition Using Inertial Sensors and Recurrent Neural Networks

Inadequate sitting posture can cause imbalanced loading on the spine and result in abnormal spinal pressure, which serves as main risk factor contributing to irreversible chronic deformity. Therefore, recognition is important for understanding people’s behaviors correcting inadequate postures. Recently, wearable devices embedded with microelectromechanical systems (MEMs) sensors, such inertial measurement units (IMUs), have received increased attention human activity recognition. In this study, a device IMUs machine learning algorithm were developed classify seven static postures: upright, slump, lean, right left bending, twisting. Four 9-axis uniformly distributed between thoracic lumbar regions (T1-L5) aligned sagittal plane acquire kinematic information about subjects’ backs during static-dynamic alternating motions. Time-domain features served inputs signal-based classification model that was using long short-term memory-based recurrent neural network (LSTM-RNN) architecture, model’s performance used evaluate relevance sensor signals Overall results from evaluation tests indicate IMU-based LSTM-RNN structural scheme appropriate