Research on Feature Extraction Method for Construction Posture Recognition Based on Wearable Sensors

Musculoskeletal disorders are common diseases with high incidence which are great impact on health in construction. They are often caused by excessive twisting or heavy load in bad posture, so it is very important to detect and identify these postures efficiently. Researchers often use inertial measurement units (IMU) to record the acceleration, Euler angle and other features from the IMU placed at some nodes to collect information of postures. However, lots of nodes and feature dimensions make the data memory very large, which will lead to the decrease of computing speed and the increase of storage cost. This study proposes a feature processing method to simplify the amount of data based on IMU, and use long short term memory (LSTM) neural network for recognition verification. The method combines the quaternions of the plane normal vectors with the coordinates of an original node. Using this method, we have achieved recognition accuracy similar to the original data while reducing the number of features by nearly half. The research results are beneficial to optimize the automatic recognition process of construction posture, which have a wide application prospect.