Using Gaussian Mixture Modelling to Analyse Dynamic Body Postures from Multiple Inertial Measurement Units

Abstract

This paper investigates the modelling of dynamic body motion and postures using multiple inertial measurement units. The ultimate goal of this work is to determine a way to model back posture during manual handling activities to prevent lower back pain. This is achieved using Gaussian Mixture Modelling to produce a model with twenty clusters. This model is then employed to predict the order in which the clusters occur during the movement. These clusters are then analysed using various methods to determine whether good or bad posture may be associated with these clusters. This is a two-fold problem and involves evaluating clusters which are statistically good or bad on their own or dynamic clusters which become good or bad after a certain sequence of clusters occur. Cluster means which indicate statistically significant postures are also discussed, which is vital for predicting bad posture use before a lift has even occurred. The key outcome of this work is the development of a decision tree which defines the posture observed, based on the order of the static and dynamic clusters associated with good and bad posture. This final decision tree has a precision of 75.3%, which is excellent considering the changes in movement between good and bad posture during a lift are very subtle.