Numerous studies have identified the importance of regular physical activity, limited sitting time and adequate sleep for the prevention and management of obesity and other lifestyle diseases. Researchers have tended to examine the health impact of these different physical behaviours in isolation; an emerging field in health research – Time Use epidemiology – has recently prompted researchers to measure these behaviours and evaluate their interactions across complete (24 hour) days. However, there are various problems associated with accurately measuring these behaviours across 24-hours using traditional measurement tools and protocols (such as insufficient wear-time compliance, and the inability to detect and differentiate postures). Advancements in technology and accelerometery have allowed researchers to utilise raw accelerometer data and develop predictive models using machine-leaning techniques. With growing interest in this field, this thesis aims to explore the utility of machine learning techniques for the accurate measurement of various human movement behaviours. To begin, a systematic scoping review was completed to summarise the current application of machine learning techniques for the accurate measurement of various physical activity behaviours. The review included studies that estimated components of physical behaviour by the combined application of supervised machine learning techniques and raw accelerometer data. Several key data points were extracted and synthesised from each study (e.g., the type of physical activity component classified, study environment, population description, device (i.e., accelerometer) specification, device placement position, ground truth method, machine learning classifier used, performance results). The review highlighted the increasing application of machine learning for predicting physical behaviours, with promising results, but their application in free-living settings was limited. To address the limited testing of machine learning with free-living accelerometer data, a validation experiment was conducted. This study investigated the performance of various dual-accelerometer placements under free-living conditions. Thirty participants (15 children, 15 adults) were equipped with three AX3 accelerometers—one to their thigh, one to their dominant wrist, and one to their lower back—alongside an automated clip camera (clipped to the lapel) that captured video of their free-living environment (criterion measure of physical activity). Participants completed several activities to represent the most common types of physical behaviours (e.g., sitting, lying, walking, running) at their private residence over a 2-hour period. A random forest machine learning classifier was then trained on features generated from raw accelerometer data. The results from the study show that the machine learning model developed using the thigh and back accelerometer performed the best and has potential to facilitate uninterrupted 24-hour monitoring of various physical behaviours This thesis revealed that accelerometery in combination with machine learning offers promise for measuring various free-living physical behaviours. However, it is essential for future studies to expand the scope of this work, by developing and validating a reliable measurement system that facilitates the continuous measurement of both intensity and type of physical behaviour in diverse free-living populations.