3D human action recognition and motion analysis using selective representations

Leightley, Daniel (2015) 3D human action recognition and motion analysis using selective representations. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (5MB) | Preview

Abstract

With the advent of marker-based motion capture, attempts have been made to recognise and quantify attributes of “type”, “content” and “behaviour” from the motion data. Current work exists to obtain quick and easy identification of human motion for use in multiple settings, such as healthcare and gaming by using activity monitors, wearable technology and low-cost accelerometers. Yet, analysing human motion and generating representative features to enable recognition and analysis in an efficient and comprehensive manner has proved elusive thus far. This thesis proposes practical solutions that are based on insights from clinicians, and learning attributes from motion capture data itself. This culminates in an application framework that learns the type, content and behaviour of human motion for recognition, quantitative clinical analysis and outcome measures. While marker-based motion capture has many uses, it also has major limitations that are explored in this thesis, not least in terms of hardware costs and practical utilisation. These drawbacks have led to the creation of depth sensors capable of providing robust, accurate and low-cost solution to detecting and tracking anatomical landmarks on the human body, without physical markers. This advancement has led researchers to develop low-cost solutions to important healthcare tasks, such as human motion analysis as a clinical aid in prevention care. In this thesis a variety of obstacles in handling markerless motion capture are identified and overcome by employing parameterisation of Axis- Angles, applying Euler Angles transformations to Exponential Maps, and appropriate distance measures between postures. While developing an efficient, usable and deployable application framework for clinicians, this thesis introduces techniques to recognise, analyse and quantify human motion in the context of identifying age-related change and mobility. The central theme of this thesis is the creation of discriminative representations of the human body using novel encoding and extraction approaches usable for both marker-based and marker-less motion capture data. The encoding of the human pose is modelled based on the spatial-temporal characteristics to generate a compact, efficient parameterisation. This combination allows for the detection of multiple known and unknown motions in real-time. However, in the context of benchmarking a major drawback exists, the lack of a clinically valid and relevant dataset to enable benchmarking. Without a dataset of this type, it is difficult to validated algorithms aimed at healthcare application. To this end, this thesis introduces a dataset that will enable the computer science community to benchmark healthcare-related algorithms.