Validation of Full-Body Markerless 3D MoCap for Paralympic AI Classification

The classification of impairments in Paralympic sport plays a fundamental role in ensuring fair competition, ensuring those who succeed do so due to their athletic excellence rather than because they are less impaired than their competitors. However, the current system still relies heavily on an authority-based decision-making process to determine both athlete eligibility and sport class allocation. To transition to a more objective and standardised approach, a new project was launched, led by the School of Human Movement and Nutrition Sciences at the University of Queensland (UQ). The aim is to develop an AI-based model that will ultimately support and complement the evidence-based Paralympic classification system. To train an AI-based model, three-dimensional motion capture data is required. Markerless motion capture is particularly appealing for this project because large-scale data collection is required and, compared with marker-based motion capture, markerless is faster and less invasive. Theia3D is a deep learning algorithm-based software that uses inverse kinematic for pose estimation. As such, it could represent a promising solution for extracting three-dimensional coordinate data from two-dimensional recorded videos, that could be used during the training and testing phases of the AI algorithm. However, the Theia3D markerless motion capture system must first be validated against a gold-standard marker-based technology. This thesis presents a validation study that was conducted in UQ's biomechanics laboratory using a synchronized setup that combined the markerless motion capture system – in which Theia3D was used with a set of eight RGB video cameras – with the VICON marker-based system. Ten participants without disabilities were recorded while performing a standardised protocol of ten movements designed to make a full-body assessment of Theia3D’s performance, evaluating both upper body and lower body tasks with varying ranges of motion and speeds. The analysis focused on inter-system discrepancies in joint kinematics across all three anatomical planes, providing insight into the level of agreement and the potential applications of markerless motion capture in research and clinical contexts. Preliminary findings suggest that while Theia3D shows reasonable accuracy for some lower limb movements, particularly those predominantly occurring in the sagittal plane, its performance was considerably less comparable to the gold standard in the frontal and transverse planes for lower body movements and overall for upper limb tasks – especially for upper body movements involving large ranges of motion. Overall, the Theia3D markerless system did not demonstrate sufficient accuracy to serve as a ground truth provider for the AI classification, at least during the initial phase of data collection. Based on the outcomes of this validation study, VICON was selected as the motion capture system for the ongoing AI-based Paralympic classification project. Alternative markerless motion capture system with improved performance can be evaluated for future applications.