Connectionist methods for data analysis and modelling of human motion in sporting activities
One or more files will be made publicly available from 2020-06-30.
MetadataShow full metadata
This research concerns automation of qualitative analysis of human motion in sports, using a novel approach related to assessment and diagnostics, which is required to provide a general user with coaching experience in next generations of motion capture video games or sport coaching software. The research comprises a framework hereinafter referred to as augmented coaching systems (ACS) and its critical components. In contrast to formative assessing of knowledge of results, which is based on predefined objective criteria, a qualitative approach to assessing knowledge of performance is linked to the questions: (1) Can qualitative assessment be automated?; (2) If so, how can such assessment be communicated from a machine to a human?; and (3) Can qualitative assessment automation be similar to human implicit, multifaceted, empirical, evolving and subjective criteria? An investigative development approach was used for automating human motion assessment. The assessment of qualitative nature incorporated a mix of objectives – such as subjective, objective, and flexible pre-defined criteria similar to a domain expert or coach. The methods of analysis and machine learning techniques included: learning-by-example from expert’s data; integrative visualisation/replay functionality for qualitative analysis and machine learning modelling; modelling and analysis utilising relatively small and larger unbalanced motion data sets; modular implementation of common-sense descriptive rules mapped to diagnostic outputs; and sub-space modelling and temporal and spatial feature extraction techniques. The introduced ACS framework is generic and it includes a critical analysis applicable to more than one sport discipline. The ACS architecture is modular, extendible and its machine learning system supports global, coaching scenario specific, personalised, evolving, and life-long learning. Using captured motion data sets representing novices towards advanced skill levels in two case studies (golf and tennis), a series of experimental modelling systems integral to ACS were developed for testing and validation using empirical, subjective, and flexible criteria. The results achieved on small and on relatively large unbalanced data sets produced human- intelligible diagnostic outputs in a qualitative fashion. The machine learning diagnostic outputs were similar to those produced by visual assessment of a tennis coach (81% ... 99.9%) and to those produced by objective measures from an embedded motion capture system in a golf club, resulting in 89.5 ±2.6%. Flexible assessment criteria were demonstrated by comparing the two different assessments for tennis swing stances that were based on different subjective criteria operating on the same motion data set using the same assessment system. The ACS framework, and developed software components for the next generation of intelligent ACS using subjective and flexible criteria, is novel in the field. This thesis has demonstrated that qualitative assessment can be automated, that assessment diagnostics can be communicated from a machine to human and that coaching insights as implicit knowledge can be modelled using connectionist and evolving connectionist approaches.