The application of wearable cameras, accelerometers and motion capture for the analysis of complex gait

Gait analysis is an increasingly useful biomedical tool, with applications including diagnosis, treatment planning, and disease monitoring. Most gait analysis is conducted in a clinical setting using motion capture, however its greatest impact and most pertinent uses are outside the laboratory in a free-living environment. To enable the transition from clinic to real-world, wearable technologies have been developed to provide affordable, accessible, remote and unobtrusive assessment. The most popular by far are accelerometry-based devices, however their reliability as free-living tools is questionable due to: the limitations of complex gait definitions; lack of robust experimental validation; an absence of ground truth in free-living environments; and their sensitivity to a range of parameters relating to the devices themselves and how they are implemented. It is the aim of this thesis to confront each of these issues, enabling a better understanding and measurement of complex gait. To achieve this, novel spatio-temporal gait definitions and visualisations are proposed to accommodate all forms of gait. It is also hypothesised that computer vision methods applied to video from a wearable camera can extract a range of gait features in all environments, providing less sensitivity in accuracy to complex movement than current accelerometry techniques. Experimentally, a study was devised and conducted using novel protocols to better imitate free-living gait in a motion capture laboratory; as well as utilising a wearable camera and range of accelerometers. It is shown that simplistic straight protocols are not representative of complex gait, and that all wearables suffer substantial loss of accuracy when submitted to such protocols. Furthermore, given robust definitions and a simple set up, complex gait can be accurately measured by motion capture; however, gait features exist on a spectrum that is highly sensitive to definition. Both points elucidate the need for unification in the field, and the questionability of the intra-study comparability and validity of gait measuring methods. Extracting gait features from wearable camera video yielded varying results. For example, turns were detected and measured with greater than 90% accuracy, step detection precision and stride length correlation with motion capture ground truth ranged from 0.78% to 0.98% and 0.55 to 0.80 respectively depending on protocol, whereas step width could not be accurately measured at all. However, these fared comparably to all research-grade and consumer accelerometers; especially when compared step-by-step. Of these, the foot-mounted devices were by far the most reliable, whereas smartwatches and smartphones showed expected high accuracy during standard validation protocols, but this in no way transposed to imitation free-living movements.