Gilbert, Thomas Jamin; (2021) Constructing a reference standard for sports science and clinical movement sets using IMU-based motion capture technology. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text The Final Thesis.pdf - Accepted Version Download (7MB) | Preview

Abstract

Motion analysis has improved greatly over the years through the development of low-cost inertia sensors. Such sensors have shown promising accuracy for both sport and medical applications, facilitating the possibility of a new reference standard to be constructed. Current gold standards within motion capture, such as high-speed camera-based systems and image processing, are not suitable for many movement-sets within both sports science and clinical movement analysis due to restrictions introduced by the movement sets. These restrictions include cost, portability, local environment constraints (such as light level) and poor line of sight accessibility. This thesis focusses on developing a magnetometer-less IMU-based motion capturing system to detect and classify two challenging movement sets: Basic stances during a Shaolin Kung Fu dynamic form, and severity levels from the modified UPDRS (Unified Parkinson’s Disease Rating Scale) analysis tapping exercise. This project has contributed three datasets. The Shaolin Kung Fu dataset is comprised of 5 dynamic movements repeated over 350 times by 8 experienced practitioners. The dataset was labelled by a professional Shaolin Kung Fu master. Two modified UPDRS datasets were constructed, one for each of the two locations measured. The modified UPDRS datasets comprised of 5 severity levels each with 100 self-emulated movement samples. The modified UPDRS dataset was labelled by a researcher in neuropsychological assessment. The errors associated with IMU systems has been reduced significantly through a combination of a Complementary filter and applying the constraints imposed by the range of movements available in human joints. Novel features have been extracted from each dataset. A piecewise feature set based on a moving window approach has been applied to the Shaolin Kung Fu dataset. While a combination of standard statistical features and a Durbin Watson analysis has been extracted from the modified UPDRS measurements. The project has also contributed a comparison of 24 models has been done on all 3 datasets and the optimal model for each dataset has been determined. The resulting models were commensurate with current gold standards. The Shaolin Kung Fu dataset was classified with the computational costly fine decision tree algorithm using 400 splits, resulting in: an accuracy of 98.9%, a precision of 96.9%, a recall value of 99.1%, and a F1-score of 98.0%. A novel approach of using sequential forward feature analysis was used to determine the minimum number of IMU devices required as well as the optimal number of IMU devices. The modified UPDRS datasets were then classified using a support vector machine algorithm requiring various kernels to achieve their highest accuracies. The measurements were repeated with a sensor located on the wrist and finger, with the wrist requiring a linear kernel and the finger a quadratic kernel. Both locations achieved an accuracy, precision, recall, and F1-score of 99.2%. Additionally, the project contributed an evaluation to the effect sensor location has on the proposed models. It was concluded that the IMU-based system has the potential to construct a reference standard both in sports science and clinical movement analysis. Data protection security and communication speeds were limitations in the system constructed due to the measured data being transferred from the devices via Bluetooth Low Energy communication. These limitations were considered and evaluated in the future works of this project.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item