Predicting lower limb kinematics and kinetics from internal measurement units using deep learning
File(s)
Author(s)
Bicer, Metin
Type
Thesis
Abstract
Motion capture using marker and ground reaction force (GRF) requires expensive and time-consuming experimental setups, trained personnel and dedicated laboratory space. In response, inertial measurement units (IMUs) have emerged as a cheaper alternative, usable in remote locations. However, IMUs present challenges in placement (calibration) and data processing (drift). To streamline IMU data collection and eliminate the need for multiple sensors, this thesis develops and validates two deep neural network (NN) models for computing lower limb joint kinematics and kinetics during walking using data from a single sacrum-worn IMU. These NNs were developed using IMU data simulated from a publicly available marker-based dataset as inputs, with joint angles and moments calculated by a musculoskeletal model as labels. Given inadequate dataset sizes for deep learning applications, generative adversarial networks (GANs) were employed to create synthetic marker and GRF datasets. Comparisons between synthetic and experimental datasets, using statistical tests, ensured realistic data generation for NN training. Subsequently, the developed NNs were initially tested on a dataset encompassing marker, GRF, and real IMU data collected for this study, showing root mean squared errors (RMSEs) of 4.48° for joint angles and 0.21 Nm/kg for joint moments. These models were then fine-tuned, reducing the RMSE to 2.57° for joint angles and 0.19 Nm/kg for joint moments on another set of real IMU data collected in this study. The significance of these NNs lies in their reliance on training with simulated datasets and testing with real datasets from different institutions with varying protocols, suggesting that the developed NNs can be deemed as inter-laboratory tools. Conversely, the fine-tuned models were refined and evaluated using the dataset collected in this study, thereby rendering them laboratory-specific in their applicability. The thesis developed and validated NNs to predict lower limb joint angles and moments using data from a single sacrum-worn IMU.
Version
Open Access
Date Issued
2024-04-03
Date Awarded
01/11/2024
Advisor
Phillips, Andrew
Modenese, Luca
Sponsor
Turkey. Millî Eğitim Bakanlığı
Publisher Department
Civil and Environmental Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)