Please use this identifier to cite or link to this item:
http://hdl.handle.net/11375/28065
Title: | ResCoCo: Residual ConvLSTM Network with Contrastive Learning for 3D Joint Angle Estimation |
Authors: | Lou, Xijian |
Advisor: | Zheng, Rong |
Department: | Computing and Software |
Publication Date: | 2022 |
Abstract: | 3D joint angle is an important indicator in human gait analysis, and musculoskeletal disease diagnosis and treatment. To accurately estimate 3D joint angle, a Deep Learning approach, Residual ConvLSTM network with Contrastive Learning (ResCoCo), is proposed in this study. A sequence shortening layer is introduced in ResCoCo to discard part of the output sequence estimated by bi-directional LSTM layers from incomplete context, and a residual block is used for network depth increase. Contrastive learning is employed in ResCoCo to ensure robust and efficient representation extraction. The model is validated on the WEVAL dataset for 3D knee joint angle estimation during walking. The experiment result shows that the sequence shortening layer and residual block benefit the 3D joint angle estimation accuracy, while contrastive learning increases the model resistance towards IMU-to-Segment (I2S) alignment and sensor placement variations. Furthermore, the sensor configuration for the model input is investigated. Using inertial data from two sensors as the model input is economical while effective, and leads to good model robustness towards I2S alignment and sensor placement variations, compared to using inertial data from a single sensor or six sensors as the model input. The model is also compared with a model-driven method. It is shown that ResCoCo not only provides accurate estimation accuracy along three rotation axes, but it is also free of calibration procedures, physical constraints, or predefined anatomical models, compared to model-driven approaches. |
URI: | http://hdl.handle.net/11375/28065 |
Appears in Collections: | Open Access Dissertations and Theses |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
Lou_Xijian_202210_MSc.pdf | 1.57 MB | Adobe PDF | View/Open |
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.