Data-driven phase extraction for anomaly detection of repetitive human movements

• Main Author: Jatesiktat, Prayook
• Other Authors: Ang Wei Tech
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2019
• Subjects: DRNTU::Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/83253; http://hdl.handle.net/10220/47998
• Institution: Nanyang Technological University

Human movements during a specific task usually consist of inconsistency and variations. They are caused by different strategies, the pace of movement, or even anthropometric structure of each subject. This dissertation aims to develop a norm modelling methodology that can model a generic repetitive movement in a data-driven way from a collection of the movement from healthy subjects without any manual data labelling or segmentation. Then, this model can be used to assess the abnormality of the movement during rehabilitation exercises of stroke patients. To make the method invariant to the variation in phase progression, a neural network is trained to extract phases from unlabeled time-series of kinematic features. This unsupervised learning method enhanced the generalization of the modelling process to be applied to a wide range of repetitive exercises without the need for handcrafting new features. The trained model is used to segment and time-warp the exercise data into standardized periods. This standardization allows the principal component analysis reconstruction to be applied for anomaly detection. The whole modelling workflow not only works on a gold-standard marker-based motion capture system, but it is also compatible with our motion capture system which integrates a Kinect sensor with wrist-worn inertial measurement units (IMU). The compactness and affordability of the system allow this method to be used in a home-based rehabilitation scenario where space and cost are the constraints. An evaluation on four different exercises shows good potential in separating anomalous movements from normal ones with an average area under the curve (AUC) of 0.9872 and the test on actual stroke survivors is also performed. In addition, a new real-time depth-based object tracking method is invented by training a neural network to learn a continuous signed distance field of a target object. Its flexibility allows the method to be applied to various rigid and articulated objects such as a human body.