Real-time human balance and gait evaluation algorithms of a robotic balance assistant using optimization and learning-based methods

• Main Author: Foo, Ming Jeat
• Other Authors: Ang Wei Tech
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Mechanical engineering::Assistive technology
• Online Access: https://hdl.handle.net/10356/152982
• Institution: Nanyang Technological University

Aging greatly degrades the balance and locomotion ability of the elderly due to weakened physiology and pathological conditions. This demands balance and fall intervention technologies that help the affected population to regain mobility, independence, and quality of life. A gait rehabilitation and assistive robot, named Mobile Robotic Balance Assistant (MRBA), is developed to assist people with balance impairment to perform activities of daily living (ADLs) in home and community environments. This student's work is to develop the balance and gait evaluation algorithms for the robot. The process of human falls and loss of balance (LoB) is examined. Human instability is described as states of fall, in which the progression differs according to the severity of the instability and the recovery capability of the falling person. The Inverted Pendulum model is examined to find the theoretical indicators for the LoB. An experiment that induces real falls on human subjects is conducted to examine how the LoB evolves in reality. The results indicate that the ground reaction forces, the subjects' kinematics, and the relationship between the Center of Mass (CoM) and the Base of Support (BoS) are effective for instability detection. The findings establish the requirements of the robot, which is thus equipped with the necessary sensors. A foot tracking algorithm is built to locate the lower limbs in 3D space using modeling and optimization in real-time. It allows the BoS to be constructed from the feet position and computes the gait parameters. Comparing with the motion capture system, the tracking errors are less than 20 degrees and 35mm for rotational and translational errors, respectively, for 0.4m/s gait speed; the spatial and temporal gait parameters are computed with errors less than 35mm and 80ms, respectively. Various instability detection algorithms are developed and examined. Two main approaches are taken, namely thresholding on a single feature and predicting the user's future state. The first method compares a single feature with a well-defined threshold such that if the data exceeds the threshold, it is determined to be an LoB. The second method is based on anomaly detection with neural networks to learn the patterns of ADLs for it to predict future data. If the prediction differs significantly from the actual data, an anomaly, hence an LoB, is flagged. The learning-based programs have better performance due to their ability to consider multiple features in both spatial and temporal domains. The best models rely on forces, CoM and feet position data. They yield a low rate of false alarms and can detect 54 - 100% of ADLs with high risk and LoB within 1s. The detection latency is compared with the average timing users recover from a fall, showing that the system allows the users to practice their balance recovery reflex while ensuring their safety. While the work reveals promising results in the development of a balance and gait evaluation algorithm, more work needs to be accomplished for the functionality to mature sufficiently for the intended use. Elderly and patient's data have to be collected to improve the system. The timeliness and accuracy of the algorithms need to be enhanced to cater to more working scenarios. More intelligent solutions for user performance evaluation and human-robot compliance are required to increase its usability.