Machine Learning Model for the Prediction of Human Movement Biomechanics

[img]
Preview
ZAROUG_Abdelrahman-thesis_nosignature.pdf - Submitted Version (15MB) | Preview

Zaroug, Abdelrahman (2021) Machine Learning Model for the Prediction of Human Movement Biomechanics. PhD thesis, Victoria University.

Abstract

An increasingly useful application of machine learning (ML) is in predicting features of human actions. If it can be shown that algorithm inputs related to actual movement mechanics can predict a limb or limb segment’s future trajectory, a range of apparently intractable problems in movement science could be solved. The forecasting of lower limb trajectories can anticipate movement characteristics that may predict the risk of tripping, slipping or balance loss. Particularly in the design of human augmentation technology such as the exoskeleton, human movement prediction will improve the synchronisation between the user and the device greatly enhancing its efficacy. Long Short Term Memory (LSTM) neural neworks are a subset of ML algoithms that proven a wide success in modelling the human movement data. The aim of this thesis was to examine four LSTM neural nework architectures (Vanilla, Stacked, Bidirectional and Autoencoders) in predicting the future trajectories of lower limb kinematics, i.e. Angular Velocity (AV) and Linear Acceleration (LA). This work also aims to investigate whether linear statistical methods such as the Linear Regression (LR) is enough to predict the trajectories of lower limb kinematics. Kinematics data (LA and AV) of foot, shank and thigh were collected from 13 male and 3 female participants (28 ± 4 years old, 1.72 ± 0.07 m in height, 66 ± 10 kg in mass) who walked for 10 minutes at 4 different walking speeds on a 0% gradient treadmill. Walking -1 -1 speeds included preferred walking speed (PWS 4.34 ± 0.43 km.h ), imposed speed (5km.h , 15.4% ± 7.6% faster), slower speed (-20% PWS 3.59 ± 0.47 km.h-1) and faster speed (+20% PWS 5.26 ± 0.53 km.h-1). The sliding window technique was adopted for training and testing the LSTM models with total kinematics time-series data of 17,638 strides for all trials. The aim and findings of this work were carried out in 3 studies. Study 1 confirmed the possibility of predicting the future trajectories of human lower limb kinematics using LSTM autoencoders (ED-LSTM) and the LR during an imposed walking speed (5km.h-1). Both models achieved satisfactory predicted trajectories up to 0.06s. A prediction horizon of 0.06s can be used to compensate for delays in an exoskeleton’s feed-forward controller to better estimate the human motions and synchronise with intended movement trajectories. Study 2 (Chapter 4) indicated that the LR model is not suitable for the prediction of future lower limb kinematics at PWS. The LSTM perfromace results suggested that the ED-LSTM and the Stacked LSTM are more accurate to predict the future lower limb kinematics up to 0.1s at PWS and imposed walking speed (5km.h-1). The average duration for a gait cycle rages between 0.98-1.07s, and a prediction horizon of 0.1 accounts for about 10% of the gait cycle. Such a forecast may assist users in anticipating a low foot clearance to develop early countermeasures such as slowing down or stopping. Study 3 (Chapter 5) have shown that at +20% PWS the LSTM models’ performance obtained better predictions compared to all tested walking speed conditions (i.e. PWS, -20% PWS and 5km.h-1). While at -20% PWS, results indicated that at slower walking speeds all of the LSTM architectures obtained weaker predictions compared to all tested walking speeds (i.e. PWS, +20% PWS and 5km.h-1). In addition to the applications of a known future trajectories at the PWS mentioned in study 1 and 2, the prediction at fast and slow walking speeds familiarise the developed ML models with changes in human walking speed which are known to have large effects on lower limb kinematics. When intelligent ML methods are familiarised with the degree of kinematic changes due to speed variations, it could be used to improve human-machine interface in bionics design for various walking speeds The key finding of the three studies is that the ED-LSTM was found to be the most accurate -1 model to predict and adapt to the human motion kinematics at PWS, ±20% PWS and 5km.h up to 0.1s. The ability to predict future lower limb motions may have a wide range of applications including the design and control of bionics allowing better human-machine interface and mitigating the risk of tripping and balance loss.

Item type Thesis (PhD thesis)
URI https://vuir.vu.edu.au/id/eprint/42489
Subjects Current > FOR (2020) Classification > 4207 Sports science and exercise
Current > FOR (2020) Classification > 4602 Artificial intelligence
Current > Division/Research > Institute for Health and Sport
Keywords machine learning; Long Short Term Memory; LSTM; lower limb; kinematics; walking; Encoder-Decoder LSTM; ED-LSTM
Download/View statistics View download statistics for this item

Search Google Scholar

Repository staff login