Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Non-invasive estimation of ground and joint kinetics through deep learning

Deep learning models driven by wearable sensor accelerometers can replace captive laboratory instrumentation to facilitate biomechanical accuracy and validity anywhere

By employing a new deep learning workbench for spatio-temporal data, we train convolutional neural networks (CNN) with archive biomechanics data to predict accurate multidimensional on-field analytics for complex sports movements. Using test sets from multi data-captures which include ground truth force plate or source modeling, we see strong correspondence between measured versus predicted ground reaction forces and moments, and knee joint moments. Driven by eight markers, study two GRF/M mean r>0.97, study three KJM mean r>0.88, and from five wearable sensor accelerometers, study four GRF mean r>0.87. The overarching hypothesis, whether it is possible to build deep learning models which can mimic the physics behind human movement, specifically to replace force plate derived kinetic output, is supported.

William R Johnson PhD CPEng CSCS (he/him/his) | July 2022 |
Caution, model files are large, you may not wish to pull the complete repository. GitHub limits file sizes to 100MB, files larger than this have been broken up using split. Instructions to reconstitute files are given inline.

Post-doc: Conference presentations, panels, & session chairs

Neural Networks Session Chair
NSCA Baseball and Sport Science SIG Performance Technology Roundtable
Data Science and Sports Biomechanics Panel

Post-doc: A comparison of three neural network approaches for estimating joint angles and moments from inertial measurement units

KeywordsMachine learning · Wearable sensors · Joint kinematics · Joint kinetics
Sensors [12]

Doctoral thesis: Non-invasive estimation of ground and joint kinetics through deep learning

KeywordsBiomechanics · Data science · Deep learning · Big data · Sports analytics · Computer vision · Motion capture · Wearable sensors
The University
of Western Australia

Study four: Multidimensional ground reaction forces and moments from wearable sensor accelerations via deep learning

KeywordsBiomechanics · Wearable sensors · Simulated accelerations · Workload exposure · Sports analytics · Deep learning
arXiv [11b]
Deep learning workbench for biomechanics Each study utilized an incremental sequence of data preparation and modeling strategies, which by study four had evolved into the "deep learning workbench for biomechanics." Although the individual data science components had previously existed in the literature, the approach was novel and unique in this configuration and application to sports biomechanics.

Multidimensional ground reaction forces predicted from a single sacrum-mounted accelerometer via deep learning
Abstract [10]
EMS HDR Conference 2018 Poster (Conference Award)
WCB-2018 Abstract (Student Bursary Award) [6]
Presentation with commentary
Artificial intelligence, data analytics and sports biomechanics: A new era or a false dawn?
Abstract [7]
MATLAB figures
Caffe models (637MB)
cat grftrain_181112104456175_mcrnet.caffemodel_* > grftrain_181112104456175_mcrnet.caffemodel # reconstitute CaffeNet donor seed model
cat grftrain_181123170749181_mcrnet.caffemodel_* > grftrain_181123170749181_mcrnet.caffemodel # reconstitute CaffeNet model

Study three: On-field player workload exposure and knee injury risk monitoring via deep learning

KeywordsBiomechanics · Wearable sensors · Computer vision · Motion capture · Sports analytics
Journal of Biomechanics [8a]
arXiv [8b]
Predicting ground and joint kinetics from wearable sensor accelerations via deep learning
Abstract [9]
Presentation (panel)
AnimationTraining set marker trajectories versus corresponding knee joint moments visualization
(supplementary figure)
MATLAB figures
Caffe models (2.6GB)

cat grftrain_180612214018112_mcrnet.caffemodel_j01_* > grftrain_180612214018112_mcrnet.caffemodel_j01 # reconstitute CaffeNet donor seed model 01
cat grftrain_190215144249130_mcrnet.caffemodel_j01_* > grftrain_190215144249130_mcrnet.caffemodel_j01 # reconstitute CaffeNet model 01
Caffe prototxt

Study two: Predicting athlete ground reaction forces and moments from spatio-temporal driven CNN models

KeywordsBiomechanics · Supervised learning · Image motion analysis · Computer simulation · Pattern analysis
IEEE TBME Paper [5]
Cover & Feature
Animation Training set marker trajectories versus corresponding ground reaction forces and moments visualization
(supplementary figure)
UWA CSSE Conference 2017 Relative performance of Caffe deep learning models for spatio-temporal sport analytics
Prediction of ground reaction forces and moments via supervised learning is independent of participant sex, height and mass
Abstract (Student Travel Grant) [3]
MATLAB figures
Caffe models (1.3GB)

cat grftrain_190215204017060_mcrnet.caffemodel_j01_* > grftrain_190215204017060_mcrnet.caffemodel_j01 # reconstitute CaffeNet model 01
Caffe prototxt
CaffeNet reference

Study one: Predicting athlete ground reaction forces and moments from motion capture

KeywordsAction recognition · Wearable sensors · Computer simulation
MBEC [4]
UWA CSSE Conference 2016Presentation with commentary
The personalised 'Digital Athlete': An evolving vision for the capture, modelling and simulation, of on-field athletic performance
Abstract [2]
MATLAB figures
R models (1.9GB)

cat grftrain_171214215406095_R_predict_model_* > grftrain_171214215406095_R_predict_model.Rda # reconstitute R model
R SPLS reference

HDR prelim: Validity of a markerless motion capture system for sporting application

The University
of Western Australia

Master's assignment: Talent identification in elite rugby union - a theoretical update to an existing predictor algorithm

KeywordsAthlete selection · Predictor variables · Algorithm · Weightings
JASC [1]

Machine Learning resources

Andrew Ng Machine Learning course
Fei-Fei Li Stanford Convolutional Neural Networks for Visual Recognition
Ian Goodfellow Deep Learning
Bill Lubanovic Introducing Python
Sebastian Raschka Python Machine Learning
Aurélien Géron Hands-On Machine Learning with Scikit-Learn & TensorFlow
Sean Driver Perth Machine Learning Group

Biomechanics resources

The University
of Western Australia
Biomechanics courses
SSEH2250 Biomechanics in Sport and Exercise
SSEH3355 Biomechanical Principles
SSEH4633 Advanced Biomechanical Methods
David Winter Biomechanics and Motor Control of Human Movement
D. Gordon E. Robertson
Graham E. Caldwell
Joseph Hamill
Gary Kamen
Saunders N. Whittlesey
Research Methods in Biomechanics
Roger Bartlett Introduction to Sports Biomechanics
Joseph Hamill
Kathleen Knutzen
Timothy Derrick
Biomechanical Basis of Human Movement
Jim Richards Biomechanics in Clinic and Research: An Interactive Teaching and Learning Course
Carl Peyton
Adrian Burden
Biomechanical Evaluation of Movement in Sport and Exercise
Youlian Hong
Roger Bartlett
Routledge Handbook of Biomechanics and Human Movement Science
The Biomechanist The Week in Biomechanics
Biomch-L Forum, sponsored by the International Society of Biomechanics (ISB)
3-D Analysis of Human Movement Technical Group of the International Society of Biomechanics (ISB)
Awesome Biomechanics A curated repository of biomechanical resources

PhD supervisors

Jacqueline A. Alderson School of Human Sciences (Exercise and Sport Science), The University of Western Australia, Perth, Australia
Sports Performance Research Institute New Zealand (SPRINZ), Auckland University of Technology, Auckland, New Zealand
Ajmal Mian Department of Computer Science and Software Engineering, The University of Western Australia, Perth, Australia
Machine Intelligence Group
David G. Lloyd Menzies Health Institute Queensland, and the School of Allied Health Sciences, Griffith University, Gold Coast, Australia


This project was partially supported by the ARC Discovery Grant DP190102443 and an Australian Government Research Training Program Scholarship. NVIDIA Corporation is gratefully acknowledged for the GPU provision through its Hardware Grant Program, Eigenvector Research for the PLS_Toolbox licence, and C-Motion Inc. for the Visual3D licence. Portions of data included in this study were funded by NHMRC grant 400937.


PhD project supplementary material






No releases published


No packages published