Deep learning workbench for biomechanics
http://digitalathlete.org
Machine learning driven by wearable sensors can replace laboratory instrumentation and bring biomechanical accuracy and validity to the sports field. By taking a new approach and training convolutional neural networks (CNNs) with archived biomechanics data, we are creating deep learning models to facilitate accurate multidimensional on-field predictive analytics. Using test sets from multi data-captures which include ground truth force plate and/or source modelling, we see strong correspondence between measured versus predicted ground reaction forces and moments, and knee joint moments, from markers (GRF/M r>0.97, KJM r>0.88, both 8 markers, 33% stance, sidestepping). Currently, we are investigating novel methods to build models for predicting GRF/M from wearable sensors (r>0.88, 4 accelerometers, 100% stance, sidestepping). William Johnson BEngHons, MExerSc, CPEng, CSCS [email protected] | |
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. |
Study four: Multidimensional knee joint moments from wearable sensor accelerations via deep learning
Keywords | Biomechanics · Wearable sensors · Synthetic accelerations · Motion capture · Sports analytics |
---|---|
Paper | In preparation |
WCB 2018 conference |
Multidimensional ground reaction forces and moments from wearable sensor accelerations via deep learning Abstract http://bit.ly/2GzYnHD Presentation with commentary http://bit.ly/2tCKHTo |
Keywords | Biomechanics · Wearable sensors · Computer vision · Motion capture · Sports analytics |
---|---|
arXiv paper | https://arxiv.org/abs/1809.08016 |
Presentation | http://bit.ly/2HS7HCv |
Animation | Training set marker trajectories versus corresponding knee joint moments visualization (supplementary figure) http://bit.ly/2yTaX1f |
MATLAB figures | https://github.com/johnsonwr/digitalathlete/tree/master/study3/figures |
Caffe models | https://github.com/johnsonwr/digitalathlete/tree/master/study3/models (1.3GB)
|
Prototxt | https://github.com/johnsonwr/digitalathlete/tree/master/study3/prototxt |
Study two: Predicting athlete ground reaction forces and moments from spatio-temporal driven CNN models
Keywords | Biomechanics · Supervised learning · Image motion analysis · Computer simulation |
---|---|
IEEE TBME paper | https://ieeexplore.ieee.org/document/8408711 |
Animation | Training set marker trajectories versus corresponding ground reaction forces and moments visualization (supplementary figure) http://bit.ly/2Is3PJx |
MATLAB figures | https://github.com/johnsonwr/digitalathlete/tree/master/study2/figures |
Caffe models | https://github.com/johnsonwr/digitalathlete/tree/master/study2/models (1.3GB)
|
Prototxt | https://github.com/johnsonwr/digitalathlete/tree/master/study2/prototxt |
CaffeNet reference | https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_caffenet |
ISBS 2017 conference | Prediction of ground reaction forces and moments via supervised learning is independent of participant sex, height and mass Abstract https://commons.nmu.edu/cgi/viewcontent.cgi?&article=1034&context=isbs Presentation http://bit.ly/2MvqW8c |
Keywords | Action recognition · Wearable sensors · Computer simulation |
---|---|
MBEC paper | https://link.springer.com/article/10.1007/s11517-018-1802-7 |
Presentation with commentary | http://bit.ly/2kcgXrw |
MATLAB figures | https://github.com/johnsonwr/digitalathlete/tree/master/study1/figures |
R models | https://github.com/johnsonwr/digitalathlete/tree/master/study1/models (1.9GB)
|
R SPLS reference | https://cran.r-project.org/web/packages/spls/index.html |