TeleRehab

This project develops an automated system for assessing physical rehabilitation exercises using RGB data.

Table of Contents

1. About The Project
   • Built With
2. Getting Started
   • Prerequisites
   • Installation
3. Usage
4. Roadmap
5. Contributing
6. License
7. Contact
8. Acknowledgements

About The Project

This project develops an automated system for assessing physical rehabilitation exercises using RGB data. The end-to-end deep learning approaches (i.e. C3D, 3D-ResNet) and the feature extraction based approaches (i.e. LR, MLP, KNN) were implemented. KiMoRe dataset is used for training and testing the models.

Getting Started

Prerequisites

• python 3.7

Installation

1. Clone the repo. Note the colab branch is the main branch.

   git clone -b colab https://github.com/claraguoguo/TeleRehab.git

2. Install common ML libraries i.e. scipy, pandas, numpy, matplotlib, seaborn, ffmpeg...

Config

• config.cfg: config file for running code locally
• colab_config.cfg: config file for running code on Google Golab

Ex1

• n_repetition = 5

Ex2

• n_repetition = 10

Ex3

• n_repetition = 15

Ex4

• n_repetition = 10

Ex5

• n_repetition = 10

Run models

• train.py: train and test deep learning models (cnn, resnet, c3d)
• train_LSTM.py: train and test LSTM model
• train_NN.py: train and test MLP and linear regression models
• train_NN_sklearn.py: train and test sklearn models
• train_weighted_loss.py: train and test deep learning models (cnn, resnet, c3d) with weighted loss implementation

sample usage:

python train.py --config config.cfg --model_name c3d

Related Repos

• Code to extract features from skeletal data can be found in TeleRehab_Utilities

• (DEPRECATED) Code used to extract skeletal joints with openpose-COCO model can be found at demo_video_KIMORE.py. This code is developed on top of pytorch-openpose.

Useful Resources

• openpose
• pytorch-openpose - pytorch implementation of openpose including Body and Hand Pose Estimation (this version works on Apple M1 chip)

Contact

Clara Guo - [email protected]