This is a step by step implementation of the method we used for human action recognition and ergonomic risk assessement in "Toward Ergonomic Risk Prediction via Segmentation of Indoor Object Manipulation Actions Using Spatiotemporal Convolutional Networks" paper that is accepted to the CASE conference 2019 and the IEEE Robotics and Automation Letters.
The UW-IOM dataset can be found here.
The TUM Kitchen dataset can be found here. We relabeled the dataset, and labels can be dounloaded in this repository under the folder "Labels_TUM".
This code has been tested on a workstation running Windows 10 operating system, equipped with a 3.7GHz 8 Core Intel Xeon W-2145 CPU, GPU ZOTAC GeForce GTX 1080 Ti, and 64 GB RAM.
-
TensorFlow, Keras (1.1.2+)
-
Tested on Python 3.6
If you creat a directory for this project and copy the code in the "Code" folder and the UW-IOM dataset in the "data" folder the rest of the required directories will be generated automatically.
.\Code
.\UW_IOM_Dataset
1- Preparing the data is described in "Preparing_the_data.ipynb"
2- The feature extraction phase is described in "VGG16.ipynb"
3- The Temporal Convolutional Network is described in "TCN_Main_GPU.ipynb"
- The TCN code was built on the code by Colin Lea presented in the Temporal Convolutional Networks for Action Segmentation and Detection.
Please cite the following article if you found the code and UW-IOM dataset useful:
B. Parsa, E. U. Samani, R. Hendrix, C. Devine, S. M. Singh, S. Devasia, and A. G. Banerjee. Toward Ergonomic Risk Prediction via Segmentation of Indoor Object Manipulation Actions Using Spatiotemporal Convolutional Networks. IEEE Robotics and Automation Letters, To appear. [Pre-print]