This repository contains the implementation of the Enhanced Spatio-Temporal Image Encoding (ESTIE), in order to perform Online Human Activity Recognition using 3D skeletons. This method is based on the use of the Spatio-Temporal Image Encoding (STIE) and the motion energy in order to encode a sequence of 3D skeletons into an image, while preserving both spatial and temporal dependencies.
Our paper can be found at:
Enhanced Spatio- Temporal Image Encoding for Online Human Activity Recognition
If you use or build on our work, please consider citing us:
@INPROCEEDINGS{10459847,
author={Mokhtari, Nassim and Fer, Vincent and Nédélec, Alexis and Gilles, Marlene and de Loor, Pierre},
booktitle={2023 International Conference on Machine Learning and Applications (ICMLA)},
title={Enhanced Spatio- Temporal Image Encoding for Online Human Activity Recognition},
year={2023},
volume={},
number={},
pages={884-889},
keywords={Training;Image coding;Three-dimensional displays;Image recognition;Time series analysis;Focusing;Streaming media;3D Skeleton Data;Spatio-temporal Image En-coding;Motion Energy;Online Action Recognition;Human Activity Recognition;Deep learning},
doi={10.1109/ICMLA58977.2023.00130}}
Before running our code, please unzip the archive data.zip provided in this repo. This archive contains skeleton data and sequence labels from the Online Action Detection dataset.
note: If you are using your own dataset, please consider adjusting the load_data_file() function.
You can start the encoding using the default parameters by running the STIE.py from the command line :
python ./ESTIE.py
Several parameters can be used to adapt the encoding according to your needs. You can find more details about these parameters using :
python ./ESTIE.py --help
The ESTIE image is the combination of the STIE and the motion energy proposed by Liu et al.
You can find the trained VGG16 (STIE and ESTIE versions) under the folder models