We augment human pose estimation (openpifpaf library) by support for multi-camera and multi-person tracking and a long short-term memory (LSTM) neural network to predict two classes: “Fall” or “No Fall”. From the poses, we extract five temporal and spatial features which are processed by an LSTM classifier.
pip install -r requirements.txtpython3 fall_detector.py| Argument | Description | Default |
|---|---|---|
| num_cams | Number of Cameras/Videos to process | 1 |
| video | Path to the video file (None to capture live video from camera(s)) For single video fall detection(--num_cams=1), save your videos as abc.xyz and set --video=abc.xyz For 2 video fall detection(--num_cams=2), save your videos as abc1.xyz & abc2.xyz & set --video=abc.xyz | None |
| save_output | Save the result in a video file. Output videos are saved in the same directory as input videos with "out" appended at the start of the title | False |
| disable_cuda | To process frames on CPU by disabling CUDA support on GPU | False |
We used the UP-Fall Detection to train the LSTM model. You can use this Colab notebook to download the download the dataset and compile the files into videos.
Please cite the following paper in your publications if our work has helped your research:
Multi-camera, multi-person, and real-time fall detection using long short term memory
@inproceedings{Taufeeque2021MulticameraMA,
author = {Mohammad Taufeeque and Samad Koita and Nicolai Spicher and Thomas M. Deserno},
title = {{Multi-camera, multi-person, and real-time fall detection using long short term memory}},
volume = {11601},
booktitle = {Medical Imaging 2021: Imaging Informatics for Healthcare, Research, and Applications},
organization = {International Society for Optics and Photonics},
publisher = {SPIE},
pages = {35 -- 42},
year = {2021},
doi = {10.1117/12.2580700},
URL = {https://doi.org/10.1117/12.2580700}
}