Fall Detection using Pose Estimation

Introduction

Fall Detection model based on OpenPifPaf

PyPI Library: https://pypi.org/project/openpifpaf/

The detection can run on both GPU and CPU, on multiple videos, RTSP streams, and webcams/USB cameras. Unlike most open-source fall detection models that work on large single subjects, this improved model integrates a person tracker that can detect falls in scenes with more than one person.

Demo Videos

Video credits: 50 Ways to Fall (Link), ran on a single NVIDIA Quadro P1000

Test Results

UR Fall Detection Dataset (Link), tested on two NVIDIA Quadro GV100s.

• Precision: 100%
• Recall: 83.33%
• F1 Score: 90.91%

Note: Due to lack of available datasets, false positives and true negatives were not tested.

Environment

• Ubuntu 18.04 x86_64
• Python 3.7.6
• Anaconda 3
• CUDA 10.2

Usage

Setup Conda Environment

$ conda create --name falldetection_openpifpaf python=3.7.6
$ conda activate falldetection_openpifpaf

Clone Repository

$ git clone https://github.com/cwlroda/falldetection_openpifpaf.git

Download OpenPifPaf 0.11.9 (PyPI)

$ pip3 install openpifpaf

Copy Source Files

$ cd {home_dir}/anaconda3/lib/python3.7/site-packages/openpifpaf
Replace ALL files in that folder with the files in falldetection_openpifpaf

Install Dependencies

$ pip3 install -r requirements.txt

Execution

For videos/RTSP streams, navigate to config/config.xml to edit the video/RTSP stream path, then run:

$ python3 -m openpifpaf.video --show
$ (use --help to see the full list of command line arguments)

For webcams/USB cameras, run:

$ python3 -m openpifpaf.video --source {CAMERA_ID} --show
$ (use --help to see the full list of command line arguments)

Citations

PifPaf: Composite Fields for Human Pose Estimation (Link)

@InProceedings{Kreiss_2019_CVPR,
    author = {Kreiss, Sven and Bertoni, Lorenzo and Alahi, Alexandre},
    title = {PifPaf: Composite Fields for Human Pose Estimation},
    booktitle = {Proceedings of the IEEE/CVF Conference on
                Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2019}
}

If you use the dataset above, please cite the following work: (Link)

Bogdan Kwolek, Michal Kepski,
Human fall detection on embedded platform using depth maps and wireless accelerometer,
Computer Methods and Programs in Biomedicine,
Volume 117,
Issue 3,
December 2014,
Pages 489-501,
ISSN 0169-2607