💧 DrownAI: Drowning Detection using Pose Estimation

This project implements drowning behavior recognition based on human pose estimation using OpenPose and deep learning classifiers.

---

Paper

This project is published at IEEE:

• Title: A Drowning Detection System Based on Human Pose Estimation and Deep Learning
• Conference: 2022 International Symposium on Computer, Consumer and Control (IS3C)
• DOI: 10.1109/IS3C53992.2022.9704884

---

Method Overview

This work proposes a real-time drowning detection system based on human skeletal motion analysis using AI techniques:

1. Pose Estimation:

   • Uses OpenPose with a lightweight Thin-MobileNet backbone to detect 13 body keypoints.
   • Camera is mounted underwater; pose estimation remains accurate despite splashes.

2. Feature Extraction:

   • Extracts joint coordinates and calculates geometric angles between limbs.
   • Converts the skeleton data into structured feature vectors.

3. Action Classification:

   • A shallow Recurrent Neural Network (RNN) is trained to classify actions as either “drowning” or “normal”.
   • The system is optimized for real-time operation with minimal GPU load.

---

Experimental Results

• Training data collected through simulated drowning scenarios in a swimming pool.
• Final classification accuracy: 89.4%
• Evaluation metrics:

Class	Precision	Recall	F1-Score
Drowning	0.98	0.67	0.79
Normal	0.96	1.00	0.98
Macro Avg	0.97	0.83	0.89

In most realistic simulated drowning situations, the system could correctly identify the drowning behavior with high precision.

---

Project Structure

• src/ - Main source code
• model/ - Pretrained classifier (.pickle)
• data_test/ - Example video for testing
• output/ - Output results and visualizations

Additional data, pretrained models and demo videos are hosted on Google Drive

---

Installation

1. Environment Setup

conda create -n drown python=3.6.12
conda activate drown
pip install -r requirements.txt

2. Install OpenPose Dependencies

cd src/depends/tf-pose-estimation
pip install -r requirements.txt
bash models/graph/cmu/download.sh
cd tf_pose/pafprocess
swig -python -c++ pafprocess.i && python3 setup.py build_ext --inplace
cd ../../../
python setup.py install

---

Quick Test

To test on the sample video:

python src/s5_test.py \
  --model_path model/trained_classifier.pickle \
  --data_type video \
  --data_path data_test/exercise.avi \
  --output_folder output

If you encounter errors related to OpenPose, run:

cd src/depends/tf-pose-estimation
python setup.py install

---

Citation

If you use this code or find it helpful, please consider citing our work:

@inproceedings{jian2022drown,
  title={A Drowning Detection System Based on Human Pose Estimation and Deep Learning},
  author={Jian, Jia-Xian and others},
  booktitle={2022 International Symposium on Computer, Consumer and Control (IS3C)},
  year={2022},
  organization={IEEE}
}

---

Acknowledgements

This work is based on:

• ildoonet/tf-pose-estimation
• philferriere/cocoapi
• Chenge Yang, Felix Chen

---

Contact

If you have any questions or suggestions, feel free to open an issue or reach me at:
Jia-Xian Jian – allensa119625@gmail.com