Skip to content

The combination of face identification and action recognition for fall detection (ICT Journal)

Notifications You must be signed in to change notification settings

DuyNguDao/Identity-Action

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

The combination of face identification and action recognition for fall detection

Introduction

Falls are a very common unexpected accident in the elderly that results in serious injuries such as broken bones, and head injury. Detecting falls and taking fall patients to the emergency room on time is very important. In this project, we propose a method combining face and action recognition for fall detection. Specifically, we identify seven basic actions in elderly daily life based on skeleton data detected using the YOLOv7-Pose model. Two deep models, Spatial-Temporal Graph Convolutional Network (ST-GCN) and Long Short-Term Memory (LSTM), are employed for action recognition on the skeleton data. The experimental results on our dataset show that the ST-GCN model achieved an accuracy of 90% higher than the LSTM model by 7%.

video demo

recog_recording.mp4

System Diagram

Dev

Member:
 - DAO DUY NGU
 - LE VAN THIEN
Instructor: TRAN THI MINH HANH

Usage

Install package

Step 1: Download and setting anaconda environment

https://www.anaconda.com/download/success

Step 2: Open Anaconda Prompt

Anaconda Prompt on Window

Step 3: Change directory to project

cd Identity-Action

Step 4: Setup python environment (GPU is required)

setup_environment_gpu.bat

Remove python environment

remove_env.bat

Download

Model yolov7 pose state-dict: (Note: the model download fee is 20$ through PayPal. Please send information to my email to download. - [email protected])

yolov7_w6_pose

Quick start

Start and config URL

pip uninstall opencv-python-headless
pip install opencv-python==4.5.5.64
python run_video.py

Start with the terminal

pip uninstall opencv-python-headless
pip install opencv-python==4.5.5.64
python detect_video.py --fn <url if is video or 0>

Run App

Step 1: Open Anaconda Prompt

Anaconda Prompt on Window

Step 2: Change directory to project

cd Identity-Action

Step 3: Run app

run_app.bat

Datasets and result model training

Dataset human action

Human action

Dataset Face Detection

Face detection

Dataset Face Recognition

Face recognition

Result face recognition

Diagram

Face detection

FDDB DATA

  • Confusion matrix of YOLO5Face

  • Confusion matrix of RetinaFace

WIDERFACE Val

  • Confusion matrix of YOLO5Face

  • Confusion matrix of RetinaFace

Result compare: Accuracy, Precision, Recall, Time processing

Config Computer:

  • CPU: AMD Ryzen 7 4800H với 16G RAM DDR4
  • GPU: NVIDIA GeForce GTX 1650 với 4G RAM DDR6

Face landmark loss

Face recognition

Backbone compare:

  • MobileFaceNet

  • ResNet18

Methods compare: LOOCV (Leave-One-Out Cross-Validation), Time processing with FaceScrub data

Diagram Accuracy Thresh

Result human action pose

Diagram

Backbone

LSTM

ST-GCN

Two Stream ST-GCN

Confusion matrix

  • Model LSTM (Long Short Term Memory)

  • Model ST-GCN (Spatial Temporal - Graph Convolutional Network)

Methods compare

Accuracy, Precision, Recall, F1-score, Time processing

Config Computer:

  • CPU: AMD Ryzen 7 4800H with 16G RAM DDR4
  • GPU: NVIDIA GeForce GTX 1650 with 4G RAM DDR6

Compare ST-GCN YOLOv7-Pose and ST-GCN YOLOv3 Alphapose

Config Computer:

  • CPU: AMD Ryzen 7 4800H với 16G RAM DDR4
  • GPU: NVIDIA GeForce GTX 1650 với 4G RAM DDR6

  • Confusion matrix of ST-GCN with skeleton data export from yolov3 alphapose

Training

Human action

Face Detection

Face recognition

Citation

@article{
  title={The combination of face identification and action recognition for fall detection},
  author={Dao Duy Ngu, Le Van Thien, Tran Thi Minh Hanh, Nguyen Thi Hong Yen, Dao Duy Tuan},
  journal={Journal of Science and Technology, Issue on Information and Communications Technology, ISSN: 1859-1531},
  Pages={37-44, Vol. 20, No. 12.2, 2022}
  year={2022}
}

Contact

The University of Da Nang, The University of Science and Technology
Address: 54, Nguyen Luong Bang street, Lien Chieu district, Da Nang City, Viet Nam

Acknowledgements