An attempt to harness the power of Deep Learning to come up with a solution that can let us detect various classes of activities an infant, toddler or a baby is performing in real-time. This POC can then be published as an end-to-end deployable cloud project.
The model does not restrict predictions for babies only, it is applicable to all entities that appears in a human posture. So temporary, this needs to be handled at project level.
Special thanks to nicknochnack/ActionDetectionforSignLanguage repository for putting up such helpful content. Without it this project might have never existed.
Data was collected from YouTube video clips. Human pose keypoints were extracted with the help of MediaPipe.
- Baby Walking
- Baby Still (no movement, can be considered as sleeping)
- Baby Crawling
Create a new environment and use below command for installing all required packages
pip install -r- requirements.txt- Rename your baby video as input.mp4 and place it inside
/rawdirectory. - Open cmd and traverse to the project directory.
- To run the prediction script, just do:
python prediction.py Downloaded Videos - Drive Link
Navigate to streamlit directory inside root project
cd streamlit
Run the application
streamlit run .\app.py
YouTube demo for the application: https://www.youtube.com/watch?v=ZIDhvSGXDmI