drat / ML_based_CCTV

neural network + raspberry pi + cctv

What's this?

Neural network based object recognition CCTV which sends email once person is detected

How it works

1. Every 1s the latest frame is requested from video_stream
2. Frame is analyzed via ML
3. If human detected, for the next 5 seconds continue to request frames from video_stream and store them locally (without analyzing via ML)
4. After 5 seconds ask for new latest frame and analyze again
5. If no human detected, send email and return to 1, otherwise continue from step 3

• gathers as many frames as camera capable to produce + writing locally
• low cpu hit

• can't show on frame where object was detected

Thanks to

• https://github.com/HackerShackOfficial/Smart-Security-Camera
• https://github.com/EdjeElectronics/TensorFlow-Lite-Object-Detection-on-Android-and-Raspberry-Pi

How to make it work on unix machines

python3 -m venv venv
source venv/bin/activate
pip install pipenv
pipenv install

Project setup

Variables that need to be set: FROM_EMAIL, TO_EMAIL, FROM_EMAIL_PASSWORD, PHOTOS_STORAGE_LOCATION. Everything else can be default

cp .env.example .env

Run script locally

python3 src/TFLite_detection_webcam.py --modeldir=coco-model --resolution=1600x1200 --framerate=30

Troubleshooting

Download photos from camera

scp -r pi@192.168.1.175:/home/pi/tflite/photos .

Commands to check storage usage

du -hs .
du -h .
df -h .
df -h

Commands to debug supervisor

sudo supervisorctl status
sudo supervisorctl restart all
tail -fn 100 /var/log/motioneye-cctv.log