The purpose of this project is to develop a basic face recognition system using picamera2 and OpenCV on the RPI. The major steps are as follows:
Here's the list of equipment that used for this project:
I have a standard camera, but there's also a noir variant, which is designed to take pictures in the dark when used with an infrared light source.
Make a dedicated directory for this project and go there:
> mkdir Recognize
> cd Recognize
Create and activate the virtual environment:
> python3 -m venv face_recognition_env --system-site-packages
> source face_recognition_env/bin/activate
When you activate the virtual environment you should see its name (face_recognition) at the beginning of the command line prompt, in parentheses.
Info: You can deactivate the virtual environment by typing deactivate at the command prompt.
Verify that the camera is working and take a test photo with these two commands:
> rpicam-hello --list-cameras
> rpicam-still -o test.jpg
The first command should show a list of cameras with a response that looks like this:
The second command opens a preview window for a few seconds, saves the picture in a file called test.jpg, followed by a command line message saying: Still capture image received. You can use the built-in viewer to verify the picture.
You can also use the following script to verify that your picamera2 integration is working:
# Quick test script
python3 -c "
from picamera2 import Picamera2
import time
picam2 = Picamera2()
picam2.start()
time.sleep(2)
picam2.capture_file('picamera2_test.jpg')
picam2.stop()
print('picamera2 test successful!')
"The script should capture and save a picture called picamera2_test.jpg in your local directory, which you can inspect and verify with the viewer on your RPi.
I decided to use pictures of Nelson Mandela and Jimmy Carter for this project. There's a wide range of pictures of the two of them on the Internet, some from their younger days and others that were taken when the two were older.
I Created a folder called reference_images in the project directory, and added five clear photos each of Mandela and Carter. To keep the program simpler, I named the pictures like so: mandela_1.jpg, mandela_2.jpg, carter_1.jpg, etc.
When you're ready to test the face recognition system, you'll have the option of directing the camera either to pictures on the computer screen or unsuspecting family members. Consider testing the following scenarios:
The first two groups will be relatively easy true-positive pictures. Groups 3 and 4 will be more interesting to test, to see whether the algorithm will score true-positives and true-negatives, and no false-positives or false-negatives. Group 5 should be a confirmation of true-negatives and hopefully won't yield any false-positives.
You can find the python script on GitHub: https://github.com/skaltounian/Recognize
Here's what the core components do:
Face Database:
face_recognition librarypickleCamera Interface:
picamera2 for native Raspberry Pi camera integrationFace Recognition:
Display System:
picamera2 captures RGB imageOnce you're save the script, make it executable by typing:
> sudo chmod +x face_rec_picam2.py
Be sure that your virtual environment is active and run the python script:
> source face_recognition_env/bin/activate
> python face_rec_picam2.py
dlibThis was a tough nut to crack. Having a slow RPI-3 made installing dlib difficult. Compilation took forever - that is, it didn't complete. Having an SSH connection to the RPI didn't help either. At some point, when I had walked away, my Windows computer went to sleep and lost network connection. When, on the next day, I tried to reconnect, the RPI failed to connect. So I rebuilt the RPI. Twice.
The solution was to:
The RPI-3 has limited RAM, so building dlib can crash without enough virtual memory.
Reconfigure the swap memory by editing the configuration file with:
> sudo nano /etc/dphys-swapfile
Change the line CONF_SWAPSIZE=100 to CONF_SWAPSIZE=2048
Save and exit, then apply the changes by restarting the swap service with:
> sudo systemctl restart dphys-swapfile
To verify the new swap is active use the command:
> free -h
Look at the Swap: row, where you should see approximately 2.0G. If it still shows 100M or less, double check the file and restart the service again.
I'm glad I had Git and my code was already on GitHub.
At one point during the project I had gone down so many rabbit holes that I decided to start with a clean slate. One of the reasons for my troubles was that both Claude and ChatGPT assumed the RPi-3 was using libcamera, and I had to remind them (after much experimentation) that since I was running Bookworm we should be using rpicam.
To clean up:
face_recognition_env> deactivate
> rm -rf face_recognition_env
> sudo apt remove libcamera-apps libcamera-tools
> sudo apt autoremove
> sudo apt autoclean # clean package cache
Then, make sure that rpicam is still available:
> rpicam-hello --version
> ls /usr/bin/rpicam*
By default, the camera in my setup captured images upside down. This apparently prevents OpenCV2 from recognizing faces, given that all training data for face recognition has faces that are straight on and right side up. Here's how you can configure your camera to rotate the images:
from picamera2 import Picamera2
from libcamera import Transform
picam2 = Picamera2()
transform = Transform(rotation=180)
camera_configuration = picam2.create_preview_configuration(
main={“size”: (640, 480), “format”: “RGB888”},
transform=transform
)