Who's at the door? Face recognition

Introduction

Face detection finds a face. Face recognition goes one step further: it works out whose face it is. That's how a phone knows it's you and not your brother.

In this lesson you'll build a pretend smart doorbell. It learns the faces of people in your family, then greets each one by name, and says "Stranger!" for anyone it doesn't know.

Working principle

• Add class as (name) from (camera) takes a photo of the face in front of the camera and saves it under a name. Each person is one class.
• Do face matching on (camera) looks at the face in front of the camera and compares it with every saved class.
• Get class of face detected reports the name of the closest match.
• Is class detected (name)? is a true/false block for one person.
• Reset class deletes every saved face.
• The saved faces are stored inside your project file, so only save faces of people who said yes.

What you'll need

• PictoBlox with the Face Detection extension
• A webcam
• Two or three family members who have agreed to take part

Working in real life

Step 1: ask first. Faces are personal. Ask each person: "Can I save your face in my project? I'll delete it when we're done." Only add people who say yes.

1. Add the Face Detection extension and make a sprite called Doorbell (a door or a bell from the sprite library).
2. Build the training scripts below. Each key saves a face for one person. Press each key three or four times while that person looks at the camera, turning their head a little each time.
3. Build the doorbell script. Click the green flag and let each person walk up to the camera.

Code

Training (one script per person; change the key and the name):

when [a] key pressed
add class as [Amma] from [camera]
say [Saved Amma] for (1) seconds
 
when [b] key pressed
add class as [Papa] from [camera]
say [Saved Papa] for (1) seconds
 
when [r] key pressed
reset class
say [All faces deleted] for (2) seconds

The doorbell:

when green flag clicked
turn [on] video on stage with (0) % transparency
forever
  analyse image from [camera]
  if <(get # faces) > (0)> then
    do face matching on [camera]
    say (join [Welcome home, ] (get class of face detected))
  else
    say [Nobody at the door]
  end
  wait (1) seconds
end

If someone who isn't saved walks up, the model still picks the closest match. Try it: does it call a visitor by the wrong name? Real smart doorbells only accept a match when it is very close, and they are trained with many more photos.

Try it

• Play a different sound for each person.
• Make the doorbell say "Stranger!" when is class detected (Amma) and is class detected (Papa) are both false.
• When you've finished, press r to delete the faces, and save your project.

Conclusion

• Face recognition compares a face with saved faces and picks the closest match.
• Each person you teach it is called a class, and a few photos from different angles make it more reliable.
• Faces are personal data: ask before saving one, and delete it when you're done.

What next?

Now let the computer see your whole body: Catch the stars: body pose detection.