It actually started talking with a friend at lunch about taking pictures from python without the motion part and I decided to add the motion detection later on. It's definitely not the best thing, motion looks much better, but it was quite fun to do and only took like 1h30 (and I wanted to play with OpenCV !)

Can anyone comment on whether it's worth using Motion or ZoneMinder or similar today? I was thinking about using ZoneMinder for a simple 4-camera setup, but it seems like there are kits (with proprietary software) which are cheaper and easier to set up.

I use motion at home.

Interestingly, the CPU usage went from 0 to ~30% when I added the first camera, but there was no noticeable increase in CPU usage adding the other 2 cameras.

I use it to record any detected motion, works a charm. It even has a feature to dynamically ignore parts of the screen. Say part of your camera's field of view catches a plant that blows in the breeze, the system will learn to ignore that area of the screen.

Edit: In fact, you can use motion to record images of the motion that it detects. You can then plug those images into OpenCV for facial recognition. Both Motion and OpenCV are very powerful pieces of software and they can compliment each other nicely.

Edit: ... or Reply.

Setting up motion is easy enough if you have some scripting skills - a lot easier than the opencv approach in the article. There are articles online using motion with raspberry pi and a USB webcam to do the sort thing described in the article (eg http://m.instructables.com/id/Raspberry-Pi-remote-webcam/).

Yes; back in 2005/2006 I used Motion to implement a system using four cameras on a 1GHz Intel P3; it worked nicely.