Realtime Terminator Salvation "Machine Vision" fx

Have you seen Terminator Salvation yet? There's a bunch of cool visual effects developed by Imaginary Forces, it shows the world as seen by machines. There's a lot of object tracking going on there, I was thinking whether I could recreate the whole thing just in pure AS3. And, well, here's the result (which I am actually very proud of) ;-)

Terminator Salvation Machine Vision in AS3 by Og2t

Click image to activate, wait for the video to buffer (1.6MB) press EDIT button to play with the filters (in full screen mode). Enable your webcam (if you have one) and play about with sliders and checkboxes – try if your face can be tracked too – but then watch for evil Terminators – they'll come and get you! ;-) Btw. you can turn histograms for every filter - thanks to Quasimondo for the code.

This is a part of the whole video filter framework I am developing just now, the inspiration came from Joa Ebert's Image Processing library (as far as I know, he's cooking a complete rewrite). The full source code (including Pixel Bender kernels and examples) will be soon released on Google Code and will feature face/eye tracking/gestures and few other things (surprise!) A lot of people are very sceptic about the whole eye tracking idea, they don't believe it's precise enough to make any use of it – I will prove that it is, and it works! (just watch closely how it tracks my eyeballs on the video!)

My approach is to make everything as much simple as I can. If something cannot be achieved using this rule, I either abandon the idea completely or look for a simpler solution.

The face tracking is actually relatively simple, I will briefly describe each step:

  1. Brightness/Contrast (HBSC filter) - initial adjustment of the input (will be replaced with auto levels)
  2. Motion Capture - works the same way as the "movement watchdog" that's implemented in brains of almost all animals (including humans) in order to survive – it finds the rectangle area of the all the differences between two frames. This step could be much more complicated (i.e. I might use face detection or Eugene's motion tracker once he decides to release the source) but simple motion capturing is good enough for Machine Vision experiment here.
  3. Shape Depth Detector – finds centres of colour local maximums, play with the levels slider carefully to get more details – it works by posterising the image then does a very fast blob detection on every result colour – thanks to Kynd and Kampei Baba for sharing this technique.
  4. Color Grading – identical to Photoshop's Gradient Map – uses paletteMap to remap the colors.
  5. Machine Vision – the final and the most complicated filter – utilises Delaunay triangulation and Voronoï diagram by Nicoptere – it's fast enough to process it realtime (thanks for sharing!). Then it plots the points and lines and applies my spotlight effect class (another blog post on that subject coming soon) to achieve the final look. Btw. I've found another very cool experiment using Delaunay for face triangulation by Neuro Productions.

Other thanks goes to Mr. Doob for his stats widget, Bit-101 for the Minimal Comps and SubBlue for lots of inspiring technical discussions we've had during lunch breaks at tictoc.

Feel free to leave any comments questions and suggestions, I am really interested what you think. You can also follow my blog updates on Twitter or RSS. It's getting very late now, so I better go.

UPDATE: I am giving up, it's just too hard to track human's head, I gonna do next experiments with chickens:

If you liked this experiment, make sure you see the new version.

11:53 PM | | 5 Comments | Tags: , , , , , ,


  1. Hi there,

    Looks good! Is the source coming availible soon? Can use it for my installation at the todaysart festival 2009 (

    Is the motiontracker the same as this guy's? :

    Beer van Geer on
  2. Is there anyway to improve tracking for those with darker skin tones?

    Mark on
  3. Wow!!! thats wonderfull!


    alfathenus on
  4. Thanks for linking back Alex, you're right about the noise, I haven't got time to implement other motion detection methods yet but I'll have a look into the camshift algorithm you've mentioned, it definitely could become a part of the framework.

    Og2t on
  5. Great work! Love it. But the motion based tracking isn't very stable when there's background noise (movement). Maybe a "weight - based" motion detection could help. The eyetracker is awesome! Unfortunately I wear glasses so I couldn't really write an eyetracker myself (coding without glasses sucks)

    Really great work!

    alex on

Adding comments disabled for now.