Succes! Of sorts:

This video demonstrates the effect that I was hoping for: moving recognizable shapes with incredibly low resolution (50 "pixels" in all). Notice that if you pause the video above, the still image shows nothing resembling a circle. The resolution is simply too low to represent such a shape. Yet, as soon as the "shadow" starts moving its shape is definitely circular and you might even spot that the shape is somewhat elliptical.

There is a lot more going on than meets the eye: First of all all the positions of the randomly distributed LEDs have been automatically registered of course, because that is what this project is all about. Secondly, the edges of the shadow circle need to be fuzzy. The effect is as if the LEDs are partially occluded when the edge of the circle moves over them. If this isn't done, then the movement is a lot more pixelated and the shape isn't recognizable anymore. Thirdly: this works a lot better for a dark circle on a light background than the other way around. Apparently, the human brain is better at discerning shapes if they occlude light sources than they are at identifying shapes "made of light". Finally, there's a minimum speed at which the object can move and still have a recognizable shape. If it moves too slow, its not recognizable as a circle anymore.

All of this is teaching me a lot about the human visual system.

In order to get the result above, the original plan had to be changed in two ways:

• Although the LEDs are in a random pattern, they weren't "draped". Instead I just drilled holes in a plywood board. This makes recognizing the LEDs an awful lot easier since now they're nicely spaced and there are no strange reflections
• In the quest for working OpenCV registration of LED positions I'd already switched to a much simpler pattern, where the LED just light one-by-one. This had mixed results with the draped LEDs, but it worked from the start with the nicely spaced and separated LEDs-onna-board.

This is the pattern that was used to register LED positions:

The code that creates this pattern using a plain AVR is in the function simple_registration(). This pattern is then analyzed using the code in LedMapping.cpp. Finally, the output of LedMapping.cpp is used to hard-code the LED positions in a demonstration program.

The board as it is now will be placed under the awning of the front door, for Christmas only and I'm not so much interested in creating floating ghost shadows there (hmm. definitely an option for Halloween). I think I'll stick with the simple location-independent random starry sky pattern for the holidays:

Given that there is quite some hardware work to do (painting, getting the 5V power to the board outside) limiting the scope at this point should increase the chances of the project actually being finished before Christmas...