Motion detecting using onboard camera

Hi,

Think simple   . Those guys are not doing anything complex, because of performance reasons.

I tried doing something like this once, for an entirely other purpose though. The key is to do hierarchical motion estimation - do it on hierarchically subsampled images and it will be fast and accurate enough.

For example:

Subsample two to N of consecutive  160x120 viewfinder frames to 80x60, then 40x30, then 20x15 "images". Compare always two to N consecutive frames of the same size, starting from the LOWEST resolution, moving the other image around the other one. The comparison operation here can be e.g. sum-of-absolute-differences of pixel values. When you have found a value below some threshold, or you have done enough comparisons (you have a fixed time budget), stop the iteration and you should have a motion vector ready

Oh, I forgot to mention that it may be useful to do lookup table to calculate the difference of two 16-bit pixels. In that way, you can use e.g. squared sum values so the end result will be more accurate.

I hope this helps, no code this time though Anyway, even the above suggestion is probably more complex than what is used in the programs that you mention.

The downsize of all this is that you have to have a good amount of non-ambiguous content in the viewfinder for the algorithm to perform well.

Cheers,
Pawel

I recently found this page:

http://www.inf.ethz.ch/~rohs/visualcodes

It has some info about this and similar subjects, along with books in .pdf and source code.

Hope this helps