Saturday, December 19, 2009

Omnidirectional stereo vision

Here I'm doing an experiment with combining omnidirectional and stereo vision. A single camera with a wide angle lens is mounted underneath five hemispherical mirrors (christmas tree decorations sawn in half). The obvious disadvantage is the small resolution for each mirror, but the advantages are:
  • You only need a single camera
  • There are no camera synchronisation issues - all the mirror images are guaranteed to be synchronised
  • The materials needed to make it are extremely cheap - you could make this for less than £50.
  • Very wide field of view conducive to mobile robot navigation problems
  • Multiple triangulations and multiple baselines permits better disambiguation
The geometry is more complicated than for traditional stereo vision, but it shouldn't be intractable and I'll be able to create a lookup table between 3D voxels and pixel coordinates.



A similar vision system in the natural world.

4 comments:

Kevin said...

Why 5 Bob? Why not two or three?

Bob Mottram said...

At present this is all very experimental, so I havn't investigated what the minimum number of mirrors is. I'm just following the intuition of "the more the merrier". The more mirrors there are the more evidence can be collected for each 3D voxel.

If we compare the colours of two rays from a pair of mirrors intersecting at an origin in space there will be a lot of ambiguity, but as more mirrors are added the chances of four or five intersecting rays having the same colour (or other properties) is reduced.

terren said...

I may be way off here but this reminds me a little of insect eyes... except instead of 5 "mirrors" you have thousands. My actual knowledge of insect eyes is pretty limited though.

Bob Mottram said...

Yes. From the geometry point of view this is quite similar to the compound eyes of insects, except in the insect case the focus point is internal rather than external. Spiders also have a similar arrangement and some have good stereo vision.