"Belco" kitchen goes live

It must surely be a good sign for the year ahead when the very first email I read on January 1, in a hung over state in Carlton, was a message informing me that a paper I submitted back in September, has been accepted for publication and presentation at this years ICRA(International Conference on Robotics and Automation). This conference is the biggest robot geeks gathering of them all, and attracts a fair bit of interest, so I am pretty happy about making the cut.

Apart from the obvious research benefits of getting more runs on the board, this particular achievement is made more significant in that the conference is in Rome this year. Having a paper to present while in Italy is very good for me (and avoids any possibility of my travel funding being suddenly taken away, not that I really feared that would happen). One slightly tricky thing is that the conference is about a week after Aff and I get married, so the honeymoon will happen straight after I give my talk (which will undoubtedly be in the last session of the last day, as per usual).

Another interesting fact about this paper is that this will be the first to feature Aff and my kitchen (referred to in the paper as the "belco" kitchen sequence). While working at home one day, with my fancy wide-angle lens camera, I walked it around our kitchen to construct an image sequence for one of my major results (though I wasn't really thinking I'd use it at the time). As it turns out, our kitchen is perfect for feature and motion detection, which is really just a nice (and slightly geeky) way of saying it's a bit of a mess.

So for your viewing pleasure (sorry Aff), I present ... the "belco" kitchen sequence.

The nitty gritty
The line you see moving around like a compass is actually indicating the estimated direction of forward motion. I use the visual motion of the scene as the camera moves along, to try and estimate which direction I am moving in. This is tricky because you can only extract this sort of information from the visual motion due to the camera's straight line motion, and so I first have to remove any motion due to the camera's rotation. Of course, I have no information other than what I see in these images to work with, so I can't just say "I know I rotated by this much, so deduct that from the visual motion". I have to estimate everything.

The blue and yellow dots around the circle indicate the direction in which the visual motion is going at that point in the image (blue is clock-wise, yellow is anti-clockwise). This information helps me "de-rotate" the visual motion.

Being a world-wide accessible blog, I should make it clear that much of what I have described above is based on some theoretical work that was done in the late eighties by Nelson and Aloimonos. As far as I know, the belco kicthen sequence is the first demonstration of this technique being used in real-time, on real image sequences. I am amazed this technique wasn't taken further a lot earlier than this.

Once I know which direction I am going, and I have removed the rotation, I can then build what are known as depth maps from the remaining visual motion. This is essentially a map showing what space is free and what is obstructed around the camera (which has a 190 degree field of view). I do this by exploiting the fact that things that are closer appear to move faster than things further way (just like bees). This is obviously pretty useful for navigating around unknown environments.

The paper's details:
"Real time biologically-inspired depth maps from spherical flow", by C. McCarthy, N. Barnes and M.V. Srinivasan. To appear in proceedings of IEEE ICRA 2007. Rome, Italy, April 10-13, 2007.