I've added a second rearward looking stereo camera to the robot. This isn't yet calibrated, but that's a fairly straightforward procedure which I always intended would be done with the cameras in situ. I'll probably also need a separate calibration procedure to characterise the pan and tilt behavior as well as possible. This will allow the grid maps to be updated properly, taking the head pose uncertainty into account.

http://farm3.static.flickr.com/2099/2205545549_e6bea41a36.jpg

One consideration when building the stereo head was whether the cameras should be rolled to a 45 degree angle or not. For simple types of stereo correspondence detecting vertically oriented features along each row of the image camera roll can be a good strategy allowing a mix of both vertically and horizontally aligned features in the environment to be ranged. However, for the stereo correspondence algorithm which I'm using rolling the cameras doesn't make all that much difference. This is mainly because I'm using simulated multi scale centre/surround fields, which turn each image into a kind of contour map, and also explicitly taking vertical context into account. In this case vertical context extends significantly beyond the usual patch matching windows. Hence features which may appear identical when narrowly viewed along each image row can be disambiguated from their wider surrounding context. There are still classic dilemmas, such as the "picket fence" effect, but vertical context tends to rule out all but the most mesmorizing situations.