We have demonstrated that in the restrictive scenario of tracking the rear of a car under the assumption of 2D affine projection, vision can produce reasonably accurate range measurements over a long sequence of images. Examining Figure 4 , there does not appear to be a clearly best method of selecting between stereo and motion estimates of range. The estimates from motion are generally smoother, but they rely on the stereo estimates to solve for the inherent scale factor. The best solution may turn out to be averaging the motion estimates, while using stereo disparity at small ranges to re-scale the motion estimates.
The next stage is to implement the vision algorithms on a network of C40s. The current speed on a 150MHz Pentium is approximately 1s per image pair, and we can expect to achieve about 3Hz performance on 3 or 4 C40s. Complementing this relatively slow, but robust tracker will be a fast and simple correlation-based tracker. We expect to achieve 30Hz performance with this layered tracking approach.
The software was written using the C library ``Horatio'' which has been
developed at Oxford and Berkeley for the purpose of supporting efficient
computer vision applications. HTML documentation for Horatio may viewed
from the first author's WWW home page at
http://www.cs.berkeley.edu/
pm/
, and the complete library may also be downloaded from that site.
Adrian F Clark