Our proposed solution is a Bayesian inference scheme, based on several layers of information processing. Starting from range images of the objects, a chain of shape analysis processes produces first a segmentation of large, reliable features on the objects, whose shape is modelled by geometric primitives (superellipsoids in the proposed test-bed solution). The shape parameters are used for part recognition (via a Bayesian discriminative classifier) and for placing ``landmark'' points upon the parts themselves. Eventually, the landmarks are used to estimate the target point position. Uncertainty in the modelization process is carried along, and expressed in the form of posterior probability distributions about the target. This effectively quantifies the confidence degree in the estimation.
Extensive tests of the above technique on test objects have given promising
results about the practical feasibility of the proposed methodology. More
details are available on line:
F. Callari, G. Soucy, F.P. Ferrie (McGill), D. Baird, D. Lamb (Hymarc Ltd.)