Constraints for Object Recognition in Aerial Images - Handling of Unobserved Features

In this paper we will show how constraint solving methods can be applied for the recognition of buildings in aerial images. Object models are transformed to constraint representations which are matched against extracted image features. To cope with disturbances caused by occlusions and noise, we distinguish between the unobservability of a) relations between object parts and b) object parts themselves. Whereas other approaches for solving over-constrained problems suggest to reduce the relaxation of a variable to the relaxation of its incident constraints, we argue that both cases have to be treated separately. Information theory is applied to derive constraint weights on a probabilistic basis. We extend constraints and variables in a way which provides for an adequate integration of constraint violation and variable elimination on the one hand, and allows the determination of the maximum likelihood estimation for the matching between model and image on the other hand.