Grouping based on projective geometry constraints and uncertainty

The process of grouping and subsequently recognising objects in cluttered images is one laden with difficulties; however, results can be greatly enhanced if the inherent uncertainty of image-features is taken into account. This paper shows that starting with the individual edgel's uncertainty it is possible to calculate covariance-information for all derived quantities. This information can be used to choose between competing algorithms, selecting the one that produces the more reliable results, but also as an aid during the recognition process. The consequent application of error-propagation leads to a new formulation for the calculation of the cross-ratio, which is both robust and efficient in dealing with measured lines, and does not require knowledge about the vanishing point. Extensive Monte-Carlo simulations as well as the application to images of cluttered street-scenes demonstrate the robustness and suitability of the approach.