Hypothesizing undetected and occluded three-dimensional features using predicate logic based spatial reasoning

A method for hypothesizing the locations of features that may have been undetected by low level three-dimensional image processing algorithms is presented. The procedure relies on a predicate logic based approach to object recognition in which viewpoint dependent (observed) data is transformed into a set of viewpoint independent assertions. The detected features in these viewpoint independent assertions are variabilized and the assertions are combined to form a clause which is then tested against known models for consistency. If the features in a known model are able to instantiate the free variables in the variabilized clause, then a subgraph isomorphism is established between the features in the observed data and those of the model. Missing or undetected features in the observed data are then sought by examining the semantic relationships between model features that are instantiated and those that are not. Results of applying the method to synthetic range data are presented.