Automatic generation of recognition strategies using CAD models

Abstract

A new scheme for generating CAP-based object recognition strategies is introduced. The strategy is used to locate a given CAD model in a scene. Similar to many previous model-based vision systems, the proposed scheme is divided into offline and online stages. In the offline stage, the CAD model is used to compile a tree, referred to as the recognition tree. The CAD model's features are organized in the tree, grouping similar features in a hierarchical manner. The recognition strategy is a series of filters derived from the recognition tree to localize the desired model in the scene. In the online stage, the scene is segmented and the detected features are grouped into a number of sets. The sets are the input to the filters; the sets where at least one member has satisfied the filter conditions are the output of the filters. The successful sets are the input to the next filter issued by the recognition tree. By issuing a series of filters, the number of successful sets decreases successively. The sets that have passed through a minimum of filters are the most probable candidates to match the desired model. >