Automated 3D object recognition and dynamic library entry/update system

A method for automatic library entry and update of three-dimensional representations of 3D rigid, planar-faced objects is presented in this paper. Geometric and structural information is obtained from multiple monocular 2D image frames and placed in relational tables. Point correspondences are found from the relational tables to reconstruct the 3D scene object. A number of global descriptors, obtained from the reconstructed object, are used as indexes to the model library to select similarly structured models for matching. A probabilistic procedure using the indexes determines whether the scene object is already represented by a library model. If the model is represented, a technique based on the hypothesize-and-test paradigm compares the scene object to the library model to update the model representation with any new geometric and structural information. If the model is not present in the library, the new model representation is added to the model library