Model-based recognition of planar objects using geometric invariants

We address the problem of recognizing possibly occluded planar objects from an arbitrary viewpoint. The primitives used either to obtain models or to recognize the planar shapes are the so called metasegments. These are groups of consecutive segments obtained from the corners of the objects' contours. From these groups of segments is that a set of geometric invariants: two new (the type and the subtype), and two well known (the four and the five point dependent affine/projective invariants used in [1] and [2]) are first calculated and then entered into a database. This provides the essential mechanism of searching. These invariants (combined as a key) are then used first to build the database of models and second to extract from the model database those models possibly present in a test image. A voting and thresholding mechanism is finally used to decide which objects are present in the scene. The system's performance is tested with a set of polygonal objects.