View planning for efficient contour-based 3D object recognition

This paper presents a method for capture planning in view based 3D recognition. Views are represented by their contours, encoded into curvature functions, which are reduced into compact feature vectors by Principal Component Analysis. These vectors are very resistant against transformations, so they can be assumed to be distributed over the surface of a sphere with the object in its center. After clustering these vectors, 3D objects are represented via Hidden Markov Models where classes are states. To recognize an object in a minimum number of steps, we propose to align candidate cluster representations and then subtracting their cluster maps to decide in which locations they differ the most. Then, a TSP is used to decide in which order these distinctive locations are visited. The proposed approach has been successfully tested with several artificial 3D object databases, even though it still presents some errors in objects with strong symmetries.