Silhouette-Based Object Recognition with Occlusion Through Curvature Scale-Space

Abstract

A complete and practical system for object recognition with occlusion has been developed which is very robust with respect to noise and local deformations of shape as well as scale changes and rigid motions of the objects. The system has been tested on a wide variety of 3-D objects with different shapes and surface properties. No restrictive assumptions have been made about the shapes of admissible objects. An industrial application with a controlled environment is envisaged. The Curvature Scale Space technique [4, 5] is used to obtain a novel multi-scale segmentation of the image contour and the model contours using curvature zero-crossing points. Multi-scale segmentation renders the system substantially more robust with respect to noise and local shape differences. Object indexing [9] is used to narrow down the search-space and avoid an exhaustive investigation of all model segments. A local matching algorithm applies candidate generation, selection, merging, extension and grouping to select the best matching models.