Shape Matching of Repeatable Interest Segments in 3D Point Clouds

Abstract

A novel approach to object recognition based on shape matching of repeatable segments is presented. The motivation is to increase the recognition system robustness in handling problems such as noise corruption at a local level, featureless surfaces, and variations in 3D data sources. Inspired by the detection of repeatable interest points, interest segments were extracted through region growing and the reconstruction of piece-wise boundary curves from connected interest points. An object pose is automatically estimated if only one of the repeatable scene segments can be matched and aligned correctly with a model segment. To demonstrate this capability, shape matching of selected segments, filtered by size, were registered using the 4 points congruent sets (4PCS) algorithm and compared with an overlap metric. Three different free-form objects were evaluated against nine different occluded and cluttered 2.5D scenes. It was found that on average 1.4 ± 0.8 scene segments can be matched correctly to a model segment in the database, indicating that a highly robust object recognition system will result.