Object Recognition by Registration of Repeatable 3D Interest Segments

Abstract

3D object recognition using depth data remains a difficult problem in computer vision. In this thesis, an object recognition system based on registering repeatable 3D surface segments, termed recognition by registration (RbR) is proposed. The goal is to eliminate the dependency on local geometry for establishing point-to-point correspondences, while maintaining the robust trait of a global technique by applying a pairwise registration process on these individual segments. The extraction of repeatable surface segments is achieved by inheriting the high repeatability of interest points, most often utilized to increase the performance of matching local shape descriptors. Precisely, dense sets of interest points are connected to form the surface segment boundaries by greedily optimizing a smoothness constraint. The reconstructed boundaries provide an effective means to facilitate fast 3D region growing on the object and scene surfaces, forming the 3D interest segments. Pairwise registration of the model and scene interest segments must then consider the imperfectly extracted segments due to data noise, occlusion, and error from the segmentation itself. An adaptation of the robust 4 points congruent sets (4PCS) registration algorithm was shown to register interest segments efficiently with great success. This is achieved by utilizing the prior knowledge of the 3D model interest segments that can be preprocessed, coupled with pose clustering of the retrieved transformation candidates. Experimentally, the interest segment repeatability, registration rate, and the object recognition rate were evaluated using a variety of free-form objects in 3D model data corrupted with synthetic noise and real 2.5D cluttered scenes. It was found that the interest segments are highly repeatable (>80% per top segment per scene), and that they can also be registered successfully within a reasonable number of RANSAC cycle of the 4PCS algorithm. Compared to other state-of-the art local approaches, RbR enjoyed superior object recognition rates in both accurate LiDAR data and noisy Kinect data (on average >90% for all objects tested in both sets of data), demonstrating that the approach provides a very attractive alternate solution to those in the current literature.