3D Keypoints Detection for Objects Recognition
Résumé
In this paper, we propose a new 3D object recognition method that employs a set of 3D local features extracted from point cloud representation of 3D views. The method makes use of the 2D organization of range data produced by 3D sensor. A detector of 3D interest points requires the expression of the local surface variation around points. In our case, we opted for a curvature-based approach. We test six methods which combine principles curvatures values under the form of: 1) a measure of the Shape Index (SI), 2) a measure of a Quality Factor (FQ), 3) a map of Shape Index (SI) and curvedness(C), 4) a map of Gaussian (H) and Mean (K) curvatures, 5) a combination of 3 and 4 (SC_HK) and 6) a combination of 5 and 4(SC_HK_FQ). For each extracted point, a local description using the point and its neighbors is done by combining the shape index histogram and the normalized histogram of angles between normals. This local surface patch representation is used to find the correspondences between a model-test view pair. Performance evaluation of the detectors in terms of stability and repeatability shows the robustness of the proposed detectors to viewpoint variations. Experimental results on the Minolta data set are presented to demonstrate the efficiency of the proposed approach in view based object recognition.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...