Fast 3D keypoints detector and descriptor for view-based 3D objects recognition

Ayet Shaiek (1) , Fabien Moutarde (1)

1 CAOR - Centre de Robotique

In this paper, we propose a new 3D object recognition method that employs a set of 3D keypoints extracted from point cloud representation of 3D views. The method makes use of the 2D organization of range data produced by 3D sensor. Our novel 3D interest points approach relies on surface type classifi-cation and combines the Shape Index (SI) - curvedness(C) map with the Gaus-sian (H) - Mean (K) map. For each extracted keypoint, a local description using the point and its neighbors is computed by joining the Shape Index histogram and the normalized histogram of angles between normals. This new proposed descriptor IndSHOT stems from the descriptor CSHOT (Color Signature of Histograms of OrienTations) which is based on the definition of a local, robust and invariant Reference Frame RF. This surface patch descriptor is used to find the correspondences between query-model view pairs in effective and robust way. Experimental results on Kinect based datasets are presented to validate the proposed approach in view based 3D object recognition.

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]

Ayet Shaiek, Fabien Moutarde. Fast 3D keypoints detector and descriptor for view-based 3D objects recognition. International Workshop on Depth Image Analysis (WDIA'2012) of 21st International Conference on Pattern Recognition (ICPR'2012), Nov 2012, Tsukuba, Japan. ⟨hal-00766679⟩