Abstract
This paper describes the results of experiments on detection and recognition of 3D objects in RGB-D images provided by the Microsoft Kinect sensor. While the studies focus on single image use, sequences of frames are also considered and evaluated. Observed objects are categorized based on both geometrical and visual cues, but the emphasis is laid on the performance of the point cloud matching method. To this end, a rarely used approach consisting of independent VFH and CRH descriptors matching, followed by ICP and HV algorithms from the Point Cloud Library is applied. Successfully recognized objects are then subjected to a classical 2D analysis based on color histogram comparison exclusively with objects in the same geometrical category. The proposed two-stage approach allows to distinguish objects of similar geometry and different visual appearance, like soda cans of various brands. By separating geometry and color identification phases, the applied system is still able to categorize objects based on their geometry, even if there is no color match. The recognized objects are then localized in the three-dimensional space and autonomously grasped by a manipulator. To evaluate this approach, a special validation set was created, and additionally a selected scene from the Washington RGB-D Object Dataset was used.