Object detection in robot picking applications using 3d cameras

This work relies on analysis and testing of two 3D cameras in robot picking applications. Continuous progress in microelectronics, micro optics and micro technology made 3D cameras affordable and competitive with respect to 2D cameras in common industrial and commercial applications. 3D cameras, in fact, could give advantages in terms of timing and performances in application involving objects normally processed with 2D cameras. These cameras are intrinsically different in the technology used and provide mono/color images, depth maps and point clouds. First camera considered, Intel RealSense D435, is a stereo camera and the second one, Basler blaze 101, is a time of flight camera. A particular focus is devoted to enhancing eventual advantages of using one technology with respect to the other. It’s important to underline that objects examined are quite small and heterogeneous in terms of material, opacity, colors, textures and transparency. Qualitative considerations will be done to a priori exclude from tests objects which depth maps result difficult to be created and exploited. Then, through computer vision algorithms, depth information are used to detect graspable objects and to get their positioning and orientations. These algorithms are capable of managing both isolated, touching or overlapping objects, establishing which one can be picked up and which not. Three different algorithms will be deeply introduced and described with a detailed focus on best camera settings to accomplish these tasks. Several analyses are done with both static and moving products on conveyor belts, varying conveyors colors and shapes.