Seite - 75 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

(a) 165 125 270 70 72 168 134 unit inmm FrameF4&F6 is shifted backward by the DH algorithm. x0 y0 z0 x1 y1 z1 x2 y2 z2 x3 y3 z3 x4 y4 z4 x5 y5 z5 x6 y6 z6 m2 (b) (c) Fig. 2: (a) ABB IRB 120 point cloud model - the (joint) frames are set-up by DH convention. Each link is implemented as an independent model but coupled to its neighbor over the respective DH transformation; (b) ABB IRB 120 kinematic structure; (c) Robot depth image captured by two IntelR© RealSense R200 in a distance of 1.5m. captured pattern in comparison to the original one to get thedepth informationby triangulation.Third,Time-of-Flight (ToF) cameras [17]where the depth information ismeasured over the elapsed time of pixel-wise emittedmodulated light signals reflected by the detectable object. RGB-D cameras provide point clouds (with position in- formation in R3) generated from depth data. Thus, it is not necessary to seek for interest points for orientation estimation since the detected objects are already available as 2.5Dobjects in theworkspace. Similar to an image-based approach a 3D model of an object can be matched to the point cloud via the Iterative Closest Point (ICP) algorithm [18–20] to find its alignment. It minimizes the distance between two point cloudswith the requirement that the two point clouds are roughly close to each other (the initial guess), until they are aligned. The ICP algorithm consists of the following phases: • Selection of point pairs, • Matching of these point pairs, • Rejectionofpoint pairs due to individual consideration, • Errormetric assignment, • Minimizing the error metric. With the ICPalgorithm, analignment canbeachievedwithin a few iterations. Now, the idea is, instead of matching the whole robot as a rigid body, to split the robot into its links and match themseparately (cf.Figure1) inaneye-to-handcomposition, such that the orientation of its joints can be estimated. In this case the use of markers can be omitted since the joint orientation can be calculated from the alignment of the links to each other, which makes this approach a versatile applicablemethod for industrial applications.Theknowledge of the robot’s kinematic chain gives the possibility of robot pose variation by well-defined joint orientations as well as the variation of the joint orientation duringmotion to correct the trajectory in case of work-flow disturbance. The goal of this approach is a visual servoing concept by depth sensing with a potential to collision protection and avoidance in a collaborative applicablemanner. This paper is organized as follows. In Section II, the applied method is described. The implementation of the robot’s link point cloud models is described in Section III. The description of the setup and the camera alignment is described in Section IV. In SectionV, the presentedwork is summarized and SectionVI concludes the paper. II. METHODS&APPROACH The goal of the presented approach is to track a manip- ulator with six Degrees-of-Freedom (DoF) by two RGB- D imaging systems for joint position perception and visual servoing.For themeasurementof the robot’s joint alignment, the cameras are placed in an eye-to-hand composition. This allows to capture the wholemanipulator from awider view and avoid occlusions. The depth sensing technology with the highest accuracy for positioning and object matching is derived by comparing two different camera technologies. Therefore, a structured light camera and a ToF camera is applied and tested. Before the pose of the robot can be estimated, the position of both cameras have to be extrinsically calibrated, to get a perfect aligned point cloud from both cameras. The camera calibration is carried out as a transformation of the camera coordinate system by its physical position relative to the robot’s base coordinate system. For amatching process of point clouds by an appropriate approximation method like ICP to receive the robot’s joint positions asmentioned in Section I, themodels of its links, generated fromComputer-AidedDesign (CAD)files, have to be prepared. This is done by aligning the linkmodels in the CADfiles in their initial position as shown in Figure 2a and 75