Seite - 17 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure6.Visualizationof the results with the objectbeingplacedatdifferinganglesanddistances on theoptical axis Figure7.Disparity imageof theobjectviewedbytheIntel RealSense camera. Occlusions induced by the stereo vi- sionsystemmakeitdifficult toaccuratelylocatethegrasp- ingareadepicted inFigure 1. theobject isbeingrotated. Figure7alsoshowsthat it isdifficult todepict thegrabbingareaoftheobjectfor creating a fitting template from the disparity image. Having a poorly chosen template leads to points be- ing incorporated into thePCAestimation thatarenot actually part of the grabbing area and therefore lead to unexpected results. Nevertheless, it can be con- cluded that the anomalies are not induced by means of the methodused forestimating thepose. Figure6showstheresults formeasurements taken withtheobjectbeingplacedatdifferentanglesonthe optical axis. Refer to Figure 3 for a visual presenta- tion of this process. The angle depicted in Figure 6 corresponds to angleα shown in Figure 3. The dis- tance mark relates to the distances shown on the X- axis labels of the graphs in Figure 6. The anomalies again can be explained by the problems mentioned before. Therightgraph inFigure6clearlyshows the limits of the proposed approach, as the structure of Table2.Regions inwhichthealgorithmyieldsresultssuf- ficientenoughfor the industrialgraspingusecaseathand. Angle [°] Distance [cm] +/-0 30 -75+/-10 +/-20 +/-10aroundcamera 30 -75+/-20aroundcamera +/-30aroundcamera theboxplotsover thegraphchanges in respect to the other twographs. Figure8showstheresults formeasurements taken with the object being placed at different angles around the camera. The angle depicted in Figure 8 corresponds to angle β shown in Figure 3. The anomalies again can be explained by the problems mentionedbefore. The results showthat theusable region for this al- gorithmcanbesummarizedwithTable2. Argument- ing, that theanomaliescanbeeliminatedbyusingthe enhancements already listed. Angles depicted with the ”around camera” suffix correspond to the object being rotatedaround thecamerawithangleα, asde- picted inFigure3. 5.Conclusion This work presented an approach to estimate the pose of a known object by using the principal com- ponent analysis. This resembles a first step to esti- mateanobject’sposewith linear statisticalmethods. The results showed that theapproach is sufficient for the industrial usecaseathand, sinceamaximumde- viation of 0.2 cm compared to the ground truth is achieved, when anomalies are ignored. The results also show the limitations of this approach. Anoma- 17