Page - 10 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure2.Poseestimatesusing theOpenCVsolvePnPalgorithm, for imagescapturedatdescribedsix locations. Different colours are for different methods with black being ground truth locations. Different symbols mark the six different locations. Unitsare incentimeters. The6coordinatesystemsgive theposeof the robot. Theaccuracy inestimatingy and errors inxare similar as with the previous experiment. The point symmetry about the origin for the iterative algorithm is visible, having incorrectlyflippedall three axis signs. Variance stays largely the sameevenwith largerz. left half are not simply mirrored across the z-axis. The distance gets consistently underestimated yet it seemswith largerabsolutevalueofx inground-truth theabsoluteestimatesinxalsoseemtoincrease. The estimatesforthelocationKleft farbreakthispattern, beingveryclose to theestimates forKamcl, the loca- tionwhere thevisionbasednavigationissupposedto takeoverafterusingamcl localization. It canalsobe observed that variance only slightly increases about the estimates in zwith increasing z distance. Clus- ters are very compact, an improvement compared to the first experiment. This can be attributed to using more precise intrinsic camera parameters. It is also visible that all algorithmsareaccurate forestimating the smalloffset iny. Unfortunately, the reason for this seemingly sys- 10