Page - 13 - in Joint Austrian Computer Vision and Robotics Workshop 2020

AutonomousGraspingofKnownObjectsUsingDepthDataandthePCA DominikSteigl,MohamedAburaia,WilfriedWo¨ber UASTechnikumWien {mr18m007,mohamed.aburaia,wilfried.woeber}@technikum-wien.at Abstract.Two main goals for automated object ma- nipulation processes are cost reduction and flexibil- ity. Time-consuming, costly object-specific fixtures can be replaced by vision systems, whereby the ma- nipulators are extended with cameras so that multi- ple objects in the environment can be precisely iden- tified. To be able to manipulate an object, it must be recognized first in the world, and then the pose must becalculated. Neuralnetworkapproachesrecognize and estimate the pose of an object in a single step and yield superior results, but rely on vast amounts of training data. This work describes an approach for estimating the pose of identified objects without pre-trained pose data. Template matching is used to recognize objects in depth images, and the pose is estimated through principal component analysis (PCA). The input to the algorithm is reduced to the template. Pre-existing knowledge about the object further improves accuracy. A maximum deviation of 0.2 cm from the ground truth has been achieved, which suffices for the industrial grasping task. The system was evaluated with real measurements taken with an RGB-D camera. This work resembles a first step to estimate an object’s pose with linear statisti- calmethods. 1. INTRODUCTION Industrial robotsareefficientatpickingupobjects in a predefined, structured environment [10]. When mobile manipulators are deployed in a factory set- ting and costly fixtures have to be avoided, robots need theability to identifyand locateobjects forma- nipulation. To overcome this problem, a vision sys- tem can be used. One way to give robot vision is to use two-dimensional imageswithdepth information, alsoknownas2.5DimagesorRGB-Dimages. RGB- D images can be used to find and localize objects by analyzing the environment. Building on top of the recognized and classified object, pose estimation tries to estimate the six degrees of freedom (DOF) pose of an object in an image. For mobile manipu- lation of objects this information is needed to accu- ratelygraspobjectswithamanipulator in thecorrect positionandorientation. The current state of the art approaches towards object recognition and pose estimation are based on deep neural networks [15]. They usually outperform human crafted features [19], but unfortunately they rely on huge amounts of training data for classifica- tionandposeestimationandaredifficult toadapt[9]. This iswhy, in thiswork,amoretraditionalapproach was chosen. The target object is recognized using template matching in a 3D space. Pose estimation is implementedusing theprincipal componentanalysis (PCA) to place an orthogonal basis in the center of thegrabbingarea. UsingPCAtoestimate theposeof the object, the needed input to the algorithm can be reduced to only the template. This work resembles a first step to estimate an object’s pose with linear statisticalmethods. In the following chapters the related work is sum- marized, the used methods are explained and the re- sults arebeingdiscussed. 2.RELATEDWORK Object recognition describes the task of localiz- ing known objects in images. Due to changes in the viewpoint or lighting, the task of mapping the huge amountof inputpixels toasmalloutput space is still complex [16]. To mitigate the influence of lighting conditions, approaches which rely on 3D informa- tion were researched [8]. The data used in these ap- proaches is usually made up of a three channel 8-bit RBG image or an additional fourth channel which represents the 3D distance of the object to the image sensor,whereeach image isdescribedusing 13