Efficient visual localization of landmarks in DLOs under robotic manipulation: A statistical approach

In this paper we present a novel paradigm for reduction of the search space for visual localization of landmarks in deformable linear objects (DLOs) for robotic manipulation tasks. This has been achieved by statistical prediction of the location of the landmarks with a multivariate multiple linear regression model, using position and orientation of the end-effectors of the robot as independent variables and the x-y coordinates of the predicted locations as dependent variables. We implemented this approach with a NAO humanoid robot and a Kinect RGB camera, the results denote a reduction in almost 90% of the search space of the position of 6 landmarks in a deformable linear object with respect to the classical methods in which the search is performed in the whole image.