Visual Servoing with Moments of SIFT Features

Abstract

Diese Arbeit ist im Sonderforschungsbereich 531, "Computational Intelligence", der Universität Dortmund entstanden und wurde auf seine Veranlassung unter Verwendung der ihm von der Deutschen Forschungsgemeinschaft zur Verfügung gestellten Mittel gedruckt. Visual Servoing with Moments of SIFT FeaturesFrank Hoffmann, Thomas Nierobisch * , Torsten Seyffarth * and Günter Rudolph † * Chair for Control System Engineering/Electrical Engineering and Information Technology/University of Dortmund, Germany {frank.hoffmann, thomas.nierobisch, torsten.seyffarth}@uni-dortmund.de † Chair of systems analysis/Department of Computer Science/University of Dortmund, Germany guenter.rudolph@cs.uni-dortmund.deAbstract-Robotic manipulation of daily-life objects is an essential requirement in service robotic applications. In that context image based visual servoing is a means to position the end-effector in order to manipulate objects of unknown pose. This contribution proposes a 6 DOF visual servoing scheme that relies on the pixel coordinates, scale and orientation of SIFT features. The control is based on geometric moments computed over an alterable set of redundant SIFT feature correspondences between the current and the reference view. The method is generic as it does not depend on a geometric object model but automatically extracts SIFT features from images of the object. The foundation of visual servoing on generic SIFT features renders the method robust with respect to loss of redundant features caused by occlusion or changes in view point. The moment based representation establishes an approximate one-to-one relationship between visual features and degrees of motion. This property is exploited in the design of a decoupled controller that demonstrates superior performance in terms of convergence and robustness compared with an inverse image Jacobian controller. Several experiments with a robotic arm equipped with a monocular eye-in-hand camera demonstrate that the approach is efficient and reliable.

Scale-only visual homing from an omnidirectional camera

“…Negre et al [24] showed that the intrinsic scale can be a measurement to the time to collision. [20] showed a direct relation between the scale and the distance to the feature point. However, for different setup and different environment, the absolute distances to the features cannot be mapped directly.…”

“…Some related early work using SIFT as main features for visual homing was proposed in [19] [20]. They considered the epipolar geometries as well as the orientation and scale of SIFT features for monocular cameras, following the similar framework proposed in [4].…”

“…We will then prove that the resulting control law converges to the Home position. [20] provided an indication that how the scale of visual features relate to the distance to the certain physical feature point. It stimulated us the possibility to perform homing via only scale information.…”

Towards Practical Visual Servoing in Robotics

“…Different type of features can be found in the literature. For example the trifocal tensor [22], the image and SIFT moments [23], [24], the cylindrical coordinates [25] and the spherical features [26], [27], [28]. Using the basic point and line features, the next section presents the key challenges related to imagebased task specification and a possible provable and intuitive specification interface.…”

Photometric moments: New promising candidates for visual servoing

“…In addition, we also obtain a large convergence domain. In [11], image moments computed over a set of SIFT keypoints are used for visual servoing. Naturally, this involves the robust extraction and matching of these keypoints at every frame.…”