1-Point Rigid Motion Estimation and Segmentation with a RGB-D Camera

RGB-D cameras like Microsoft Kinect that provide color and dense depth images have now become commonplace. We consider the problem of estimation and segmentation of multiple rigid body motions observed by such a camera. On the basis of differential geometry of surfaces and image gradients, we present a method for completely estimating the Euclidean transformation of a rigid body by using just a single surface point correspondence. This is facilitated by two methods of removing the sign ambiguity of principal curvature directions which is the main contribution of the paper. Further, we apply state-of-the-art rotation/translation averaging techniques to achieve refined Euclidean transformation estimates and segmentation. Results using both synthetic and real RGB-D data show the validity of our approach.