Sensorimotor Coordination in a Humanoid Robot: Building Intelligence on the iCub - Bioinspired Approaches for Human-Centric Technologies - page 190

Fig. 6.16 Objects used for testing with their respective success rates over 20 trials
(reaching and grasping). One appealing visual cue is motion and we have been
recently able to devise a method which provides motion segmentation independent
from the movement of the cameras. Another important source of information is 3D
vision which enables the extraction of partial object shape which in turn is useful for
the control of grasping. For 3D vision we used a standard method and we refer the
reader to the paper by Hirschmuller ( 2008 ). The estimation of visual motion is
instead described next.
6.7.1 Optical Flow
Our method is based on the analysis of failures of the standard Lucas-Kanade
algorithm ( 1981 ). As a general rule, in order to verify that the instant velocity v of a
point p has been correctly estimated, the patch W around that point in the image I t is
compared to the patch of the same size at p+v in the new image I t+1 (where the