Fast and accurate unknown object segmentation for robotic systems

Abstract

Object segmentation is the first step towards more advanced robotic behaviors, as robots need to localize objects before attempting tasks such as grasping or manipulation. A robot vision system should be able to provide accurate object hypotheses in reasonably high frame rates, using images and possibly also depth data. This work proposes a fixation-based object segmentation algorithm able to cope with unknown objects, and run on a real-time robot. We show that a balanced combination of moderately accurate, when considered independently, but at the same time computationally inexpensive building modules can yield remarkable results both in terms of accuracy, but also of execution speed. We describe our algorithm and present both qualitative and quantitative experimental results that indicate significant speed-up over the state-of-the-art. © 2013 Springer-Verlag.