Figure-ground segmentation using multiple cues

The theme of this thesis isfigure-ground segmentation.We address the problem in thecontext of a visual observer, e.g. a mobile robot, movingaround in the world and capable of shifting its gaze to andfixating on objects in its environment. We are only consideringbottom-up processes, how the system can detect and segment outobjects because they stand out from their immediate backgroundin some feature dimension.Sinte that implies that thedistinguishing cues can not be predicted, but depend on thescene, the system must rely onmultiple cues. The integrated use of multiple cues formsa major theme of the thesis. In particular, we note that an observer in our real environment has access to 3-D cues. Inspiredby psychophysical findings about human vision we try todemonshate their effectiveness in figure-ground segmentationand grouping also in machine vision.An important aspect of the thesis is that the problemsare addressed from asystems perspective:it is the performance of the enbresystem that is important, not that of component algorithms.Hence, we regard the processes as part of perception actioncycles and investigate approaches that can be implemented forreal-time performance.The thesis begins with a general discussion on theproblem of figure-ground seg mentation and thereafter the issueof attention is discussed. Experiments showing someimplementations of attentional mechanisms with emphasis onreal-time performance am presented. We also provideexperimental results on closed-loop control of a head-eyesystem pursuing a movmg object. A system integrating mo tiondetection, segmentation based on motion and segmentation basedon stereo is also presented. Maintenance of an already achievedfigure-ground segmentation is discussed. We demonstrate how aninitially obtained figure-ground segmentation can be maintainedby switching to another cue when the initial one disappears.The use of multiple cues is exemplified by a metbod ofsegmenbng a 2-D histogram us ing a multi-scale approach. Thismethod is further simplified to suit our real-time performancerestrictions. Throughout the thesis the importante of havingsystems with a capacity of operating continuously on imagescoming directly from cameras is stressed, thus we prove thatour systems tonsist of a complete processing chain, with nolinks missing, which is essential when designing workingsystems. Keywords:computer vision, figure-ground segmentation,multiple cues, real-time systems, tracking, active vision,attention, image flow, phase-bases stereo, 2-DhistogrammingNR 2014080