Dynamic segmentation of small image windows for visual servoing

Abstract

Processing of image windows rather than complete images is useful for robots incorporating visual servoing as high image processing rates are required. Each image window is segmented, often by thresholding, to identify features of interest. To adapt to changing conditions and to achieve the thresholding of low contrast and shadowed windows, a sophisticated method for performing dynamic segmentation is required. Segmentation methods from the pattern recognition and optical character recognition fields were studied to determine their effectiveness at thresholding 32 by 32 pixel image windows of circular hole features. The segmentation technique must be capable of preserving the centroid location with sub-pixel accuracy. To this end a new morphological preprocessing method is introduced to improve the performance of most thresholding algorithms. It was found that this new preprocessing method was able to improve the centroid location error by nearly 40% when Yasuda's thresholding algorithm was used. The preprocessing algorithm in combination with Yasuda's thresholding algorithm was able to segment the holes with an average centroid location error of 0.423 pixels and a standard deviation of 0.328 pixels.