Gradient based splitting of blocks of houses in high-resolution ortho-images

The described building block segmentation is one module in an approach to automate object extraction from high resolution image data. The level of detail in high resolution (sub-meter scale) images is adequate to be able to discriminate directly adjacent houses visually. Thus it seems obvious to automate not only image segmentation at the scale of building blocks but also at the scale of single houses. Based on a segmentation into single blocks of buildings, a further split into single house units is aspired. Skeletal pixels serve as the starting points for setting up split line candidates. The gradient orientation is used to determine the split line orientation. Thus, the split line candidates, spanning from one building segment border to another, are defined. Among all the candidates, the best ones are selected by introducing a set of quality measures. The segments resulting from these split lines are accepted, if they have a predefined minimum extension.