ABSTRACT
Infrared (IR) images depict thermal radiation of physical objects. Imaging the building fašades and the roofs with an IR camera, thermal inspections of the buildings can be carried out. In such inspections a spatial correspondence between IR-images and the existing 3D building models can be helpful. Texturing 3D building models with IR images this spatial correspondence can be created. Furthermore in textures heat leakages can be detected and the heat bridges can be stored together with 3D building data. However, before extracting leakages, the windows should be located. In IR images glass reflects the surrounding and shows false results for the temperature measurements. Consequently, the windows should be detected in IR images and excluded for the inspection. The most common algorithms for window detection were developed for the images in the visual band. In this paper, an algorithm for window detection in textures extracted from terrestrial IR images is proposed. In the first step, small objects have to be removed by scaling down the image (texture). Then in the scaled image, regions are detected using a local dynamic threshold. Morphological operations are used to remove false detections and unify substructures of the windows. For every extracted region, which is a candidate for a window, the center of gravity is calculated. It is assumed that windows on fašades are ordered in regular rows and columns. First the points are grouped into rows using histogram of height created from extracted gravity centers. Then masked correlation is used to detect the position and size of the windows. Finally, the gaps in the grid of windows are completed. For the first experiments we use a dataset from densely build urban area captured in Munich, Germany. The IR image sequences were taken from a vehicle driving on the street around the test area. Camera was directed to the building in oblique view. According to the acquisition geometry, no fašade could be completely seen in one frame. Therefore, we combine the textures from many frames. To these textures we applied our algorithm for window detection. First results are promising. Applying the method for our test dataset, 79% completeness and 80% correctness could be achieved.