A Cascaded Classifier Approach to Window Detection in Facade Images

A major part of recent developments in civil engineering in the urban context evolved around building and city models. Especially for a precise risk assessment of damages to existing buildings induced by ground movements, accurate models are inevitable. Beside the shape of a building, the focus is also on components compromising a building’s stiffness. Particularly, by including wall openings such as windows into risk analyses, these can be improved to provide more reliable predictions. However, most publicly available data sources only provide simple block models of existing buildings sometimes extended by roof shapes. As a consequence, any information concerning the windows of a building must be integrated into the model using other data sources. Whereas numerous approaches address the refinement of building shapes, their windows and other components are commonly disregarded. Although cascaded classifiers already turned out to yield good results in general and applying them to window detection seems promising, such approaches are yet insufficient to reliably extend building models. Drawing on previous findings, we present an approach to window detection in facade images satisfying the needs of risk assessment analyses. Our detection system combines a soft cascaded classifier consisting of thresholded Haar-like features with a sliding window detector extracting image patches for classification. The soft cascaded design improves the detection rate over previously made approaches while coincidentally reducing the amount of required features. Further, we evaluate the effect of a rectified dataset on the classification results compared to its counterpart with images taken from varying angles.