Discrimination between Urban Area and Vegetation in High- Resolution Images Using Markov Random Fields (mrf)

The automatic extraction of traffic data from digital images is a challenging task and implies an extensive pre-processing of this data. An important information which simplifies the further extraction of traffic objects is the recognition of active traffic regions. A possible approach is masking of streets using platform attitude and position together with a digital street map (i). But often such digital maps are not available or at least inaccurate (ii).To improve differentiation between streets and environment, image based methods can be used. The aim of this work is the detection of active traffic regions (e.g. discrimination of urban area and vegetation) in highresolution aerial photographs with image processing methods. The segmentation of images was realised on the basis of texture features. There are various texture features which are suggested for segmentation in published literature. But unfortunately most of these approaches are barely applicable for the segmentation of real digital photos. In this paper Markov Random Fields (MRF) for feature determination are used, because MRFs enable the modelling of texture specific interactions between pixels. From the different existing MRF models the autobinomial model was chosen, which is known in image processing since the 80ies (iii). For obtaining reliable results the parameterization of the MRFs with synthesised images was investigated. Afterwards different model parameters, that influence the segmentation, were studied using standardised Brodatz textures (iv). Thus the capabilities and limits of this method can be characterised. It was possible to segment precisely the collages of Brodatz textures with MRFs. However, the quality of segmentation depends significantly on some additional factors that have to be considered, e.g. noise and image normalisation. Concerning these parameters the realised segmentation of aerial photos is likewise good. Nevertheless, the segmentation suffers from radiometric poor photographs. Moreover heterogeneous areas, e.g. urban areas with lot of single trees, make segmentation incorrect.