Enhancment of dense urban digital surface models from VHR optical satellite stereo data by pre-segmentation and object detection

Kurzfassung

The generation of digital surface models (DSM) of urban areas from very high resolution (VHR) stereo satellite imagery requires advanced methods. In the classical approach of DSM generation from stereo satellite imagery, interest points are extracted and correlated between the stereo mates using an area based matching followed by a least-squares sub-pixel refinement step. After a region growing the 3D point list is triangulated to the resulting DSM. In urban areas this approach fails due to the size of the correlation window, which smoothes out the usual steep edges of buildings. Also missing correlations as for partly – in one or both of the images – occluded areas will simply be interpolated in the triangulation step. So an urban DSM generated with the classical approach results in a very smooth DSM with missing steep walls, narrow streets and courtyards. To overcome these problems algorithms from computer vision are introduced and adopted to satellite imagery. These algorithms do not work using local optimisation like the area-based matching but try to optimize a (semi-)global cost function. Analysis shows that dynamic programming approaches based on epipolar images like dynamic line warping or semiglobal matching yield the best results according to accuracy and processing time. These algorithms can also detect occlusions – areas not visible in one or both of the stereo images. Beside these also the time and memory consuming step of handling and triangulating large point lists can be omitted due to the direct operation on epipolar images and direct generation of a so called disparity image fitting exactly on the first of the stereo images. This disparity image – representing already a sort of a dense DSM – contains the distances measured in pixels in the epipolar direction (or a no-data value for a detected occlusion) for each pixel in the image. Despite the global optimization of the cost function many outliers, mismatches and erroneously detected occlusions remain, especially if only one stereo pair is available. To enhance these dense DSM – the disparity image – a pre-segmentation approach is presented in this paper. Since the disparity image is fitting exactly on the first of the two stereo partners (beforehand transformed to epipolar geometry) a direct correlation between image pixels and derived heights (the disparities) exist. This feature of the disparity image is exploited to integrate additional knowledge from the image into the DSM. This is done by segmenting the stereo image, transferring the segmentation information to the DSM and performing a statistical analysis on each of the created DSM segments. Based on this analysis and spectral information a coarse object detection and classification can be performed and in turn the DSM can be enhanced. After the description of the proposed method some results are shown and discussed.