Image Matching and Outlier Removal For large Scale DSM Generation

Digital surface models can be efficiently generated with automatic image matching from optical stereo images. Detailed reconstruction from very high resolution stereo images requires the use of dense stereo matching algorithms, such as Semiglobal Matching. If only a single stereo pair is available, the results contain a small amount of outliers, mostly due to changing reflectance behaviour, clouds, water and moving objects. This is especially critical for large scale and fully automated DSM generation with dedicated stereo satellites, such as Cartosat-1 and ALOS/PRISM. An effective outlier removal algorithm for dense stereo matching algorithms, based on region based consistency checks between to matching results with slightly different parameters is presented in this paper. Ground truth has been manually extracted on 66 test areas, covering different land cover types, seasons and elevation heights. On this diverse and demanding dataset more than 99\% of all outliers are removed automatically, with minimal influence on the correctly matched regions.}