Reliable automatic three-dimensional surface extraction incorporating image matching and object recognition

A stereo image matching algorithm based on feature points, multi-scale regions, areas, and recognized objects, such as buildings and roads was implemented. The algorithm includes several key parts: (1) Feature points extraction and reliable point matching; (2) Image segmentation and vectorization; (3) Object recognition/classification; (4) Matching using multi-scale regions, area, and recognized objects; (5) Bare earth surface reduction; (6) Manual editing of the automated matching result.;Some independent preprocessing steps were conducted before the matching. In one step, the zero-crossing images were produced, and the zero-crossing points were matched as feature points to extract reliable matching points. In another step, the images were segmented, and shape analysis was done to the segmented regions to extract man-made objects. The regions were then matched by a region based, multi-scale cross-correlation and least square matching to find the disparity plane for the regions. In the multi-scale matching, the large and prominent objects of large scale, mainly fields, buildings and roads were matched first, then smaller objects and regions of smaller scale. If a region contains reliable matching points, those points were used to estimate the initial disparity plane for the region. At this stage, the matching result is still sparse. A reliable growth matching starting from the reliable points was then applied to densify the matching result. The result from the above matching is a DSM (Digital Surface Model), i.e., elevations of the tops of the objects. It was reduced to a DEM (Digital Elevation Model) using the plane-cut non-ground point filtering algorithm. As the generated DEM was not perfect, manual editing was done to fix any remaining problems.;The contributions of this research are: (1) innovatively incorporate the image object recognition into matching process; (2) a new algorithm was developed to extract reliable matching points; (3) a novel way to reliably match segmented regions from larger scale to smaller scale was implemented; (4) the matching result were densified by growth matching; (5) a stereo 3D point editing system was implemented.;From the experimental results, it can be concluded that the algorithms implemented generates DEMs with higher accuracy than area-based matching alone.