Considering the stereo model registration in large-scale 3D terrain reconstruction from aerial imagery

we propose an approach for registering all stereo models to realize automatic large-scale 3D terrain reconstruction. Individual stereo models were constructed from the adjacent images. The tie points of adjacent models were extracted according to the pixel coordinates of the corresponding points based on feature matching in the shared image. All models were linked and registered through loop traversal search. Subsequently

the large area terrain of the overall aerial photography area was produced through triangulation

Tin generation

differential correction

elevation interpolation

and color blending. The above processes were sequentially and automatically performed. Two groups of data were applied in the experiment. Results showed that the root-mean-square (RMS) registration errors of all 3D models of the two data sets were 5.20 and 2.63 pixels. The relative accuracy of the produced large-scale terrain was high. The absolute orientation was determined based on the results generated from the second data set with ground control points

and the absolute positioning accuracy was evaluated using checkpoints. The RMS errors of the check points in the planimetry and altitude were 0.326 and 0.502 m

respectively

illustrating a high absolute accuracy. Couclusion: Large-scale 3D terrain that was organized in a special format was automatically produced by incorporating a series of overlapped aerial imagery. The visualized and realistic 3D terrain validated the feasibility and effectiveness of the proposed solution.