Most traditional 3D terrain reconstruction algorithms cannot represent the accurate structure of the terrain and are time consuming as well, thus the technological development is seriously hindered . In order to realize the 3D terrain reconstruction accurately by using pictures taken by unmanned aerial vehicle (UAV) with the advantages of high resolution, wide camera range and low demand of shot environment, we propose a method that generates digital elevation model (DEM) data respectively in different superpixel terrain areas by segmenting the images at the preprocessing stage. Firstly, the simple linear iterative clustering (SLIC) algorithm, which shows good performance in superpixel generation and is convenient to use, is adopted to segment the images into multiple superpixel terrain areas which contain just a single terrain type. Then the adjacent superpixel areas containing the same terrain are fused according to the LAB color information, in which way the number of superpixel areas is decreased and the speed of the algorithm is improved. Thirdly the scaleinvariant feature transform (SIFT) feature points’ extraction and matching are done in each area. Based on the matching results, the 3D coordinates are computed with the method of binocular stereo vision and DEM data are generated in the end to reconstruct the terrain. Comparing with the satellite map, the proposed algorithm can reconstruct 3D terrains effectively, and it can present the boundaries information accurately in contrast with traditional 3D terrain algorithms.
