The Study On Key Technologies Of Sparse Depth Map Matching

The 3D data acquisition is very important for 3D Geographic Information System (3DGIS). Image matching based on Photogrammetry and multiple view geometry has become one of the main means of 3D data acquisition which can obtain ground elevation and texture data. Dense depth map as the result of image matching contains 3D information of each image pixel. But the 3D data extracted from this way need to be simplified if used for 3D modeling in 3DGIS. On the other hand, if sparse texture regions exist in the images, it's difficult to overcome the error matches. Therefore, for the need of 3DGIS modeling, the author studies the key technologies of sparse depth map matching. The objective is to quickly generate sparse depth map automatically and avoid mismatching in texture-less regions.The main work is as follows:1) The imaging principle of binocular stereo vision, the basic theory and solution method of epipolar image generation are introduced. Then the image matching constraints, difficulties and evaluation means are analyzed.2) The author proposes an extraction algorithm of feature regions based on normal vector. The method judges whether the pixel is feature point or not according to the change degree of normal vectors in pixel neighborhood. It effectively extracts the regions in which the gray scale changes greatly and filters the smooth areas with no remarkable feature.3) After comparing and analyzing the matching algorithms in current reference, the author presents a sparse depth map matching algorithm based on feature regions. The algorithm takes one image as a continuous 3D surface, and mainly matches feature regions. The summary of cosine values of corresponding normal vectors in two match windows are adopted as the cost functions. The algorithm based on feature regions can filter sparse texture regions to avoid mismatch and save computer time.4) C++ language is used to implement this algorithm, and the real image data is used for checking it. The results show that the proposed algorithm is effective to obtain the sparse depth maps which can satisfy 3DGIS modeling.