High accurate 3D reconstruction method using binocular stereo based on multiple constraints

In order to make an improvement of accuracy, completeness and robustness of the image based 3D reconstruction algorithm, we present a stereo matching approach under the condition of multi-resolution combined with multi-constraints to compute the 3D model of an object in this paper. In the first step, we build the image pyramid and stereo matching happens from the top level to the lowest level. In each level, we execute the following procedures: NCC matching, multi-constraints checking, re-matching and weighted median filter, disparity refinement. At last, we acquire an accurate and smooth disparity map in the bottom level. Combine with the calibration parameters, we can compute the 3D point cloud and use the Possion reconstruction method to generate the mesh model. The contributions of this paper are the proposal of choosing the pixel in the eight-neighborhood which has the highest match cost to guide the matching of center pixel during re-match, and design an anisotropic smoothing kernel to smooth the depth map and keep the detail at the same time in the disparity refinement. Experimental results show that the algorithm has the advantages of high precision, strong robustness and good model integrity.