Technology for fast 3d-scene reconstruction from stereo images

We propose a fast algorithm for disparity maps construction from stereo images. It is known that the main problem in 3D-reconstruction is to find the corresponding points on different views of the scene. The search area greatly extends when shifts and scale differences in stereo images are great. To improve the performance we offer to use initial image matching by using an affine transform. The reliability and efficiency of image matching in subsequent steps is achieved by using epipolar constraints and an image pyramid. The developed method was implemented on a parallel computing platform CUDA. The results of experimental studies show high performance of the proposed approach, while maintaining the high-quality reconstruction of 3D-scenes.