Abstract: In recent years, it is possible to perform real-time 3D reconstruction with the development of the GPU technology and maturity of parallel algorithms. This paper presents an interactive dense 3D reconstruction system for small scenes, which can accurately estimate the real-time position of the camera by means of advanced mobile tracking technology. An improved multi-view depth generation algorithm is proposed, which can calculate scene depth in real-time under GPU acceleration. The cumulative cost of sub-pixel semi-global matching in the improved algorithm improves the accuracy of multi-view stereo matching, and combines the global optimization method to calculate the accurate scene depth information. Depth map is converted into distance field, and real-time depth fusion is realized by using globally optimized histogram compression fusion algorithm and parallel primal-dual algorithm. The experimental results prove the feasibility of the reconstruction system and the correctness of the reconstruction algorithm.

Key words: reconstruction, real-time, parallel algorithms, depth, fusion