Multi-View 3D Reconstruction Technology Research Based On Noise Estimation

With the increasing demand of 3D model in the field of film and television animation, virtual reality, heritage protection and other fields, more and more attention is focused on the technology of 3D model recovery from multiple images. Compared with the traditional laser scanning 3D modeling technology, multi-view 3D reconstruction method has the characteristics of reducing the equipment requirements, controlling costs and so on. But in the multi view 3D reconstruction, the image and the map between 2d image and 3D model space coordinates are subjected to the noise and the external point, which make the reconstruction process face many challenges.For the multi angle view feature, propose a multi view 3D reconstruction method based on noise estimation, which can be divided into four parts: image feature extraction, feature matching, sparse reconstruction, dense reconstruction. Firstly, the input images are extracted feature points by using Surf algorithm, and the KD tree is used to store the feature points. Then use the feature point matching method based on minimum spanning tree generating feature point trajectories. Because feature point trajectories generated by the previous step may contain wrong matching, use the residual consistency method accurately eliminate error matching points in feature trajectory. At the same time, triangulate the right matching points and bundle adjustment to solve the camera parameters and sparse 3D points cloud. Lastly, the sparse 3D point as seed node is used to reconstruct the dense 3D model and a complete 3D model is obtained.Compared with the traditional multi view feature points matching and sparse 3D reconstruction algorithm, the improved algorithm is validated. Compare the proposed multi view feature matching method based on minimum spanning tree with traditional matching algorithm in four indicators. The four indicators are needed for the number of images, the total ratio of successful matching images, the number of successful extracted features, the number of multi view matching. The experimental results show that the improved algorithm is 1/5 time of traditional algorithm. Next use consistent residual to remove wrong matched points in the feature track and use bundle adjustment method to make sparse reconstruction. The experimental results show the proposed method is obviously improved than the original algorithm.