A Robust Targets Matching Algorithm in 3D Reconstruction from Multi-views

To find the corresponding target points among multi-images of the same scene is a premise of three dimensional (3D) reconstruction automatically from multi-view images. The geometrical constraints among multi-images of the same scene are analyzed and it is concluded that the error matching cannot be removed if only using 2D information. In this paper, we present an iterative fractional step matching algorithm: (1) pre-matching the target point using constraints of epipolar geometry firstly; (2) removing the error matching points using the principle of the uniqueness of a point in 3D space; (3) bundle adjustment is added to refine the camera external parameters. After the external parameters of every image refined, repeat step (1) to step (3) until matching occurred at the same point between the adjacent iteration. Real experiments showed that error matching can be effectively removed using our algorithm, resulting in improved efficiency and accuracy in 3D reconstruction.