Structure from motion for 3D object reconstruction based on local and global bundle adjustment

Abstract

Structure from motion approach allows to recover both the 3D structure of the scene and the camera motion. It uses a global bundle adjustment by minimizing a non-linear criterion (often through the use of the Levenberg-Marquardt algorithm) to adjust the various entities initially estimated, which requires a long calculation time and can converge to a local solution due to a bad initialization. In this paper, we used the Structure from Motion approach for 3D object reconstruction from images taken by a single camera moving around the object. The proposed approach is based on the combination of the local bundle adjustment (LBA) and the global bundle adjustment (GBA) which is very useful to ensure effective and rapid convergence to the optimal solution. Our reconstruction system is initialized from two calibrated images. After each insertion of a new uncalibrated Image, we integrate a LBA to adjust the new estimated parameters and avoid as much as possible the accumulation of errors which can affect the system's stability. After the insertion of the last image, a GBA is performed to adjust as best as possible all the estimated entities already refined locally (3D points and camera parameters). Experimental results show the reliability and rapidity of the proposed approach compared to the classical approach.