Reconstructing 3D Scenes from UAV Images Using a Structure-from-Motion Pipeline

This paper aims to apply grid-based motion statistics strategy for 3D reconstruction and promote the reconstruction results. Hence, a 3D reconstruction system that utilizes normal collections of unmanned aerial vehicle images using a Structure-from-Motion pipeline is provided. A typical incremental SfM is performed in this paper. It starts from an initial two-view reconstruction (the seed) that is iteratively extended by adding new views and 3D points, using pose estimation and triangulation. Later on, Bundle Adjustment (BA) is performed to minimize the accumulated error (drift). It is shown that reconstruction results have been improved and grid-based motion statistics strategy significantly improve the completeness and accuracy by mitigating drift effects. In addition, to evaluate our approach without ground truth, several different measures have been estimated. To assess the result of feature correspondence estimation and its effect on the SfM reconstruction result, this paper has measured the residual of the robust estimation and the root mean square error of the residuals of the SfM scene. While the incremental system has many advantages in robustness and accuracy, the efficiency remains its crucial challenge. This remains a problem to resolved in future works.