Accurate and Robust Scene Reconstruction in the Presence of Misassociated Features for Aerial Sensing

Abstract

Georeferencing through aerial imagery has several applications, including remote sensing, real-time situational mission awareness, environmental monitoring, rescue and relief, map generation, autonomous hazard avoidance, and landing and navigation of unmanned aerial vehicles (UAVs). In aerial imagery, structure from motion (SfM) is often used for three-dimensional (3D) point reconstruction (i.e., ground locations) and for camera pose estimation (i.e., airborne position and orientation) from a set of geometrically matched features between two-dimensional (2D) images. In order to improve the correspondence estimation, outliers (defined as gross misassociated features) are determined and excluded using an outlier rejection algorithm such as random sample consensus (RANSAC); however, in practice, it is impossible to ensure that outlier rejection algorithms always lead to perfect feature correspondences. Consequently, there remain misassociations between matched features in different views that significantly lower the accuracy and robustness of 3D scene reconstruction algorithms. This paper introduces an adaptive resection-intersection bundle adjustment approach that refines the 3D points and camera poses separately after the gross misassociations are removed by an outlier rejection algorithm. For each iteration, the proposed approach identifies the potential misassociated features independently in the resection as well as the intersection stage, where these potential outliers, contrary to previous studies, are reexamined at later iterations. Compared to state-of-the-art algorithms, the proposed approach leads to smaller 3D reconstruction errors by rejecting almost all of the misassociations as outliers while the maximum number of inlier matched features are retained. The results from several numerical simulations and real data sets are presented, and it is shown that the proposed approach outperforms existing bundle adjustment (BA) approaches in the presence of misassociated features although the convergence rate is slower.