Robust alignment for UAV images based on adaptive adjustment

Abstract

Abstract—In this paper, we investigate thoroughly the problem of aligning sequential images taken from low-altitude unmanned aerial vehicles (UAV). It is difficult because the photographic scene is no longer an approximate plane from the UAVs at low flight altitude. On this occasion, we assume the ground regions of the scene share the dominant plane, so sequential images are aligned based on the registration of the ground as a landmark, which precludes the obstruct of other objects of various altitudes. Specially, considering the occasional accumulating error and deformation degree, an adaptive bundle adjustment based on the updated reference image is creatively designed to guarantee the quality of incremental alignment in an efficient way. Finally, the global refinement on transformation models is performed under the constraint of no perspective distortion engendered, to improve the aligning accuracy further. Experimental results on several challenging datasets illustrate the superiority of our approach.