Evaluating Influence of Nonlinear Disturbances on Image Registration Based on Virtual Forces

In case of a disaster it is of utmost importance to obtain an overview of the affected area in order to coordinate rescue operations. Recent research and development has led to affordable unmanned aerial vehicles (UAVs) and enabled (semi-)autonomous surveillance and mapping of a disaster areas by UAVs. Mapping an area from many images incorporates image registration. Commonly used registration techniques expect static images to be registered. However, while mapping a disaster area, it is very likely that objects like vehicles or persons are moving on the ground, thus leading to dynamic scenes. We formerly introduced a fast and robust approach to register large amounts of successively arriving images, e.g., from UAVs. The approach only uses rigid-body transformations, but the usage of virtual forces enables the registration to tolerate small perspective distortions. This paper investigates the performance of our approach when applied to dynamic scenes, which are represented by nonlinear disturbances within image correspondences. We show that our approach not only tolerates small perspective distortions but also is able to register dynamic scenes.