Defining Aerial Camera Pose and Solving Scale Ambiguities for Features’ Positioning and Dimensioning

This paper proposes a novel method for determining aerial camera pose and resolving scale ambiguities in aerial imagery when GNSS data is not sufficiently accurate. By relying on a monocular Perspective-n-Point (PnP), the method enhances the identification of matching points between real images and synthetic renderings of 3D models. This can be achieved by recreating a camera view in a 3D graphics environment that closely approximates the available intrinsic and extrinsic parameters of the real image. Additionally, it is proposed to limit to two the number of required points by employing Pitch and Roll angles (Euler) in the calculation. This may further enhance the likelihood of finding a solution to the PnP problem, even in extremely complex environments. The 3D graphics environment allows managing both 3D models and camera view in the same environment to solve the scale ambiguity. The method was tested using 24 aerial images from a DJI Inspire PRO drone in combination with a high-resolution LiDAR-derived 3D model of a building. The results demonstrated the capability to achieve sub-meter accuracy in terms of target positioning and centimetric-level accuracy in terms of target dimensioning. As a future development, Google Tiles could serve as a source of reference 3D models, expanding the method’s applicability to numerous locations. Moreover, inertial navigation systems (INS) could provide initial reference parameters for generating the synthetic camera view, enabling the method to be used in GNSS-denied or unreliable scenarios.