ATASR: Air-to-Air Scene Reconstruction

The widespread deployment of uncrewed aerial vehicles (UAVs) has raised serious public-safety concerns due to the increasing prevalence of illegal UAV operations. This article presents an air-to-air scene-reconstruction framework that simultaneously reconstructs the target UAV, its trajectory, and the ground background from air-to-air tracking image sequences. Air-to-air imagery poses significant challenges for pose registration because the target UAV typically appears small and exhibits weak texture. To address this issue, we design an iterative UAV-reconstruction pipeline that integrates cyclic image-pose estimation with progressive Gaussian-field optimization. The Gaussian field renders target color and depth images, which are used together with an end-to-end image matcher and a perspective-n-point (PnP) solver to estimate the poses of target-focused frames. The registered image set is then employed to optimize the Gaussian field, and pose estimation and scene reconstruction are performed in an alternating manner. To mitigate image degradation caused by high-speed rotor-blade motion, we train a new Gaussian field with blade masks and generate the blade disks by clustering and fitting the reconstructed sparse Gaussian point cloud. Finally, background is reconstructed and fused with the target model and trajectory to obtain the complete scene. Experiments demonstrate the effectiveness of the proposed method for both static and slow-moving targets. Code and data will be released upon acceptance.