Uncertainty-Aware UAV Trajectory Adaptation for Object-Centric Active 3D Reconstruction

Active 3D reconstruction with small uncrewed aerial vehicles (UAVs) is constrained by strict limits on battery life, flight time, and computational payload. Under a limited observation budget, an evolving reconstruction model can maximize quality by guiding decisions made during the mission. Our acquisition strategy addresses image validation, object-centric modeling, and uncertainty-driven viewpoint selection jointly. After a short initialization flight, acquired images pass through region-based geometric validation. A masked 3D Gaussian Splatting (3DGS) representation of the target object is then fitted on the surviving views via a tethered base station. Its parameter-space uncertainty, derived from the Fisher information of the Gaussian parameters, scores candidate viewpoints by their expected contribution, selecting the most informative poses as the UAV’s next waypoints. Evaluated on a real small-UAV testbed, our approach demonstrates that restricting 3DGS to the target accelerates fitting, while validation improves reconstruction metrics despite using fewer images. Moreover, model-driven viewpoint selection consistently outperforms uniform acquisition across varying budgets and localization modalities, with D-optimality emerging as the most stable criterion.