AeroDGS: Physically Consistent Dynamic Gaussian Splatting for Single-Sequence Aerial 4D Reconstruction

Researchers Hanyang Liu and Rongjun Qin have introduced AeroDGS, a novel framework for 4D dynamic scene reconstruction from single-view aerial sequences, accepted for presentation at CVPR 2026. The system addresses the critical challenge of monocular aerial reconstruction, which has remained inherently ill-posed due to severe depth ambiguity and unstable motion estimation when capturing wide spatial ranges with dynamic objects. Traditional approaches struggle with the limited spatial footprint and large motion disparity typical in drone footage, but AeroDGS leverages Gaussian splatting—a recent advancement in neural rendering—combined with physical constraints to produce stable, coherent reconstructions from challenging UAV videos.

The technical breakthrough lies in two core modules: a Monocular Geometry Lifting component that reconstructs reliable static and dynamic geometry from a single sequence, and a Physics-Guided Optimization module that incorporates differentiable priors for ground-support, upright-stability, and trajectory-smoothness. These physical constraints transform ambiguous 2D image cues into physically consistent 3D motion, enabling joint refinement of both static backgrounds and dynamic entities. The researchers validated their method on a newly created real-world UAV dataset spanning various altitudes and motion conditions, demonstrating superior reconstruction fidelity compared to existing state-of-the-art techniques. This advancement opens new possibilities for applications in urban planning, disaster response, and autonomous systems that require accurate spatiotemporal understanding from limited aerial footage.