A System for Multi-View Mapping of Dynamic Scenes Using Time-Synchronized UAVs

Abstract

Recent advances in 3D scene reconstruction, such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting, have demonstrated remarkable results in novel view synthesis and dynamic scene representation. Despite these successes, existing approaches rely on time-synchronized multi-view imagery captured using specialized camera rigs in controlled environments. This reliance limits their applicability in uncontrolled, unbounded dynamic scenes. In this work, we propose a novel Unmanned Aerial Vehicle (UAV) based multi-view capture system that leverages GNSS Pulse Per Second (PPS) signals for precise frame synchronization across multiple cameras. Our system eliminates the need for fixed infrastructure, enabling flexible and scalable data collection for dynamic scene reconstruction in diverse environments. In addition to the system architecture, we also introduce a dataset of synchronized multi-view images captured in unbounded outdoor scenes from four synchronized UAVs, each carrying a stereo camera rig. We benchmark several 3D and 4D representation methods on our dataset and highlight the challenges associated with data collection in unstructured outdoor settings such as sparse views, varied lighting conditions, visual degradation etc. Our hardware configuration details, software details and dataset is available at https://github.com/neufieldrobotics/Dynamic_Mapping.