UAV-Based Large-Scale Active Mapping with Implicit Neural Representations

Active reconstruction enables robots to autonomously explore scenes, achieving full coverage and alleviating the burden of manual data collection. Existing methods, however, are limited by complex scene representations, particularly in large-scale environments. Implicit neural representations, especially Neural Radiance Fields (NeRF), offer significant advantages for high-fidelity reconstruction of large-scale scenes due to their ability to capture intricate details and optimize storage. In this paper, we propose an implicit neural representation-based framework for large-scale active reconstruction, enhancing NeRF to identify low-quality regions and performing online quality evaluation to guide efficient exploration. We also design an active reconstruction strategy with smooth flight trajectories to improve reconstruction quality and completeness. The quadrotor platform is chosen for its high flexibility. Extensive experiments and ablation studies in simulated environments validate the effectiveness of our method.