NeRF Representation-Based Efficient AAV Path Planning Framework for High-Fidelity 3-D Ship Reconstruction

As an efficient and inexpensive acquisition platform, Unmanned Aerial Vehicles (UAVs) are particularly suitable for image acquisition in 3D ship reconstruction tasks. However, path planning in ship reconstruction tasks is often required to consider the energy consumption of UAV batteries and the applicability of the reconstruction method, which introduces challenges to the efficient reconstruction of high-fidelity 3D ship models. To address the above challenges, this paper proposes a Neural Radiance Field (NeRF) representation-based efficient UAV path planning framework for high-fidelity 3D ship reconstruction. First, a multi-modal adaptive generation strategy for rays in NeRF is proposed by fusing the features of the point cloud and image. Second, a novel reward function and a viewpoint partitioning optimization guidance strategy for the deep reinforcement learning of the UAV path planning framework are proposed to ensure the efficiency and robustness of ship reconstruction tasks. To further optimize the 3D reconstruction result of the ship, a local adaptive distance-aware sampling policy is proposed to boost the overlap rate of the image data, which enhances the framework’s applicability for downstream reconstruction tasks. Finally, we demonstrate the efficiency and applicability of the UAV path planning framework through a series of experiments.