Spatial–Temporal Diffusion Model for Underwater Scene Reconstruction With Application to AUV Navigation

autonomous underwater vehicles (AUVs) have been extensively utilized in subsea exploration and surveying. However, accurately perceiving the surrounding environment remains a significant challenge for AUVs due to the complexities of subsea terrains. To address this issue, we propose a novel generative scene reconstruction method to enhance AUVs’ perception capabilities. Our method is primarily designed for reconstructing dense subsea terrain from 3-D multibeam echosounder data. We leverage local diffusion and denoising strategies to reconstruct complete subsea terrain at the scene scale directly, without requiring normalization from point clouds. Considering the motion dynamics of AUVs and the overlap between consecutive sonar frames, we introduce a spatial–temporal attention mechanism to aggregate features from consecutive point clouds and guide the reconstruction process as a condition. Then, the reconstructed point cloud is utilized for probabilistic terrain modeling through Bayesian updating, enabling path planning. Experiments conducted on simulation and real-world datasets demonstrate that our method can generate more accurate and complete terrain maps. Furthermore, path planning based on our reconstruction method achieves the shortest and smoothest motion path, further validating that our reconstruction method can provide more complete perception information for AUV navigation.