基于单视图的非合作航天器三维体素重构算法

On-orbit relative navigation tasks in space confrontation situations, such as fast and high-precision target perception, intent recognition and threat assessment, rely on accurate three- dimensional model of the target. However, the perception field of space-borne sensors is limited, and traditional 3D reconstruction algorithms rely on multi-angle views, resulting in complex processes, poor real-time performance, and low attitude estimation accuracy. To solve these problems, this paper proposes a single-view-based 3D voxel reconstruction model for non-cooperative spacecraft, which can efficiently restore the spatial 3D information of the target based on a single frame image. In order to solve the common data hungry problem in visual perception tasks in space scenes, a small-scale local dataset is constructed by collecting 3D model of the spacecraft and using image rendering technology. The whole dataset contains the 3D model of the spacecraft, corresponding voxels, and images sampled in multi-pose. The network uses a single frame of non-cooperative target spacecraft image as input, and completes the three sub-module calculations of “Encoding-Decoding-Tuning” through 2D-3D hybrid convolution to restore the target 3D voxel structure. The training and quantitative evaluation results of the model on the local data set show that the proposed model can accurately reconstruct the 3D voxel (0. 895 MIoU) based on a single-view image in real time, and the reconstruction on different non-cooperated spacecraft shows a strong generalization ability.