Intelligent 3d reconstruction method for small celestial bodies based on texture model and noise point cloud cleaning

Abstract

The intelligent three-dimensional reconstruction method for small celestial bodies based on a texture model and noise point cloud cleaning disclosed by the application belongs to the technical field of aerospace. The method of the application is as follows: for the three-dimensional appearance perception problem of deep space small celestial bodies, a MVSNet neural network model with good depth estimation and depth fusion performance is selected, unstructured images containing 1 reference image and N source images are taken as input, and the depth image of the reference image is taken as output; the MVSNet neural network model is trained by using a self-built small celestial body data set, the complex nonlinear relationship between the input two-dimensional optical image and the output depth image is learned, the optimal weight parameter is determined, the restored depth is predicted, and a three-dimensional point cloud model containing texture is generated, the noise point cloud is removed based on the deep space background characteristics to optimize the point cloud model, and high-precision and fast intelligent three-dimensional reconstruction of small celestial bodies is realized. The application has the advantages of high reconstruction solving precision, high efficiency, strong expansion capability and the like, and can improve the relative pose estimation and small celestial body gravity field inversion precision.