Point cloud registration method for linear spatial diffusion

The invention discloses a point cloud registration method for linear spatial diffusion. The method comprises the following steps: acquiring a source point cloud and a target point cloud to be registered; performing global feature extraction on the source point cloud and the target point cloud to obtain a source point cloud global feature and a target point cloud global feature; reasoning prior rigid transformation by using a diffusion model; the diffusion model is obtained by training based on a plurality of noise rigid transformations and corresponding real rigid transformations, the real rigid transformations are rigid transformations between registered sample source point clouds and sample target point clouds, and the noise rigid transformations are rigid transformations correspondingly obtained by performing forward diffusion noise addition on the real rigid transformations; and predicting a point cloud registration result of the source point cloud and the target point cloud according to the source point