The invention relates to the field of computer vision, in particular to a roof segmentation method, a system and a device based on multi-resolution three-dimensional statistical information, aiming atproviding a fast and robust roof segmentation method. The roof segmentation method of the invention comprises the following steps: acquiring high-precision depth and dense three-dimensional point cloud information from multi-view aerial photograph images; based on the statistical information of 3D point cloud features with different resolutions, the 3D point cloud scene is semantically classifiedby global energy optimization, and the point cloud of buildings is obtained. Taking the point cloud of the building as the initial a priori, the fine building segmentation is carried out on the multi-view aerial photograph image. Building transition region based on depth information is computed, and the fine roof segmentation result is obtained by removing the interference of building side to fine roof segmentation from fine building segmentation result. The invention realizes fast, robust and universal roof segmentation, effectively removes the influence of the lateral transition area of thebuilding, and improves the segmentation fineness.