For robotic smooth assembly of cylindrical shaft and hole parts, a system for robotic automatic assembly based on 3D vision, monocular vision and admittance control is established. The strategy for assembly of cylindrical shaft and hole parts is proposed, integrating axis pose estimation based on 3D point clouds, object detection using image deep learning, and admittance control. Aiming at the pose estimation of cylindrical hole parts based on 3D vision, an algorithm of axis pose estimation based on 3D point clouds is studied. Firstly, the method of keypoint selection on 3D point clouds is introduced. Then, based on the geometric constraints of point cloud surface normal and axis, the algorithm of coarse axis estimation is proposed and analyzed. After that, based on the coarse axis estimation, the algorithm of axis pose optimization based on iterative robust least squares is proposed and analyzed. The experimental results show that the angle RMSE of estimated axis pose is 0.248$^\circ$ and the position RMSE of that is 0.463 mm. Compared to the existing popular methods of axis estimation, the proposed method has the higher accuracy. The assembly strategy can meet the requirements of high precision, stability and reliability in robotic assembly of cylindrical shaft and hole parts.