Monocular depth estimation plays a pivotal role in critical domains such as autonomous driving and robot navigation, holding an extremely high application value. Particularly, research on monocular depth estimation using self-supervised deep learning imposes more flexible and lenient requirements on training datasets, enabling model training without actual depth information. However, compared with outdoor datasets dominated by translational motion, indoor datasets have more rotational motions. This will make it difficult to predict accurate camera poses, thereby affecting the accuracy of depth estimation. To address this problem, we propose an indoor depth estimation method based on iterative pose refinement, thereby improving the rotation processing ability through hierarchical optimization. Using the self-supervised architecture MonoDepth2, we creatively reconstruct the pose estimation module into a cascaded structure. On the one hand, an initial network is designed to predict camera transformations from adjacent frames; on the other hand, a residual network is constructed to iteratively correct rotation drift using photometric reprojection loss. Through this iterative fusion mechanism, the geometric constraints between frames are effectively strengthened. We built a prototype system and conducted experimental evaluations on the KITTI dataset and a self-built indoor dataset rich in rotation scenarios. The experimental results show that compared with the baseline method, our framework effectively reduces the error by 14.5%, significantly improving the robustness to complex rotation scenarios.