<p>Monocular depth estimation is a critical component in various vision tasks, including robotic navigation and autonomous driving. Traditional methods rely on expensive hardware, limiting their applicability. Self-supervised learning has emerged as a promising alternative, yet challenges remain in weakly textured regions and object boundaries. To address these, we propose SFP-Depth, a novel framework integrating semantic segmentation and planar constraints. SFP-Depth employs a multitask learning paradigm, a boundary-aware gradient mask, and a planar depth constraint to enhance depth estimation accuracy. Extensive experiments on the KITTI dataset demonstrate the superiority of our approach, achieving an absolute relative error of 0.095, a 6.8% improvement over FSRE-Depth, and outperforming state-of-the-art methods in accuracy and robustness. Here, we show that leveraging semantic and geometric priors can significantly advance self-supervised monocular depth estimation, with broad applications in real-world scenarios. Source code available at <a href="https://github.com/hwlf/SFDepth">https://github.com/hwlf/SFDepth</a></p>