Depth estimation and dense reconstruction method for monocular SLAM in intracavity environment

• Endo-MVSNET enhances scene continuity and edge detail preservation • Provides robust depth measurement prior for subsequent 3D reconstruction. • Keyframe selection strategy based on common view and disparity mapping, enhancing the accuracy and efficiency of position measurement. • Hierarchical iterative strategy integrates the geometric and color consistency constraints to reduce the negative impact of cumulative errors on the reconstruction quality. • Construct a light-weight TSDF optimization network to enhance the reconstruction completeness of thin-walled structures and complex intracavities. The key bottleneck in precision geometric measurement of intracavity lies in high-accuracy three-dimensional (3D) reconstruction. 3D reconstruction technology based on depth measurement and pose measurement serves as the core support for overcoming this bottleneck. The paper proposes a dense reconstruction method for monocular simultaneous localization and mapping (SLAM) tailored to intracavity environments. This method addresses issues such as depth blurring in weakly textured regions, positional drift due to dynamic deformation, and insufficient reconstruction integrity in monocular endoscopes within intracavity environments. The development of a depth estimation network that incorporates transfer learning, attention mechanisms, and neighborhood smoothing loss has enabled the creation of a robust depth measurement prior. Subsequently, a deep learning-enhanced truncated signed distance Function (TSDF) optimized SLAM framework is constructed, with algorithm optimization and system integration to enhance 3D reconstruction accuracy in intracavity scenes. According to the experimental results of the Hamlyn dataset, the absolute relative error of depth estimation reaches 0.245. Furthermore, simulation experiments substantiate the efficacy of the optimization method in intracavity scenes, exhibiting an absolute trajectory error of 0.007444 across 3,000 simulation frames. The reconstruction results from the Hamlyn datasets and real-world data demonstrate that the proposed method achieves optimal surface completeness and reconstruction quality for fine texture details. The proposed method provides a comprehensive solution for depth estimation, keyframe optimization, and 3D reconstruction in monocular SLAM for intracavity scenarios. It offers theoretical guidance and technical validation for the practical application of 3D geometric measurement technology.