Improved ORB-SLAM Based 3D Dense Reconstruction for Monocular Endoscopic Image

Monocular visual simultaneous localization and mapping (SLAM) performs effectively in camera pose estimation and 3D sparse reconstruction of natural scenes. However, in monocular endoscopic environment, serious distortion of the images and the inconstant illumination, even the lack of surface texture, make SLAM-based tracking and 3D dense reconstruction still a challenge. In response to the above problems, it is proposed to use local features to match adjacent frames in ORB-SLAM system for the endoscopic poses estimation and keyframes selection, then combined with the probabilistic monocular stereo technology to calculate the dense depth map from keyframes, and finally complete the 3D dense reconstruction of the endoscopic scene. The experimental results proved that this method can track the endoscope robustly and reconstruct a 3D point cloud with high density and smoothness.