Camera Pose Estimation Based on Feature Extraction and Description for Robotic Gastrointestinal Endoscopy

The application of robotics in gastrointestinal endoscopy has gained more and more attention over the past decade. The localization and navigation of the robotic gastrointestinal endoscopy is very important in robot-assisted gastrointestinal examination and surgery. The camera pose of the robotic gastrointestinal endoscopy can be estimated directly from the image sequence. However, due to the texture-less nature and strong specular reflections of the digestive tract surface, it is hard to detect enough keypoints to estimate the camera pose when using the traditional handcrafted method. In this paper, we propose an end-to-end CNN-based network to deal with this problem. Our network is trained in a self-supervised manner, and the network plays two roles, a dense feature descriptor and a feature detector simultaneously. The network takes the image sequence as input, and the featured keypoints and their corresponding descriptors as outputs. We demonstrate our algorithm on images captured in stomach phantom. The experimental results show that our method can effectively detect and describe the featured keypoints in challenging conditions.