Unsupervised Learning based Relative Localization for WCE in a Deformable Tubular Environment

Wireless capsule endoscopy is a promising technology for screening the entire gastrointestinal tract, but the lack of the capbility to locate the capsule is still one of its limitations. To solve the problem, in this paper, we propose a novel relative localization method based on unsupervised learning. The proposed algorithm can estimate the pose and the depth of each image from a image sequence. A Shape from Shading (SfS) based depth constraint is proposed to overcome the problem of “fake far view” and increase the localization accuracy. The proposed method is verified in an image dataset consisting of several pig colons with the average position error rate of 6.71%, which is 4.01% lower than that of the baseline.