Visual Odometry in Stereo Endoscopy by Using PEaRL to Handle Partial Scene Deformation

Stereoscopic laparoscopy provides the surgeon with the depth perception at the surgical site to facilitate fine micro-manipulation of soft-tissues. The technology also enables computer-assisted laparoscopy where patient specific models can be overlaid onto laparoscopic video in real-time to provide image guidance. To maintain graphical overlay alignment of image-guides it is essential to recover the camera motion and scene geometry during the procedure. This can be performed using the image data itself, however, despite of the mature state of structurefrom-motion techniques, their application in minimally invasive surgery remains a challenging problem due non-rigid scene deformation. In this paper, we propose a method for recovering the camera motion of stereo endoscopes through a multi-model fitting approach which segments rigid and non-rigid structures at the surgical site. The method jointly optimizes the segmentation of image and uses the rigid structure to robustly estimate the motion of the laparoscope. Synthetic and in-vivo experiments show that the proposed algorithm outperforms RANSAC-based stereo visual odometry in non-rigid laparoscopic surgery scenes.