Deep point cloud landmark localization for fringe projection profilometry

Point clouds have been widely used due to their information being richer than images. Fringe projection profilometry (FPP) is one of the camera-based point cloud acquisition techniques that is being developed as a vision system for robotic surgery. For semi-autonomous robotic suturing, fluorescent fiducials were previously used on a target tissue as suture landmarks. This not only increases system complexity but also imposes safety concerns. To address these problems, we propose a numerical landmark localization algorithm based on a convolutional neural network (CNN) and a conditional random field (CRF). A CNN is applied to regress landmark heatmaps from the four-channel image data generated by the FPP. A CRF leveraging both local and global shape constraints is developed to better tune the landmark coordinates, reject extra landmarks, and recover missing landmarks. The robustness of the proposed method is demonstrated through ex vivo porcine intestine landmark localization experiments.