Stereoscopic Scene Flow for Robotic Assisted Minimally Invasive Surgery

Information about the 3D shape and motion of tissue surfaces at the surgical site during minimally invasive surgery is important for providing metric measurements that enable the deployment of image-guidance and enhanced robotic control. This article presents a scene flow algorithm that recovers the deformation and 3D structure of the surgical field-of-view from stereoscopic images by propagating information starting from a sparse set of candidate seed matches. By imposing spatial and temporal constraints the proposed algorithm is able to reconstruct dense 3D scene flow accurately and efficiently. Validation is performed using simulation data to evaluate the method against varying levels of image noise and results are also presented for benchmark phantom model data. The practical value of proposed method is shown by qualitative results for in vivo videos from robotic assisted procedures.