ABSTRACT
This paper presents a high-accuracy method for globally consistent surface reconstruction using a single fringe projection profilometry (FPP) sensor. To solve the accumulated sensor pose estimation error problem encountered in a long scanning trajectory, we first present a novel 3D registration method which fuses both dense geometric and curvature consistency constraints to improve the accuracy of relative sensor pose estimation. Then we perform global sensor pose optimization by modeling the surface consistency information as a pre-computed covariance matrix and formulating the multi-view point cloud registration problem in a pose graph optimization framework. Experiments on reconstructing a 1300 mm × 400 mm workpiece with a FPP sensor is performed, verifying that our method can substantially reduce the accumulated error and achieve industrial-level surface model reconstruction without any external positional assistance but only using a single FPP sensor.
