Stereo vision and range image techniques for generating 3d computer models of real objects

One topic of research interest today in three-dimensional (3D) model reconstruction is the generation of a complete and photorealistic 3D model from multiple views of an object. This dissertation addresses the problem of generating 3D computer models of real-world objects. We present Stereo Vision systems and Computer Vision techniques for complete 3D model reconstruction through a sequence of steps: (1) Multi-view range image acquisition (2) Registration and integration of multi-view range images (3) Pose estimation of 31) models (4) Integration of two-pose 3D models and (5) Photorealistic texture mapping. We present two stereo vision systems to obtain multi-view range images and photometric textures of an object. Each system consists of a stereo camera and a motion control stage to change the view of the object. Calibrations of both stereo cameras and motion control stages are presented. Range images obtained from multiple views of an object are registered to a common coordinate system through the calibrations of the vision systems. In order to refine the registration of multi-view range images, we introduce a novel registration refinement technique. The proposed technique combines Point-to-Tangent Plane and Point-to-Projection approaches for accurate and fast refinement. In order to merge registered range images, we present two different integration techniques. A mesh-based technique integrates range images through merging of multiple contours on a cross section of a volumetric representation of the object. A slice-by-slice integration on all cross sections reconstructs a complete 3D model represented by a set of closed contours. We also present a volumetric multi-view integration technique. In order to remove erroneous points outside of the visual hull of an object, Shape-from-Silhouettes technique is combined. A 3D grid of voxels is classified into several sub-regions based on the signed-distances of a voxel to overlapping range images. The iso-surface of the object is reconstructed by a class-dependent technique of averaging the signed distances. Marching Cubes algorithm then converts the iso-surface representation of the object to a 3D mesh model. For many real objects, using a single pose yields only a partial 3D model because some surfaces of the object remain hidden from a range sensor due to occlusions or concavities. In order to obtain a complete and closed 3D model, we generate two 3D models of the object, register and integrate the 3D models into a single 3D model. By placing the object in different suitable poses and sensing the visible surfaces, we reconstruct two partial 3D models. We then merge the partial 3D models by novel pose registration and integration techniques. Registration of two pose models consists of two steps, coarse registration, and its refinement. A pose estimation technique between two 3D models is presented to determine coarse registration parameters. The pose estimation technique finds a stable tangent plane (STP) on a 3D model which can be transformed to the base tangent plane (BTP) of the other model and vice versa. After pose estimation, the two pose models are integrated to obtain a complete 3D model through a volumetric pose integration technique. The integration technique merges two iso-surfaces of the corresponding partial 3D models. Texture mapping finally generates photorealistic 3D models of real-world objects.