Research On The Registration Techniques Of Mutli-view Point Cloud With Texture Atttibutes

3D profile measurement of objects with complex surfaces can be widely used in aerospace, automotive, shipbuilding and machinery manufacturing, etc. Registration of multi-view measurement data is an important part of obtaining the whole profile. Registration techniques for multi-view point cloud data with texture attributes based on binocular stereo measurement system have been fully studied in this thesis, including image matching, data registration, texture addition and texture fusion. The thesis's main contributions are summarized as follows:Several common methods of image matching have been analyzed, and the one based on feature is adopted to establish corresponding point matches between images in different views which are generated by binocular stereo measurement system. Then false matches are rejected according to disparity gradient constraint. The steady matches will be used in the following registration.The advantages and disadvantages of existing registration methods have been analyzed. An automatic registration algorithm is proposed and its superiority is also explained. Firstly, the Five-Point algorithm is employed to compute essential matrix E between two views through selecting five pairs from established point matches, and random sample consensus is adopted to guarantee the algorithm's stability. Secondly, singular value decomposition is used to acquire rotation matrix R and translation vector T. Finally, auto-registration is realized by coordinate transformation of data. The algorithm is simple and reliable, what's more it is unnecessary to use hardware auxiliary equipment such as turntable or adopt markers for assistance. It can ensure high measurement accuracy and has wide applicability at the same time.The algorithms of texture addition and texture fusion have been researched. Firstly, regular triangular mesh model is established according to the correspondences of 3D point and image point, and then a weighing scheme of gravity center is adopted to realize texture addition and smooth display. Secondly, the registered mesh models are simultaneously merged. Each mesh model is in turn defined as the benchmark, and the overlapping areas between the benchmark with the other mesh models are detected simultaneously and adjusted optimally. Then the redundancy is merged and removed, and the gaps are connected smoothly. Thirdly, a weighing scheme of texture values is used to acquire the texture of new point, according to the distance of correspondences from the boundary. This method can ensure smooth transition of texture in the overlapping areas. Finally, texture addition and smooth display are re-operated for new triangles produced by gap connection. Thus, a complete manifold mesh surface with smooth texture attributes can be generated.