3D Reconstruction Of Small Object Based On The Structured Light

Recently,various computer vision applications have emerged in endlessly.2D image information has been unable to meet the needs of human production and life.Reconstructing the 3D model of real scene veritably and objectively has become one of the main tasks of computer vision.3D reconstruction technology based on structured light can reconstruct a high precision 3D model by projecting coded light with certain regularity and using the information of coded light to match corresponding points in stereo vision.This technology has been a research hotspot because of its low cost of device and high accuracy of reconstruction.This thesis firstly describes the basic principle of binocular stereo vision and builds the experimental system.Then,it studies different kinds of encoding and decoding schemes of structured light,and achieves 3D reconstruction with sub-pixel accuracy.Finally,registration of 3D point clouds from multiple perspectives is carried out to obtain a complete 3D model.The specific studies and innovation works are as follows:1.This thesis studies the basic principle of binocular stereo vision,that is,the method of recovering 3D information from 2D images.It mainly includes imaging model of camera,epipolar geometry constraint,stereo matching,triangular intersection and calibration of camera's interior and external parameters.2.In this thesis,the encoding and decoding principle of gray code and phase code are studied respectively.Two reconstruction methods are adopted,which are the combination of gray code with phase code and high-low frequency phase code.The first method solves the problem that the resolution of gray code can't reach one pixel width.It realizes one-to-one correspondence matching and achieves 3D reconstruction with pixel accuracy.On this basis,this thesis uses the linear property of phase code to fit the phase line,and achieves the 3D reconstruction with sub-pixel accuracy.The second solves the problems that periodic ambiguity of high-frequency phase and large noise of unitfrequency phase.In the dual-frequency mapping method,there will be mapping errors when over-high frequency is directly mapped to unit frequency.This thesis proposes a three-frequency mapping method to solve this problem and improve the reconstruction accuracy.3.The point cloud registration is studied,which combines multiple point clouds from different perspectives into a complete 3D model.Aiming at the problem that ICP algorithm is sensitive to the initial position of point cloud and easy to fall into local optimum,this thesis uses rotating platform to achieve rough registration of point cloud.The rough registration result is used as the initialization of ICP algorithm,which can improve point cloud registration accuracy.For the loop closure problem in multiple point clouds registration,a global optimization algorithm is used to distribute the cumulative errors reasonably to each point cloud to offset the impact of registration errors.Finally,point clouds are merged by voxel filtering,and a complete 3D model with uniform density is obtained.In this thesis,the accuracy of the improved method is verified by reconstructing standard geometrical objects and fitting them.Then,by reconstructing some irregular small objects,the practicability of this method and the reliability of the experimental system are tested.