【文章內容簡介】 They are regarded as matching features. For improving measurement precision, twocamera with converging opticalaxis is chosen. And the twocamera and the small selfmade projector constitute a flexible measuring head. A sketch of the 3D measurement principle based on stereo vision is shown in . 3 Establishment of the Mapping Relation Between Image Point and Special Point Actually, obtaining 3D information of object from a pair of two images is by mapping relation between image point and special point, but until now no approach can pletely describe the nonlinear mapping relation since there are many plex nonlinear influencing factors including radial distortion and lateral distortion of camera. However, neural network can simulate human vision to establish plex mapping by simple nonlinear processing cells, so this paper regards the middle process from image point to special point as a black box. And BP network with a middle layer of six neural cells is used to set up the mapping relation between image point and special point. Point A in left image and a point in right image are input into the BP network, a special point is output. In other words, the structure of BP network is 463. Using neural network, the choosing of training samples is important The training samples not only lie in the measurable range, but also show measurement range of measurement system.While twocamera is used to grab object, the object and the part of object only in jointing viewing field can be able to be grabbed. So 3D information of object from a pair of stereo images, lens focus, measurement precision, once measuring area and the distance between object and baseline of twocamera control 3D measurement range of the system are obtained.In this paper, the structure and function of the two cameras that are posed symmetrically are identical, and the image area is , just as . The lens focus is f。 the line between two image centers is perpendicular to . The mon part is regarded as joining viewing field of twocamera. And the part out of is known as blind area. If 2 is viewing field angle, on the basic of imaging relation, the formula is (1) An inscribed circle is done in the joining viewing field, if β is included angle of twocamera optical axis, 2 is the distance between two image centers, its ratio is (2)In this way, a 2R2R sample template with 88 grids is made. The sample template is put worktable. Three pairs of stereo images are grabbed respectively, while the sample template is moved to three different heights (0, R, 2R) along the vertical direction to simulate 3D measurement range. The three pairs of stereo images are regarded as training samples, and they are input network.4 Stereo Precise Matching at Subpixel Level Stereo precise matching is much more difficult in stereo vision, so the applying of stereo vision is restricted in a way. In this paper, wavelet transform is applied to detect edge points, searching nonsupervisor clustering approach is proposed to distinguish the different edge point groups. The edge points in the same point group are fitted quadratic curve, and then stereo precise matching is achieved at subpixel level based on geometry invariance. Stripe Edges Fitting Based on Searching Nonsupervisor Clustering Generally, image often contains random noise, and wavelet transform can restrain noise and detect edge, while different structure image edges are described by the information spreading in all resolutions. Since translating invariance is the most important in stereo matching based on edge feature. Quadratic Bspine is selected for a multiscale generating element to detect edge points of stripe.Act