Reasearch On Depth Estimation Of Stereo Vision For Zoom Camera

Stereoscopic technology has achieved rapid development in recent decades.Because of its non-contact,high speed and high precision,it has been widely used in areas such as non-destructive testing,scene sensing,robot navigation,medical assistance,virtual reality,military detection,etc.,which require deep recovery,3D target reconstruction and description.Stereoscopic vision has three manifestations: stereo vision based on binocular camera,stereo vision based on monocular camera,and stereo vision based on camera plus structured light.Each of the three implementation forms has advantages and disadvantages,wherein monocular stereo vision refers to the use of a set of imaging systems or two imaging systems that are separately imaged on the same optical axis,and stereoscopic image acquisition of spatial objects in the same orientation,and then Two or more monocular images are analyzed by stereo processing,and the depth estimation information of the spatial points is calculated according to the vector displacement of the image points on different images,the pixel gray value,etc.,and a new technique for realizing three-dimensional description is realized..Compared with the traditional binocular depth measurement method,monocular stereo vision can effectively avoid the increase of the baseline distance between the two cameras in the binocular stereo vision system,and the measurement coverage of the stereo image pair is small,subject to ambient illumination.The characteristics affecting the texture of the image pair and the like affect stereoscopic matching and the like.Aiming at the main problems of monocular zoom stereo vision,this paper mainly studies the distortion correction technology of zoom camera image acquisition,the decoupling calibration technology of zoom camera and the depth estimation technology of zoom camera.The main research results include:(1)According to the vanishing point geometric constraint characteristic formed by the line points in the plane checkerboard image,the decoupling lens distortion model with single parameter first-order form and the minimum fitting error constraint are used to solve the focal length of the zoom lens.Several sets of lens distortion parameters at different focal lengths are then used to fit the straight line of the corrected points by least squares method.The Levenberg-Marquardt algorithm is used to re-optimize the distortion coefficient and distortion center coordinates according to the principle of minimization of the deviation error between the point and the fitted line.Then,the spatial distribution of these distortion parameters with the focal length is plotted and the curve fitting is used to linearly fit these points.The empirical formula of the distortion parameter distribution with the focal length change or the training of these parameters according to LS-SVR is constructed.The distortion parameter lookup table at each focal length of the zoom lens finally realizes image correction for the nonlinear distortion of the real image.(2)Based on the radial distortion separation model and the corner sub-pixel coordinate distribution,the homography matrix is calculated according to the mapping relationship between the corner coordinates of the image and its corresponding world coordinate system coordinates,and then passes through the homography matrix.The internal and external parameters of the camera are decomposed.Since the camera distortion center and the first-order radial distortion coefficient have been iteratively optimized by using the vanishing point constraint,the decoupling between the parameters of the camera's distortion coefficient can be finally realized.(3)This paper uses SIFT feature matching technology to match the image pair features acquired by the zoom camera at two focal lengths,and obtains the pixel position of the matching point.Then,according to the bifocal imaging principle,the actual depth of field information corresponding to each pair of matching points is calculated.Delaunay triangulation method meshes the coordinate data of each set of matching points,and performs color mapping on each triangulation area according to the depth value to achieve a rough estimation of the target object depth.