The Study Of Three-dimensional Measurement Based On Binocular Vision

Computer vision technology has the characteristics of non-contact and a high degree of automation, so it has broad application prospects in the surface quality inspection, measurement and shape recognition and so on. Computer vision research aims at getting structure information of three dimensional objects from two dimensional images. However, the current measurement techniques based on machine vision mostly deal with the two-dimensional images obtained of projection or a surface of objects. This greatly reduces the efficiency of measurement and narrows the range of objects which can be measured automatically using the computer vision.3d size detection method of complex shaped objects based on binocular vision in this paper can simultaneously detect the target size characteristics in different planes of object, so it can greatly improve the measurement and auto-recognition efficiency of different geometric primitives and expand the scope of visual measurement applications and has important theoretical research value and broad application prospect.In this paperⅠconducted the theoretical analysis and experimental study on binocular CCD camera calibration, sub-pixel edge detection, image registration and the recognition and fitting of some geometric primitives—lines, circles and ellipses around three-dimensional measurement using object digital images.Binocular stereo vision system uses two cameras to simulate the human eyes to obtain two digital images of the detected objects from different angles at the same time, and then the restoration can be done based on the parallax principle. It is the physical basis of implementing the three-dimensional image size measurement. This paper firstly analyzes the characteristics and applicability of different binocular vision system models and selects the model which can achieve higher precision and requirements, and then rationally designs and selects the structural parameters of the visual system according to their influences on the measurement accuracy, finally builds a simulation of binocular stereo vision system with two digital cameras and creates the non-linear imaging camera model with distortion which is consistent with the actual situation. After the initial calibration of the camera using black and white checkerboard template, Levenberg-Marquardt algorithm is used for nonlinear optimization, and finally high-precision camera calibration is achieved; Sub-pixel edge detection based on wavelet transform and least square fitting can extract the edge of the object to sub-pixel accuracy. This will help the measurement results to achieve a more accurate. But the experiment proved that the sub-pixel edge detection based on least squares fitting is easier to implement, has wider applicability, and is more suitable for cross-edge detection; Image registration is the key and also is the difficulty to achieve stereoscopic vision. SIFT image matching algorithm can match feature points of left and right images effectively and accurately thereby increasing the accuracy of three-dimensional coordinate space point; Complex shape object often contains a variety of geometrical element, such as linear, round, arc, elliptical and various curve, etc. This paper only studies the recognition testing and the fitting method with contour data of linear, round, and elliptic limited to the research conditions and technical level. Image correction can be conducted using 3d coordinate points in the same coordinate which were obtained after image registration, and then detect and compute related size characteristics by using 2-d image size calculation, such as round radius, elliptical long axis and short axis, etc. This research work has some certain significance to the development of 3d digital image measurement and guiding the engineering application of image measurement techniques.