Vision Measurement For Rotation-Symmetric Targets Using Monocular Camera

The purpose of vision measurement is to obtain the position, pose and other parameters of the target by using images captured by imaging systems. Most of the vision measurement methods today are based on local features on the target, such as points, lines and many other local features that are prone to disruption of noise when being extracted, which affects the accuracy of measurement. However, a number of global features are available in both industrial and life scenarios. Rotation-symmetric features, such as wheels and propellers, are common among these global features. In this paper, a vision measurement method for rotation-symmetric targets using monocular camera is proposed, which uses the global features of the rotation-symmetric targets without extracting local features. Therefore it avoids the errors generated by extracting local features such as points and lines. The major research work is as follows:(1) A vision measurement method based on single image is proposed. By using single image of the rotation-symmetric patterns, the proposed method is able to calibrate the focal length and correct the radial lens distortion, as well as obtain the position and pose parameters of the target. The method is based on the fact that when the camera is conducting virtual parallel imaging with its image plane parallel to the plane of rotation-symmetric patterns, the obtained image is rotation-symmetric. Setting the symmetric center of the image as the center of a circle and expanding the image along its radial direction (Frieze Expansion Transform, FET), the obtained matrix is low-rank. However, when the camera is imaging with a tilt angle, the matrix obtained after conducting FET is not low-rank. Thus, the relationship between the camera parameters obtained by imaging with a tilt angle and camera parameters obtained by conducting virtual parallel imaging can be established. Further, a low-rank matrix recovery method is harnessed to recover the rotation-symmetric image from the image obtained by tilt imaging, meanwhile the focal length, radial distortion, pose and position of the camera can be estimated. The proposed vision measurement method that is based on single image has the merits of using a relatively simple model and having a fast processing speed. Besides, it’s easy for users to operate as only one image is required to complete the measurement.(2) A vision measurement method based on multiple images is proposed. By using multiple images of the rotation-symmetric patterns, the proposed method is able to calibrate the focal length and main point, as well as correct the radial lens distortion and obtain the position and pose parameters of the target. First the proposed method establishes a homographic relationship between the obtained images and the rotation-symmetric image. Then by using low-rank matrix recovery method, the rotation-symmetric image can be recovered from the images obtained by tilt imaging, meanwhile the homographic matrix can be estimated. The homographic matrix computed by the images obtained from different viewpoints would pose different constrains on the camera intrinsic parameters which can be estimated if these constrains are combined. Further, the pose and position parameters of the target can be estimated after the intrinsic parameters are obtained. The method based on multiple images has the merit of measuring more parameters of the camera with high accuracy.(3) An automatic installation measurement system for industrial work pieces is designed by using the proposed vision measurement methods. This system has a complete software interface and is capable of automatic measurement. Besides, it overcomes some problems including uneven illumination and is able to provide pose and position parameters for the automatic installation of industrial large work pieces.The proposed vision measurement method uses the global features of the rotation-symmetric targets without extracting local features. Therefore it avoids the errors generated by extracting local features such as points and lines. Extensive simulation and real experiments prove that the proposed method is accurate, robust and the processing speed can meet the practical requirement. The implement to industrial scenarios proves the practical value of the proposed method.