Optical three-dimensional coordinate measurement using stereo vision has systematic errors that affect measurement quality. This paper presents a scheme for measuring, modelling and correcting these errors. The position and orientation of a linear stage are measured with a laser interferometer while a stereo vision system tracks target points on the moving stage. With reference to the higher accuracy laser interferometer measurement, the displacement errors of the tracked points are evaluated. Regression using a neural network is used to generate a volumetric error model from the evaluated displacement errors. The regression model is shown to outperform other interpolation methods. The volumetric error model is validated by correcting the three-dimensional coordinates of the point cloud from a photogrammetry instrument that uses the stereo vision system. The corrected points from the measurement of a calibrated spherical artefact are shown to have size and form errors of less than 50 μm and 110 μm respectively. A reduction of up to 30% in the magnitude of the probing size error is observed after error correction is applied.

使用立体视觉的光学三维坐标测量具有影响测量质量的系统误差。本文提出了一种测量，建模和纠正这些误差的方案。线性平台的位置和方向是用激光干涉仪测量的，而立体视觉系统则跟踪移动平台上的目标点。参考更高精度的激光干涉仪测量，评估跟踪点的位移误差。使用神经网络的回归用于根据评估的位移误差生成体积误差模型。回归模型显示出优于其他插值方法。通过使用立体视觉系统的摄影测量仪器校正点云的三维坐标，可以验证体积误差模型。来自校准球形伪像的测量的校正点显示为分别具有小于50μm和小于110μm的尺寸和形状误差。进行纠错后，可以观察到探测尺寸误差幅度最多降低30％。