A three-dimensional computer vision system for inspecting the geometric tolerances of circular machine features

In manufacturing, it is impossible to produce a part with ideal features. Manufactured features always exhibit geometric deviations from their nominal geometrical properties both for systematic and random reasons. Part inspection is therefore crucial to verify that the geometric deviations of the machined features remain within the tolerance limits. This research focuses on developing a three-dimensional (3D) machine vision system for inspecting the compliance of circular machine features to 3D geometric tolerances.;Characterizing the compliance of a machined part to 3D geometric tolerances from its image involves the following basic issues: image feature extraction, 3D feature re-construction, error propagation, and tolerance inference. Feature extraction involves detecting two-dimensional (2D) geometric features from the part image. The detected features are used for measurement and camera calibration. Through this research, we have developed statistical techniques for extracting 2D features like corners, lines, and ellipses. Three-dimensional feature reconstruction is concerned with reconstructing the corresponding 3D features from the detected 2D features. This is necessary since dimensioning based on the 2D features may produce significant measurement errors due to geometrical distortion caused by the perspective projection. For 3D feature reconstruction, we have developed several techniques, including: an integrated technique for camera calibration, a constraint-based technique for 3D reconstruction from monocular images, an efficient and accurate technique for point matching, and an improved Bayesian triangulation technique for 3D reconstruction from multiple images.;Since an image is never noise free, tolerance measurements are subject to image errors. We have developed techniques for analytically studying the impact of image errors on the estimated camera parameters, the reconstructed 3D points, and on that estimated tolerance measurements. Finally, given the reconstructed 3D features, we introduce a statistical framework for modeling image and manufacturing errors and for statistically inferring the geometric tolerances.;We have demonstrated using two images for each part with a resolution of 512 x 384 over a field of width of 3 inch. the shape and size tolerances can be determined to within an average of 0.8 mm respectively, and the position tolerance to within an average of 1.5 mm.