High-Precision 3D Pose Estimation Of 3C Components Based On Multi-2D Vision

There are a lot of actions involving 6 degrees freedom(6-Do F)adjustment in the automatic assembly of 3C products.The traditional 2D vision is featured by high precision and fast processing speed,but when it comes to pose estimation,the traditional 2D vision prevails in plane pose estimation instead of 3D pose estimation,and the research regarding this is insufficient;since it is time-consuming to collect point cloud data through 3D vision sensors and it is difficult to meet industrial demands.Therefore,this dissertation explores how to realize high-precision 3D pose estimation by multi-2D camera system taking the advantage of high precision and small computational complexity of the traditional 2D vision.In this dissertation,three 2D cameras are placed in an orthogonal layout to construct a multi-camera system,and images with different characteristics of 3C components are taken for pose estimation.In so doing,the problem of high-precision pose estimation is transformed into the analysis and fusion of the estimated parameter value obtained by every single camera.Through simulated analysis of common pose estimated algorithms,this dissertation verifies that iterative Pn P algorithm can better meet the project requirements in terms of time and accuracy,establish an error analysis model for single 2D camera pose estimation,estimate error sources and explore their degrees of influence on accuracy,finally reach the conclusion that there are advantaged and disadvantaged parameters in single 2D camera pose estimation.After that,feasibility of multi-amera fusion is demonstrated by combining dimension features of 3C components.Later,the selection of marker is discussed and the layout of the overall features is designed.In terms of the image extraction accuracy problem,we propose the ELSDc rough detection-Zernike moment sub-pixel detection-least square ellipse fitting framework.Considering the center projection deviation of two-dimensional image under perspective projection,this dissertation combines the perspective projection model with concentric circle properties to compensate the deviation of two-dimensional marker,and achieves the accurate extraction of feature points.The performance of the pose estimation system of single 2D camera is simulated and verified though the blender rendering platform.After the above analysis and optimization,this dissertation introduces common multi-camera systems and calibration methods,fuses the estimated results of multi-camera systems based on coordinate system conversion,analyzes the pose estimation performance of 2D camera system through simulation,and compares with that of single 2D camera.Finally,the overall framework is verified by experiments.Results show that the translation error is within 0.2mm,the rotation error is within 0.3°,the time-consuming is less than 2s,which basically achieves the expected goal of high-precision 3D pose estimation for 3C components.