Research On Fast Modeling Method Of Complex Structures Based On Computer Vision

High-quality geometric models are an essential part of scientific research and industrial production,and are a prerequisite for ensuring smooth operations.Currently,when building complex structures,it is usually necessary to first complete measurements and then use modeling software to build 2D or 3D models.When the structure contains many parts or irregular shapes,the modeling workload is extremely large and manual modeling efficiency is low.Therefore,it is of great engineering significance to study rapid,efficient,accurate,and reliable modeling methods for complex structures and complex cross-sections,digitize and parameterize the size information of all models,which not only conforms to the current national trend of digitization,but also can improve the design and development efficiency of enterprises.In this article,we propose a rapid modeling method for complex 2D cross-sections(using gear and tire cross-sections as examples)and complex 3D models(using cars,electric bicycles,and laptops as examples),which integrates computer vision technology,secondary development technology of Auto CAD software and Abaqus software,deep learning algorithms,and image matching algorithms.We have built a complete modeling workflow framework and achieved efficient and accurate modeling of 2D and 3D models.The main research content includes:(1)A fast modeling method based on edge recognition in computer vision technology was proposed for complex 2D models.An optical imaging system was set up to capture images,and image enhancement was performed using image processing techniques.After comparative research,the Canny edge detection operator was selected to extract the contour features of complex 2D sections.Accurate dimensions were obtained using camera calibration and camera imaging models.Based on the secondary development technology provided by modeling software,a dedicated program for fast modeling of complex sections was developed.(2)A quick modeling method for a 2-module 24-tooth involute gear was achieved through secondary development of Abaqus software,taking only 5.6 seconds,which is about 50 times more efficient than manual modeling in 3D CAD software.Compared with the standard gear model,the absolute errors of key features such as the inner hole diameter,tooth root circle diameter,and tooth top circle diameter obtained by the fast modeling method based on computer vision were all less than 1%,and the absolute error of maximum stress was 3.5%.A fast modeling program for tire sections was also developed using the secondary development technology of Auto CAD software,which can be completed in less than one minute,compared to at least one hour for the same model of tire using traditional methods,resulting in an efficiency improvement of about60 times.(3)For complex three-dimensional models,a fast modeling method based on the multi-method fusion of image matching algorithms was proposed.This method combined the advantages and characteristics of two-stage image matching algorithms(SIFT,Super Point+Super Glue)and end-to-end image matching algorithms(Patch2Pix)to explore methods to improve image matching performance and modeling accuracy by combining algorithms(SIFT+Super Point+Super Glue,SIFT+Patch2Pix,SIFT+ Super Point+Super Glue+Patch2Pix,and multi-algorithm and multi-scale fusion)and verified the average matching accuracy and homography estimation on the Hpatches public dataset.The results showed that combining different image matching algorithms could improve matching performance.Among the fusion methods proposed in this study,the second fusion method achieved the highest matching performance and provided a solid foundation for three-dimensional model reconstruction.(4)Replace the SIFT algorithm in the COLMAP 3D reconstruction framework with the matching algorithm proposed in this paper,and quickly model cars,electric bicycles,and laptops on a workstation.The number of images used were 17,18,and 15,respectively,and the time taken was 20 minutes,26 minutes,and 13 minutes.The point cloud increased by 209%,134%,and 131%,respectively.The modeling time for complex3 D models that would normally take a day or even several days was reduced to less than an hour,greatly shortening the product modeling time and improving product development efficiency.