Incremental Multi-view Point Cloud Extraction And Optimization

3D reconstruction is an important branch of computer vision research,with significant application value in industries such as urban planning,medical diagnosis,and map surveying.Common methods for 3D reconstruction include active vision-based and passive vision-based.Active vision-based require the active emission of lasers or structured light from measurement devices,but have limited coverage and are suitable for high-precision measurement of small scenes.Passive vision-based directly use camera sensors to capture images and reconstruct 3D models by extracting point clouds from multi-view,making them suitable for complex scene reconstruction.Over the years,researchers have proposed many passive vision-based 3D reconstruction methods with good performance.However,there are still significant challenges in addressing issues such as varying outdoor lighting conditions,image distortion,and point cloud drift.In this paper,we propose modifications,fusion,and optimization of image feature point detection,feature point matching,and point cloud reconstruction by using methods such as color invariant models and similarity measurements,aiming to improve multi-view three-dimensional reconstruction in terms of reducing reconstruction time,expanding reconstruction area,and enhancing reconstruction accuracy.The main contributions of this paper are summarized in the following three aspects:(1)A color invariant image matching algorithm is proposed to address the problem of fewer feature points extracted and poor matching capabilities under complex outdoor lighting conditions,which in turn affects the effectiveness of point cloud extraction.Firstly,a color invariant feature detection algorithm based on Kubelka-Munk theory is designed,which distinguishes feature regions that can be captured in scenes based on color invariance and combines AKAZE and SIFT algorithms to generate more comprehensive feature descriptors.Then,Tanimoto similarity are used to screen feature point pairs,followed by bidirectional feature matching,and finally,random sample consensus algorithm is used to remove outliers.This method not only alleviates the problem of insufficient features but also improves feature matching accuracy,thereby improving image matching results.(2)Aiming at the issues of slow extraction speed and more outliers in incremental 3D point cloud reconstruction in the case of a large amount of images,this paper improves incremental motion recovery and reconstruction.A score function is designed based on the sum of the angles between the rays leading out from the feature points and the area of the minimum enclosing circle of the feature points.The image pair with the highest score is selected as the initial image pair,which tends to select images with larger feature point coverage areas and wider baselines between images.The extracted 3D point cloud is filtered by a mixed outlier filtering method to improve reconstruction accuracy.This improved incremental structure from motion method not only reduces the time consumed by bundle adjustment but also reduces the root mean square error of the extracted point cloud,optimizing the extraction result.(3)A 3D reconstruction software based on the incremental multi-view method is designed and developed on the PyQt4 platform.The software has a complete graphical operation window and a lightweight architecture,integrates the improved algorithms for realizing functions such as image import,feature processing,point cloud generation,and 3D structure reconstruction.Experimental results on the Pix4d dataset show that the software designed in this paper achieves optimal performance in terms of the number of extracted point clouds,average reprojection error,and chamfer distance,with a decrease in average reprojection error of 31%to 67.2%compared to other three-dimensional reconstruction methods,and an increase in the number of reconstructed point clouds by 5.4%to 37%.