Research And Application Of 3D Reconstruction Technology Based On Multi-View And Low Overlap Rate Point Clouds

Point cloud-based 3D reconstruction is the process of utilizing point cloud data to create accurate and detailed 3D models of real-world objects or scenes,which can be used for various applications such as visualization,analysis and simulation.In recent years,point cloud-based 3D reconstruction has received continuous attention due to its wide application in different fields,and many research results on point cloud-based 3D reconstruction algorithms and technologies have been achieved3D reconstruction based on point cloud data typically involves steps such as point cloud acquisition,preprocessing,registration,and surface reconstruction.Among them,point cloud registration is the alignment and stitching of multiple point clouds acquired from different perspectives or sources to generate a complete 3D model,which is one of the main challenges of 3D reconstruction.At present,point cloud registration methods are mainly suitable for high overlap rate point cloud processing,and their processing accuracy and speed are affected by the low overlap rate of multiple perspectives,which makes them unable to meet the requirements of low overlap rate point cloud registration.This paper conducts in-depth study on low overlap rate point cloud registration methods and investigates the 3D reconstruction technology based on multi-view low overlap rate point cloud and its application in the 3D reconstruction of subway vehicle body point clouds.The main research contents are as follows:(1)The noise characteristics of the vehicle body point cloud are analyzed,and the statistical filtering algorithm is selected for denoising of the vehicle body point cloud through comparative experiments.In order to solve the problem that the traditional voxel filtering algorithm may lead to the loss of surface feature points and edge contour points in the down-sampling process of point clouds,an improved down-sampling method based on voxel filtering is put forward,using the actual point nearest to the voxel gravity center as the down-sampling point for the voxels that meet the constraint of minimum-number of points,and its effectiveness and accuracy are verified through experiments.(2)To address the shortcomings of point cloud registration algorithms in low overlap rate point cloud registration,the principle of the traditional Tr ICP algorithm and its problems in practical applications are studied profoundly.An automatic calculation method for point cloud overlap rate is offered,aiming to reduce the human estimation error of the overlap rate before point cloud registration and to reduce the impact of the predicted overlap rate on the registration process and results;In order to improve the registration accuracy and speed,the point cloud registration algorithm based on variable sequence-length least trimmed squares,i.e.,VSLTS-ICP algorithm,has been studied and proposed.Experiments have shown that for low overlap point cloud registration,the processing accuracy and speed of this method have been significantly improved.(3)A consecutive adjacent frames stitching strategy using a camera as a processing unit is presented to solve the problems of multi-view vehicle point cloud stitching.This strategy achieves fast and accurate stitching of multi-view vehicle point clouds,and achieves surface 3D reconstruction and visualization of vehicle point cloud models through methods such as point cloud refinement,point cloud smoothing,and surface reconstruction.This study provides a new method and idea for the research on 3D reconstruction based on multi-view low overlap point clouds,and has good application potential in practical engineering.