Research On 3D Point Cloud Stitching Method Based On Multi-view Camer

In the trend of increasingly diverse 3D vision applications such as autonomous driving,remote sensing mapping and medical 3D imaging,3D data processing technology has attracted much attention.Due to the limitation of the geometry of the object and scanning equipment,usually,the 3D measurement technology cannot complete the overall morphology measurement of the object at one time.Therefore,to establish a complete 3D model of the object,the 3D point cloud data of the object acquired by the sensor needs to be collected from multiple views and then unified under the same coordinate system through registration processing.Due to the limitations of the actual acquisition conditions,there are problems such as missing,large deflection angle,or low overlap rate between point clouds from different views.Therefore,this thesis investigates the registration method of human point clouds with a low overlap rate and verifies the effectiveness and feasibility of the method through experiments.In this thesis,point cloud acquisition and registration experiments are conducted with real human subjects to explore the registration method of multi-view point clouds.To collect multi-view human body point cloud data,a multi-view 3D point cloud acquisition hardware system is built and point cloud acquisition experiments are carried out to obtain data-rich point clouds from multiple views.For the characteristics of low overlap between point clouds and missing features,ISS and NARF methods are used for feature point extraction,and for the characteristics of obvious curvature changes at the overlapping regions of point clouds,a kind of curvature-sensitive saliency constraint feature points are added to extract,and the three feature points of the original point clouds under different view are taken and set separately as their composite feature point sets.To screen out the feature points in the non-overlapping part of the point cloud from different views and reduce the data processing volume,the extracted composite feature point set is taken as the intersection set in the k-neighborhood as the common feature and used as the input point set in the point cloud registration step.Based on the point-to-plane ICP method,the normal constraint is added to the matching point pair relationship to filter the noisy point pairs and improve the registration accuracy,and the corresponding point cloud registration transformation relationship is obtained by using the extracted feature point set as the input for registration experiments using the improved registration method in this thesis.The data volume of the point cloud collected by the built hardware system is about 400,000 points per view,and the experimental results show that the composite feature points extracted by the improved feature point extraction method are rich in feature information,which can fully and effectively represent and simplify the original point cloud,reduce the data processing volume,and improve the running efficiency;the improved point cloud registration method in this thesis has experimented on the low overlap human body point cloud and Bunny data set,and the Compared with other methods,its root-mean-square error is the smallest and the running time is shorter,0.073 m and 0.119 s for human data and 2.182 mm and 23.024 s for Bunny data,respectively,with relatively better comprehensive performance,which confirms the effectiveness and feasibility of the proposed method for low overlap rate 3D point cloud registration and meets the requirements of application scenarios.