Research And Application Of Dynamic Human Point Cloud Registration Algorithms Based On Clustering And Quaternion Representation

Point cloud registration of 3D human body[49,50]is one of the most basic problems in generating 3D human body model,and It is a key step of 3D human body reconstruction[51].3D human body data can be widely used in virtual reality,game making,online medicine,clothing design and other fields.If using scanning equipment to obtain 3D human body data is very expensive,complex and inconvenient to carry,depth cameras such as Kinect[1]have become a research hotspot of many scholars in the field of computer vision because of their low cost,portability and easy operationAlthough there are many traditional 3D scanning methods and 3D point cloud registration technology,these traditional methods all face the same problem:the 3D scanning process needs the human body to remain absolutely still,and the computing speed is slow.All of these make them not suitable for dynamic human registration Based on this,this project aims to realize point cloud registration and application of dynamic human body.This method can obtain multi angle human body data at the same time by constructing a multi Kinect 3D scanning environment,and solve the problem of unconscious movement of human body in the process of multiple scanning of camera rotation and translationIn this paper,a method of dynamic human body point cloud registration based on clustering and quaternion representation is proposed.This method uses human body point cloud as processing unit to perform local rigid body registration for dynamic human body data.First of all,based on the point cloud clustering algorithm,a method of clustering human skeleton points is proposed.The correctness of skeleton points should be considered when clustering human skeleton points.We need to correct the occluded skeleton points.Then,local registration is carried out for the dynamic human point cloud after clustering.In the process of registration,the rotation matrix expressed by quaternion is used to reduce the complexity of registration calculation and improve the speed of registration.Using the distributed network,we record the dynamic human body sequence of multi angle continuous frames,and do a lot of experiments to prove the efficiency and accuracy of this algorithm.Experimental results show that the accuracy of this algorithm is better than other methods.The main work and innovations of this paper are as follows:1.In order to improve the accuracy of dynamic point cloud registration,this paper uses human skeleton point data to cluster each human point cloud.In order to ensure the consistency of clustering criteria,the SVD is used to correct the rotation and translation of the occluded skeleton data.For the clustered human point cloud,local registration is performed in space(adjacent frames at the same time)and time(the same frame at the same time).2.In the process of coarse registration of point cloud,the color information and spatial information in the generated human point cloud data are synthetically used in this paper.After getting the corresponding relationship of the adjacent frames by 3D SIFT[2]and K-D Tree,this paper first uses the result of point cloud clustering as the prior information,as the clustering constraint of the corresponding relationship.Then,the color information and spatial information of point cloud are combined as another constraint condition to get the relatively correct coarse registration results.3.In order to improve the efficiency of dynamic point cloud registration,an improved ICP fine registration method is proposed.In the process of dynamic human registration,we use quaternion instead of rotation matrix to calculate the result,which is less time-consuming.For human dynamic data,the clustering results and rotation matrix of the corresponding frame at the previous time can be taken as the initial input of the next frame,which can reduce the number of iterations and improve the registration speed.