Research On Key Technologies Of Sheep 3D Reconstruction Based On 3D Point Cloud

The body size parameter of sheep is the main index to measure the growth and development of sheep.According to the body size parameter,the growth and development characteristics,production performance and genetic characteristics of sheep can be evaluated.The traditional sheep body parameters are measured by using tools manually,which requires a large amount of work and is prone to stress reaction.The three-dimensional reconstruction model of sheep can be used to realize non-contact measurement of sheep body size parameters,it can provide accurate three-dimensional models for sheep digitization,and provide a solid foundation for sheep body evaluation,health assessment and three-dimensional visualization.In this thesis,the key technologies of three-dimensional reconstruction of sheep based on three-dimensional point cloud are studied.Through point cloud normal estimation,feature point detection,segmentation and registration,a complete three-dimensional sheep point cloud data model is obtained,and on this basis,accurate and rapid three-dimensional reconstruction of sheep point cloud is carried out.It mainly includes the following aspects.(1)In order to solve the problem of inaccurate estimation of normal vector at the edge or detail features of point cloud by traditional methods,principal component analysis,hierarchical clustering and iterative least square method are used to estimate the normal vector of point cloud.The initial normal vector of point cloud is calculated by principal component analysis;Gaussian mapping and hierarchical clustering algorithm are used to calculate the characteristic coefficients of each point in the point cloud;The initial normal vector is adjusted by iterative least square method.The experimental results show that the normal vector estimation method based on iterative least squares(ILS)can effectively estimate the normal vector of point cloud with edge or detail characteristics,reduce the root mean square error of the estimation result and the number of wrong points,and effectively suppress the influence of external noise point.(2)Aiming at the problem that the traditional methods can not accurately detect the detailed features in the point cloud and can not reflect the real information of the object,this thesis proposes point cloud feature point detection method based on hierarchical clustering(HCFPD).The minimum spanning tree and depth first traversal algorithm are used to adjust the direction of the normal vector of each triangle formed by each point in the point cloud and its neighborhood points;Non feature points and candidate feature points in the point cloud model are detected by Gaussian mapping of normal vector;For candidate feature points,hierarchical clustering algorithm is used to judge whether they are feature points.The experimental results show that point cloud feature point detection algorithm based on hierarchical clustering can accurately detect the feature points in the feature area in the scattered point cloud data,and can also effectively detect the feature points with unclear details.(3)Aiming at the problems of over segmentation or under segmentation and noise sensitivity in traditional region growing methods,this thesis proposes two point cloud segmentation methods.1)Point cloud segmentation method based on improved region growing(IRGBM).Firstly,find the best point subset of each point in the point cloud model;Secondly,use the best point subset combined with Mahalanobis distance to remove the outliers in the local neighborhood,and use principal component analysis to calculate the normal vector of each point.Finally,the improved region growing method is used to segment the sheep point cloud and extract the sheep region data.The experimental results show that IRGBM can accurately segment sheep point cloud and remove the external noise point.The segmentation accuracy rate of noise free point cloud data is 0.9950,the segmentation accuracy rate of noisy point cloud is 0.9847,and the noise removal rate is 0.982.2)Point cloud segmentation algorithm based on mean shift clustering(MSCM).The normal vector of sheep point cloud is calculated by iterative reweighted least square method;The k-means algorithm is used to divide the point cloud data initially,and the Gaussian projection of the normal vector is used to project the point cloud onto the Gaussian sphere;The points on the Gaussian sphere are clustered by Gaussian kernel mean shift algorithm to realize the segmentation of sheep point cloud,and accurately separate the sheep region in the point cloud from other data.The segmentation accuracy rate of two different types of sheep point cloud are 0.9963 and0.9981 respectively.(4)Aiming at the problems that the point cloud registration algorithm is easy to be affected by noise,external points and sampling rate,and the ICP algorithm has high requirements for the initial position of point cloud,a point cloud registration method based on feature matching is proposed.The covariance matrix descriptor is used to describe the feature points in the point cloud,the improved nearest neighbor distance ratio method is used to match the initial features of the feature descriptor,and the distance constraint is used to remove the mismatched point pairs,so as to optimize the corresponding relationship between the feature points in the matching pair,and the transformation matrix is calculated by the quaternion method to complete the rough registration of point cloud.Finally,the improved ICP algorithm is used to further optimize the results of rough registration.The experimental results show that using the covariance matrix to describe the characteristics of key points can accurately complete the registration of point cloud,and also has a good registration effect for the point cloud model with noise,external points and low sampling rate.(5)Aiming at the problem that the 3D reconstruction algorithm is time-consuming and sensitive to noise and cannot obtain smooth surfaces,this thesis proposes a 3D reconstruction method based on projection triangulation(PTRM)for noisy point clouds.A smoother manifold surface is obtained by moving least squares(MLS)point cloud smoothing,and a more accurate surface normal vector is obtained.The surface reconstruction of sheep point cloud is completed using projection triangulation.The experimental results show that MLS smoothing can effectively improve the ability of the reconstruction algorithm to deal with the noise point cloud,and the PTRM method in this thesis significantly improves the reconstruction speed.