Research On RGB-D Point Cloud Registration Method Based On Mixed Feature Representation

Three-dimensional reconstruction of real objects is a hot research topic in the field of computer vision and computer graphics.It has a wide range of applications in reverse engineering,three-dimensional animation,digitized cultural relics protection and medical image processing.When performing 3D reconstruction of real objects,it is necessary to scan multiple views to obtain the surface data of objects.Therefore,point cloud registration with different views is an important step in 3D reconstruction.With the development of science and technology,the 3D scanning equipment is continuously improved,not only acquiring the depth information of the real object but also obtaining the color information of the object.Therefore,the research on the registration of RGB-D point cloud is great significance in the field of 3D reconstruction.For the coarse registration of RGB-D data,this dissertation proposes the RGB-D data registration method based on mixed features.Firstly,the curvature of two-point clouds is defined as geometric features,and the color features of two-point clouds are calculated.The geometric features and the color features are weighted to form the mixed features.Secondly,the feature points are obtained from the source cloud according to the mixed features,and the feature vectors are formed from the mixed features and the normalized RGB values.Corresponding points are searched according to the feature vectors of the feature points to form the corresponding point pairs.Finally,the distance of the corresponding point pairs is calculated in the feature vector space.The distance set is formed and the average distance is calculated.And the point pairs are left whose distance are less than the average distance.Then,the remaining feature points are implemented partition.Every partition is registered and the registration transformation of each block is applied to the entire source point cloud.The rigid transformation with the smallest error is used as the optimal transformation of the coarse registration and the coarse registration is completed.In fine registration phase,firstly the RGB color values are converted to grayscale values.Then the range of grayscale values are mapped to the range of the geometric data by linearly transformation.And Four-dimensional vector is composed of the mapped gray value and the geometric information of the point cloud.Then the feature points are obtained according to the mixed features.The corresponding points are searched in 4D vector space by k nearest neighbor algorithm.And we define a 4D Euclidean distance ICP algorithm that uses the 4D-ICP iterative process to achieve fine registration.Experimental results show that the proposed method is fast and effective.The coarse registration method of the partition registration based on mixed features is compared with that of LFSH method.The proposed method can achieve good registration performance for different models and point clouds with different degrees of overlap.The 4D-ICP fine registration method for RGB-D data achieve registration of RGB-D data.The proposed method combined the color information with the traditional ICP method.Compared with other improved ICP methods,the proposed fine registration method is more effective in registration accuracy and color texture preservation.Therefore,the point cloud registration method in this dissertation,whether it is coarse registration or fine registration,has a good registration effect in the registration process of RGB-D data.