Local Feature Based Registration Research For Images And Point Clouds

Registration technology is the process of matching and aligning several images or point clouds acquired by different sensors in different time or under different conditions.It is a very crucial and important problem in the field of computer vision and pattern recognition.It has a wide range of applications in computer vision tasks such as 3D reconstruction,visual synchronization and positioning,target recognition and tracking,stitching,retrieval,etc.The feature-based image and point cloud registration method is mainly based on the feature description of the local area of the key points and the matching of the key points.Feature description aims to generate a high-dimensional vector to represent the neighborhood information of key points.Feature matching determines the correct correspondence relation for the feature point set of the images or point clouds to be registered.In practical application,the robustness of feature description and feature matching is poor and the accuracy of registration is low due to the interference of moving objects,repeated structures,noise and occlusion.In order to solve the aforementioned problem,this paper studies the registration algorithm for image and point cloud,mainly including the following two aspects:(1)In order to assemble a series of local images and get a scene with a wider field of vision,the rotation translation matrix between the local images needs to be obtained.Aiming at the problem that the accuracy and efficiency of the random sampling consistency algorithm are affected by the mismatching in the feature matching set,this paper proposes a mismatching elimination algorithm based on the assumption of spatial consistency.Firstly,the SURF algorithm and the nearest neighbor ratio method are used to calculate the feature matching between images.Secondly,the relative motion slope and relative motion distance were calculated for the preliminary matching point set.DBSCAN algorithm based on density clustering was used to cluster the feature matching point set,and the outliers in the clustering results were eliminated.Finally,the space transformation matrix is calculated to transform the two images to the same plane for fusion.The actual data proves that the proposed algorithm can effectively purify the feature matching point set and improve the registration accuracy.The proposed algorithm is further verified by applying it to the aerial image sequence of UAV.(2)In order to reconstruct the 3D point cloud scene better,it is necessary to study the registration of 3D color point cloud.The essence of point cloud registration is to solve the spatial rotation translation matrix between point clouds to be registered.Aiming at the problem that the traditional point cloud description algorithms seldom use the color information of point cloud,this paper proposes a local descriptor for the color point cloud data,which is applied to the point cloud registration.Firstly,the color point cloud data to be registered is projected into an image.Secondly,SURF algorithm is used to calculate the key points respectively.Then,the convolutional neural network and directional gradient histogram are used to calculate the local features of key points and generate matching point pairs.Finally,according to the corresponding relationship between the pixel and the point cloud data,the rotation translation matrix between the two point clouds is calculated to realize the coarse registration of the point cloud.In this paper,the actual 3D color point cloud data is compared with a variety of registration algorithms to verify the effectiveness of the proposed method.The mismatching elimination algorithm based on density clustering proposed in this paper can be applied to UAV sequence image Mosaic,remote sensing image multi-modal data fusion and other tasks.The 3D color point cloud registration algorithm based on local descriptor can perform 3D reconstruction of 3D color point cloud data in large scenes well.The theoretical and engineering application value of the proposed algorithm is proved by experiments in this paper.