Feature Description For Object Recognition And Reconstruction With Point Clouds

Computer vision has always been a very important topic in the field of artificial intelligence.As two major tasks in computer vision,target recognition and three-dimensional reconstruction are widely used in robotics,object tracking,scene understanding,and navigation guidance.The 2D image processing technology has been widely used due to its advantages such as structured pixels and has achieved considerable development.However,it shows many limitations,such as being susceptible to light and scale changes.Therefore,3D image processing has gained more attention and has become a new popular research topic in the area of computer vision.3D images carry depth information naturally and the acquisition is more and more easy.Point clouds are the most common representation of 3D images.Strengthening the research on point clouds can make the processing of 3D information better.In view of these,this thesis presents in-depth research on point cloud feature description,3D target recognition,and 3D reconstruction.Firstly,the research significance and research status of 3D image processing technology are introduced from three aspects of point cloud description,target recognition,and 3D reconstruction.The advantages and disadvantages of the existing point cloud feature descriptors,recognition,and reconstruction algorithms are summarized.Then,the principles and application scenarios of the acquisition and preprocessing algorithms of point cloud are introduced in detail,including point cloud normal estimation,filtering,and segmentation techniques.Secondly,Frame-SHOT hybrid descriptor is proposed for point cloud feature description.The local sub-feature is described using the well-known Signature of Histograms of Orientation(SHOT)descriptor,and the global sub-feature is encoded by the structural frame points of the object.This algorithm combines the advantages of traditional local and global feature descriptors.The parameter selection and feature descriptor matching experiments on the Bologna dataset are performed to verify the strong robustness and descriptiveness of this method.This Frame-SHOT hybrid description algorithm provides the theoretical basis for the research of the rest chapters in this thesis.Thirdly,a novel 3D object recognition framework based on Frame-SHOT descriptor is proposed.And we compares this algorithm with the existing traditional methods on the Kinect dataset and the Challenge dataset respectively.The validity of this recognition framework and the high performance of the Frame-SHOT descriptor are verified.Finally,in the aspect of 3D reconstruction,a novel reconstruction framework based on the Frame-SHOT descriptor is proposed.According to descriptor matching,the appropriate feature points are selected as corresponding point sets.The registration between two consecutive views point clouds is gradually completed.Through this method,the three-dimensional structure,surface texture,and spatial scale information of the target object are presented.The validity of this algorithm is verified on the public datasets and the point clouds collected by ourselves.