Research On Three-Dimensional Feature Description Method And Application

Three-dimensional(3D)feature description is important for intelligent perception of 3D scenarios.It is also the fundamental research of pattern recognition and 3D vision,that widely used in industrial measurement,medical diagnosis,heritage conservation,robot and automatic driving.With the development of the sensing technology,the 3D application scenarios become more and more rich and complex,which makes the 3D data massive and diversified.The 3D vision tasks also develops from 3D reconstruction to the high-level perception.This makes the traditional 3D feature description no longer suitable for the current technological development in terms of accuracy,efficiency and high-level semantic representation.In view of the main challenges of this research,the relevant theories,methods and applications are studied in this academic dissertation.Focusing on the multi-level robust 3D feature structural representation in complex scenarios,this research studies how to efficiently extract 3D local geometric features,object features and semantic features for scene perception.The 3D feature description involves the three levels of computer vision,i.e.,the pointwise geometric description of the low-level vision,the object-wise reconstruction of the middle-level vision,and the semantic perception of high-level vision.This dissertation carries out related research from these levels successively.The main contents are as follows:(1)A local height image feature description method was proposed to solve the problem of high complexity of the traditional 3D space meshing methods.In this method,the local height images are efficiently generated through the inner product.In order to improve the accuracy of the descriptor under complex and noisy conditions,the method is optimized by discretization,mean calculation and Gaussian filtering.This method achieves a more compact 3D geometric feature description in 2D space,without causing information loss.The experimental results show that the proposed method is close to the state-of-theart method in accuracy,but about 7 times more efficient,achieving a balance between the efficiency and accuracy.(2)With regard to the low efficiency of the Local Reference Frame(LRF)estimation and the partial invalid feature in 2D projection methods,a simplified LRF and a wellconditioned optimization theory are studied,respectively.The method improves the efficiency of the LRF estimation by simplifying the covariance matrix and reducing the calculation range.The capability of the feature description is strengthened by nonlinear weighting the original ill-conditioned feature space to make it closer to well-conditioned.Based on these optimization methods,a weighted height image descriptor is proposed from the local height image descriptor.Compared with the traditional methods,the LRF estimation and information coding in this method are no longer independent but integrated,such that the efficiency is improved.In addition,the enhanced descriptive capability also reduces the dimension of the descriptor and makes it more compact.The experimental results show that the proposed weighted height image descriptor not only outperforms the existing methods in efficiency and accuracy,but also improves the compactness by 6.89 times compared with the state-of-the-art method,achieving the synchronous improvement of the efficiency,accuracy and compactness.(3)To solve the problem of the object feature description in 3D reconstruction,this research proposes a coarse-to-fine point cloud registration method based on the weighted height image descriptor and a 3D reconstruction algorithm based on a single-view point cloud.The point cloud registration method uses the local feature to achieve a fast and robust coarse registration.Then,the iterative closest point algorithm is used for the fine registration.The proposed 3D reconstruction algorithm utilizes the above registration method to preprocess the original point cloud.Then,the method combines the processes of the computer-aided design model processing,surface reconstruction and edge reconstruction to achieve the single-view 3D reconstruction.The proposed algorithm addresses the issue that the traditional methods cannot reconstruct the occlusion area.This research takes the 3D pantograph reconstruction as an example to verify the efficiency and accuracy of the proposed method.It takes less than 20 seconds for 3D pantograph reconstruction and detection.The reconstruction error is about 0.2 mm.The measurement error is less than 0.7 mm,meeting the engineering requirements and showing the application value.(4)As for the high-level semantic feature representation of 3D objects,a point-tosurface representation based point cloud feature description method is proposed.This method utilizes deep learning network to learn a series of global reference surfaces,and transforms the traditional Euclidean coordinates representation into the relative relationship between the point and the global reference surface,i.e.,the point-to-surface representation.This representation can transform the point coordinates to the geometric information.Compared with the traditional point cloud,depth image or voxel representation,this method directly represents the local and global geometric characteristics.Through the solution uniqueness of the linear equations,it is proved that the representation will not cause the loss of the effective information.In addition,this method is more consistent with the mechanism of human perception of 3D objects.The point-to-surface representation is easy to implement and can be embedded into the existing 3D deep learning network by plug-and-play for point cloud classification,segmentation and scene understanding.The experimental results show that the method achieves the best performance in both the public ModelNet classification dataset and ShapeNet segmentation dataset,which indicates that the point-to-surface representation is helpful to the global semantic feature description.