基于压缩感知的点云数据编码与重建

发布时间：2018-04-17 17:43

本文选题：压缩感知 + 点云数据　；参考：《北京工业大学》2014年硕士论文

【摘要】：三维激光扫描设备的飞速发展已经使得三维点云数据成为多媒体数据非常重要的组成部分。然而，随着三维扫描设备精度的不断提高，通过扫描设备所获取的三维点云数据也越来越大、越来越复杂，这给三维点云模型在网络资源有限的情况下，对其存储、传输、处理带来了很大困难，因此，关于三维点云数据的高效压缩编码方案一直被国内外学者广泛关注。近几年，Candes、Donoho等人提出的压缩感知理论（Compressive Sensing/Compressed Sampling，CS）指出对于稀疏或者在某变换基下稀疏的信号可以对信号采用非线性下采样的方法来进行观测，利用低维的观测结果可以用与变换基满足非一致性的观测矩阵来对原始信号进行高概率精确重建。区别于传统的奈奎斯特采样定理，压缩感知理论结合信号的稀疏特性，利用观测矩阵来对信号进行观测，从而使得信号的采样过程不依赖于信号的带宽，而是信号的内容和结构。因此，压缩感知理论为多媒体信号的压缩编码开辟了一条崭新的途径。本文从压缩感知理论的最新成果出发，利用三维点云数据局部空间的相似性，建立了三维点云数据的规格化方法，提出了基于过完备字典的点云数据稀疏表示模型和编码、重建模型。具体完成的工作如下：第一，通过研究三维点云模型的几何空间特性以及局部相似特性，提出了一种基于K近邻的点云数据规格化方法，该方法有效的利用了点云数据的局部空间相似性，提高了三维点云数据在坐标数值上的相似性，为三维点云数据的稀疏表示提供了重要保证。第二，考虑到规格化后的点云数据之间具有自相似性，因此，本文首先提出基于K-SVD的字典训练算法来获得规格化点云数据的稀疏表示基，使得规格化后的点云数据能够在过完备字典下稀疏表示。但是，，传统的过完备字典训练算法不能很好的适用于三维点云数据，因此，本文结合点云数据的空间几何特性，提出一种基于K均值的三维点云数据过完备字典训练算法，为基于压缩感知的点云数据编码与重建奠定了基础。第三，以压缩感知相关理论为指导，在点云数据规格化以及信号稀疏表示基础下，本文针对过完备字典稀疏表示的三维点云数据，提出了基于随-维1点云数据观测编码，方法，并且提出了基于1机观测的三范数最小化重建模型，以及基于TV方法的重建模型实验结果表明，本文所述的三维点云数据处理方法具有良好的结果，为基于感知的三维点云数据编码与重建提供了一条崭新的思路，具有较高的创新性和实用价值。
[Abstract]:With the rapid development of 3D laser scanning equipment, 3D point cloud data has become a very important part of multimedia data.However, with the continuous improvement of the accuracy of the 3D scanning equipment, the 3D point cloud data obtained by the scanning device is becoming larger and more complex, which gives the 3D point cloud model storage and transmission in the case of limited network resources.Therefore, the efficient compression and coding scheme of 3D point cloud data has been widely concerned by domestic and foreign scholars.In recent years, Compression Sensing/Compressed sampling theory proposed by Candesl Donoho et al., points out that the signal can be observed by nonlinear down-sampling for sparse signal or sparse signal based on a transform basis.The low dimensional observation results can be used to reconstruct the original signal with high probability and precision by using an observation matrix which is not consistent with the transform basis.Different from the traditional Nyquist sampling theorem, compression sensing theory combined with the sparse characteristics of the signal, using the observation matrix to observe the signal, so that the signal sampling process does not depend on the bandwidth of the signal.It is the content and structure of the signal.Therefore, the theory of compression sensing opens up a new way for the compression and coding of multimedia signals.Based on the latest achievements of compression perception theory and the similarity of local space of 3D point cloud data, the normalization method of 3D point cloud data is established in this paper, and the sparse representation model and coding of point cloud data based on overcomplete dictionary are proposed.Reconstruct the model.The work accomplished is as follows:First, by studying the geometric spatial characteristics and local similarity of 3D point cloud model, a point cloud data normalization method based on K-nearest neighbor is proposed, which effectively utilizes the local spatial similarity of point cloud data.The similarity of coordinate values of 3D point cloud data is improved, which provides an important guarantee for sparse representation of 3D point cloud data.Secondly, considering the self-similarity between normalized point cloud data, this paper first proposes a dictionary training algorithm based on K-SVD to obtain sparse representation base of normalized point cloud data.The normalized point cloud data can be represented sparsely in an overcomplete dictionary.However, the traditional algorithm of over-complete dictionary training is not suitable for 3D point cloud data. Therefore, combining the spatial geometric characteristics of point cloud data, this paper proposes an algorithm for over-complete dictionary training of 3D point cloud data based on K-means.It lays a foundation for point cloud data coding and reconstruction based on compressed perception.Thirdly, under the guidance of the theory of compressed perception, based on the normalization of point cloud data and the sparse representation of signals, this paper proposes the observation coding based on in-dimension 1-point cloud data for over-complete dictionary sparse representation of three-dimensional point cloud data.In addition, a tri-norm minimization reconstruction model based on 1-machine observation and a reconstruction model based on TV method are proposed.The experimental results show that the 3D point cloud data processing method presented in this paper has good results and provides a new way of thinking for perceptual 3D point cloud data coding and reconstruction. It has higher innovative and practical value.
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.7

【参考文献】