【作者】 张军；

【导师】 李宗民；

【作者基本信息】 中国石油大学(华东) ， 信息与通信工程， 2020， 硕士

【摘要】 随着激光雷达、深度相机等三维传感器在各类三维场景上的广泛应用,基于点云数据的三维场景感知学习越来越受到人们的关注,针对传感器直接输出的三维点云数据的研究也逐渐从底层次的几何特征提取向高层次语义理解过渡。深度学习特别是深度卷积神经网络在特征学习及语义信息提取上有巨大优势,如何设计高效的网络模型并利用点云数据驱动模型进行三维场景感知学习成为最近的研究热点。三维点云的深度学习方法研究主要受以下几个方面的限制:首先,点云受采集设备以及坐标系的影响,点的排列顺序千差万别,因此点云数据难以直接作为深度网络的输入。其次,点与点之间的关联性难以表示,这让基于点云数据的高层语义感知变得更加困难。最后,点云数据包含的信息量有限,只能提供片面的几何信息。本文通过分析点云数据的特点和大量的相关工作,基于这些研究的热点和难点,设计并搭建了一种依赖于邻域聚合框架的端到端的深度网络模型用来处理原始的点云数据。本文主要的研究内容如下:1.对基于点云数据的三维场景感知学习的相关技术进行了充分的调研,并进行分析对比,总结了该领域的发展概况和未来需要面临的挑战,同时对场景感知学习的相关技术原理进行了详细的介绍。2.设计并搭建了一种新的适合原始点云作为输入的图注意力深度网络模型。网络模型将传统的图神经网络作为基础的网络框架,结合自然语言处理中的注意力机制进行局部特征的提取。3.提出描述点云关系的关联度的概念。通过获得的关联度在聚合节点特征时来聚合点云的结构化特征,最后再结合网络中全局特征,以获得更加鲁棒、信息更加丰富的特征描述子,同时在一定程度上能够克服点云的稀疏性问题。4.改进图神经网络中的节点状态更新策略,设计并使用全新的节点更新函数用来更新点云中每个数据点的特征。随着网络的加深,多个图注意力层进行多次特征更新,以便可以获得丰富的节点特征表示和局部结构化的特征。最后,通过在流行的基准数据集上与其它的点云感知算法进行实验对比,验证了本文提出的算法的有效性。更多还原

【Abstract】 With the widespread application of 3D sensors such as Lidar and depth camera in various3 D scenes,3D scenes perception learning based on point cloud has attracted more and more attention,and the research on native point cloud data has also gradually shifted from the bottomlevel geometric feature extraction to the high-level semantic understanding.Deep learning,especially deep convolutional neural networks,has great advantages in feature learning and semantic information extraction.How to efficiently design and utilize data-driven models for perceptual learning of 3D scenes has become a research hotspot.The research progress on deep learning for 3D point cloud which is mainly limited by the following aspects: firstly,due to the influence of acquisition equipment and coordinate system,the order of points in point cloud varies greatly,so point cloud can not be used as the input of deep neural network.Secondly,point cloud is sparse,and the correlation between points is also difficult to be represented.Because these reasons make the high-level semantic perception of point cloud more difficult.Finally,point cloud contains a limited amount of information and it can only provide one-side geometric information.Based on the difficulties of these research and application,recently,some works have found some ideal methods to deal with raw point cloud.By analyzing the characteristics of point cloud and the latest excellent works,this thesis proposes a new end-to-end deep neural network model for handling the raw point cloud.The main research contents of this thesis are as follows:1.The relevant technologies of 3D scene perception learning based on point cloud are fully investigated and compared,the development situation of this field and the challenges in the future are summarized.The relevant technical principles of scene perception learning are introduced in detail.2.The thesis designs and proposes a graph attention depth network model which suitable for original point cloud as input.The model takes the traditional graph neural network as the basic network framework and combines the attention mechanism of natural language processing to extract local features.3.The thesis first proposes the concept of relational degree describing the relationship between points.Through the obtained relational degree,the structural features of point cloud are aggregated when the features of nodes are aggregated.Finally,the structural features combine the global features in the network are to produce more robust and informative feature descriptors.At the same time,the sparsity of point cloud is overcome to some extent.4.The thesis improves node status update strategy in graph neural network.A new node update function is proposed to update the feature of each data point in point cloud.As the network gets deeper,multiple feature updates are carried out in multiple graph attentional layers to obtain rich node feature representations and locally structured features.Finally,the effectiveness and efficiency of the proposed algorithm are verified by comparing the popular benchmark dataset with other point cloud perception algorithms.更多还原