海量遥感点云数据的表达与压缩

With the development of various devices for 3D information capture, e.g. light detection and ranging (LiDAR) sensors, RBG+D and light field cameras, remotely sensed point cloud is experiencing explosive growth. It represents very large volumes of data with challenges in terms of storage, transmission and processing. This project will focus on the issue of representation and compression for massive remotely sensed point cloud, including three aspects as follows. 1) Application- driven simplification for massive point cloud. The captured point cloud data is firstly simplified by removing repeated and unimportant nodes, with respect to the follow-up applications of point cloud; 2) Rate-distortion optimized representation for point cloud. The geometry information of simplified point cloud is represented as a set of depth images. In order to obtain an optimal representation, rate-distortion optimization is used in the rendering task. Based on these depth images, we will propose an alternative graph structure based representation for point cloud, in which the graph edges connect nodes with high correlation; 3) Graph transform based compression. Thanks to the topological connections provided by the graph representation, graph transforms can be directly applied on the point cloud data to exploit the redundancy in transform domain. A traditional coding scheme consisting of quantization and entropy coding is then used for encoding the transformed coefficients to compress the point cloud data.

随着激光扫描（LiDAR，RGB+D）、光场等三维信息获取技术的发展，遥感点云数据的数据量呈现指数级增长，如何高效地表达和压缩这些海量点云数据已经成为点云处理领域亟待解决的问题。本研究将提出基于图结构的点云数据表达与压缩方法，主要研究内容有：1）考虑遥感点云数据海量的问题，根据后续应用场景的具体参数，如视角位置、尺度等，研究海量遥感点云数据的去冗余方法；2）在压缩点云三维空间几何信息方面，本项目研究将点云节点三维信息投影到二维深度图像进行压缩的方法。考虑节点空间分布、局部结构复杂程度等因素，提出基于率失真模型的最优投影方法，并在此基础上构建点云的图结构表达；3）在压缩点云节点内容信息方面，本项目将改进现有基于图结构的信号处理方法（如变换域方法），在变换域挖掘点云数据中的冗余信息，从而对点云数据进行压缩。

针对遥感点云数据量巨大，导致后续点云数据处理以及信息挖掘困难的问题，本项目提出了特征保持的点云精简压缩方法，以较少的信息损失为代价，得到较高的点云数据压缩率，为海量遥感点云的实时处理、自动化解译提供支撑，并将成果直接用于基于激光的机器人SLAM技术中用于减少点云数据量。此外，本项目还将研究成果应用于多源遥感图像数据的目标检测、变化检测、目标跟踪等自动解译任务中，包括基于图结构信号处理的三维灰度共生感知技术、基于图变换的频域注意力机制、多尺度信息融合技术等，取得了较好的效果。在项目资助下，项目组发表科研论文6篇，全部为SCI检索论文；参加国际遥感学术会议1次，项目组骨干人员获得市级人才项目1项，培养研究生9人，其中硕士毕业生2人，在读博士1人，在读硕士6人。