【计算机科学】【2017】点云压缩与低延迟流

本文为美国密苏里大学堪萨斯分校（作者：KARTHIK AINALA）的硕士论文，共37页。

随着三维深度传感器的商品化，我们现在可以很容易地将真实的物体和场景建模成数字域，然后在游戏、动画、虚拟现实、沉浸式通信等领域中广泛应用。现代传感器能够捕捉细节非常精细的物体和大面积的场景，因此可能包括数百万个点。在实时传输的情况下，这些点数据通常占用较大的存储空间或需要较高的带宽。因此，有必要对这些巨大的点云数据进行有效的压缩。

点云通常是由基于八叉树的结构组织和压缩的。八叉树细分序列通常以字节序列化，然后使用距离编码、算术编码或其他方法对其进行熵编码。这种基于八叉树的算法只在一定的细节层次上有效，因为它们在细分层次的数量上具有指数式的运行时间。此外，随着细分层数的增加，压缩效率降低。

在这项工作中，我们提出了一种划分点云数据的替代方法。点云是在数据分割的基础上，用kd树二值分割代替八叉树的空间分割方法，形成一个基础层。在基层叶节点中，根据节点的平面度，考虑点的分布并投影到二维平面上，基于八叉树和四叉树的分区用于进一步将数据转换为位流。这些是可伸缩的点云比特流，因为对于特定的视角，每次只需要特定数量的kd节点。使用案例是自主车辆中的导航，在这种情况下，它需要在不同速度下达到特定距离的点云信息。这些可扩展的kd节点位流可用于低延迟的实时传输。结果表明，对于点云中的几何压缩，压缩性能得到了提高，并且为导航用例提供了一个可扩展的低延迟流模型。

With the commoditization of the 3D depth sensors, we can now very easily model real objects and scenes into digital domain which then can be used for variety of application in gaming, animation, virtual reality, immersive communication etc. Modern sensors are capable of capturing objects with very high detail and scene of large area and thus might include millions of points. These point data usually occupy large storage space or require high bandwidth in case of real-time transmission. Thus, an efficient compression of these huge point cloud data points becomes necessary. Point clouds are often organized and compressed with octree based structures. The octree subdivision sequence is often serialized in a sequence of bytes that are subsequently entropy encoded using range coding, arithmetic coding or other methods. Such octree based algorithms are efficient only up to a certain level of detail as they have an exponential run-time in the number of subdivision levels. In addition, the compression efficiency diminishes when the number of subdivision levels increases. In this work we present an alternative way to partition the point cloud data. The point cloud is divided based on the data partition using kd tree binary division instead of Octree’s space partition method and forming a base layer. In base layer leaf nodes, the distribution of points is considered and projected to a 2D plane based on the flatness of the node points. Octree and Quadtree based partition is used to further convert the data to bitstreams. These are scalable point cloud bitstreams as we need only specific number of kd nodes in each time for a specific point of view. The use case is navigation in autonomous vehicles where it requires point cloud information up to a specific distance at different speeds. These scalable bitstreams of kd nodes can be used in real time transmission with low latency. Results show that compression performance is improved for geometry compression in point clouds and a scalable low latency streaming model is shown for navigation use case.

1 引言

2 项目背景

2.1 MPEG点云压缩PCC

2.2 主成分分析PCA

2.3 基于平面投影近似（PPA）的几何压缩

2.4 本文提出的kd树与PPA逻辑

3 实验与计算细节

3.1 基于PPA的有损几何压缩

3.2 无损几何压缩

3.3 可扩展点云比特流

4 结论

附录A LDPC译码时间估计程序