Towards a Software System for 3D Modeling of Urban Road Environments using Mobile Laser Scanning Data

移动的激光扫描（MLS）系统是一种新兴技术，在关键基础设施库存、智能交通系统、应急响应规划和智慧城市发展等方面具有巨大的应用潜力。然而，如何直接从三维MLS点云数据中提取几何和语义信息是一个非常具有挑战性的任务，仍然是一个悬而未决的问题。本研究解决了这两个问题。一种是由原始MLS点云中对象的不完整性、重叠性、遮蔽性和相似性引起的，所述原始MLS点云覆盖具有高水平遮挡和杂乱的复杂城市道路场景。另一个是由于原始MLS点云的大体积、混合密度和空间离散性的特点导致的高计算复杂度。受这些挑战的激励，本研究将有助于MLS技术，进一步开发一个软件系统，能够处理大体积，混合密度的MLS点云。该研究团队将开发创新的算法和软件工具，用于点云分类，对象提取和变化检测，所有这些都在3D对象空间中进行。通过利用深度学习的最新进展，特别关注开发基于深度卷积神经网络的3D点云分类算法及其在基于GPU（图形处理单元）的并行处理环境中的实现。拟议的项目还通过将深度玻尔兹曼机和霍夫森林模型与可见性估计相结合来开发3D对象提取器。覆盖相同道路场景的至少两个点云可以由不同的MLS系统在不同的时期采集，用于检测不同时期之间的几何变化。3D变化检测将通过基于新的局部特征描述符的成对点云配准以及随后的鲁棒多尺度张量投票来实现。研究成果将提升加拿大在移动的激光扫描技术方面的全球领导地位，不仅在硬件方面，而且在软件方面。拟议的软件系统将克服目前点云处理系统的缺点，并将提供更多的定量报告，随着时间的推移，在城市道路环境的变化。报告制度还将导致及时和有效地维护加拿大的运输基础设施。*****

Mobile laser scanning (MLS) systems are a new and emerging technology that has demonstrated the great potential applications in critical infrastructure inventory, intelligent transportation systems, emergency response planning, and smart city development. However, how to extract geometric and semantic information directly from the 3D MLS point cloud data is a very challenging task and still remains unsolved. This research tackles the two problems. One is caused by the incompleteness, overlapping, sheltering, and similarity of the objects in the raw MLS point clouds covering the complex urban road scenes with high levels of occlusion and clutter. The other is caused by the high computational complexity due to the characteristics of large-volume, mixed-density, and spatial discreteness of raw MLS point clouds. Motivated by these challenges, this research will contribute to the MLS technology by further developing a software system that is able to handle large-volume, mixed-density MLS point clouds. The research team will develop innovative algorithms and software tools for point cloud classification, object extraction and change detection all in the 3D object space. By taking advantage of recent advances in deep learning, a special focus is given to develop algorithms for 3D point cloud classification based on deep convolutional neural networks and its implementation in a GPU (Graphics Processing Unit)-based parallel processing environment. The proposed project also undertakes development of 3D object extractors by integrating deep Boltzmann machines and Hough forests model with visibility estimation. At least two point clouds covering the same road scene that may be acquired by different MLS systems at different epochs are used to detect the geometrical changes between different epochs. The 3D change detection will be implemented through the pairwise point cloud registration based on new local feature descriptors followed by robust multi-scale tensor voting. The research results will enhance Canada's global leadership in mobile laser scanning technology, not only in its hardware but also in its software. The proposed software system will overcome the shortcomings of the current point cloud processing systems and will provide more quantitative reports of the identified changes in the urban road environment over time. The reporting system will also lead to timely and effective maintenance of the Canadian transportation infrastructures. *****