基于深度学习的城市场景点云数据分类识别与建模

Reconstructing 3D digital model of urban scene has a wide spectrum of applications, such as urban planning and surveillance, intelligent transportation, 3D maps, virtual battlefields, the immergence processing of natural calamities and terrorist attack, the simulation of pollutant diffusion and so forth. In order to explore the ability of the depth learning architecture on feature capture, semantic segmentation, category annotation and scene layout inference in the point cloud geometric processing, and to further address the intractable issues within the state-of-the-art reconstruction methods concerning computational efficiency, accuracy and interaction, we aim to study large-scale data-driven deep learning algorithms for intelligent understanding and quality reconstruction methods on urban scenes based on large-scale raw point clouds (e.g., LiDAR or Structure from Motion). The urban scene understanding algorithm is to be designed based on machine learning techniques to automatically recognize the typical targets/objects in urban scene. To reconstruct the main structure of urban building, low rank matrix recovery and subspace clustering techniques are to be deeply investigated, based on which the modeling algorithm is devised. As for the auxiliary structures of urban building, we are to design a symmetry detection algorithm to facilitate the following 3D reconstruction. In terms of non-building targets, we plan to study 3D model matching techniques and component-correspondence-deformation algorithms to accomplish rapid reconstruction. On this basis, the high-quality 3D models of urban scenes can be reconstructed efficiently and automatically. As a consequence, the outcome of this project would extend and enlarge the development and applications of urban digital city in a number of fields, which possesses substantially theoretical significance and practical values.

构建大规模城市场景三维模型在城市规划与监控、智能交通、三维地图导航等方面发挥着重要作用。为探索深度学习架构在点云处理问题中的语义分割、类别标注、场景理解等方面能力以及基于此解决目前城市场景建模方法在效率、精度、自动化程度等方面普遍存在的问题，本项目拟面向城市场景，研究基于大规模点云数据深度学习方法支撑下的场景快速理解技术与高质量建模方法。首先研究基于机器学习的场景自动识别方法，实现场景目标快速识别。针对建筑物主体，研究基于低秩矩阵恢复和子空间聚类技术的快速重构方法，解决数据严重缺失情况下模型精确重构难题；针对建筑物附属结构，研究对称结构自动识别及重构方法，实现建筑物快速、高精度整体重构。针对非建筑物，研究模型匹配技术及组件约束变形方法，快速生成非建筑物的三维模型，最终实现城市大场景快速、高质量整体重构。本项目的成功实施将会进一步拓展城市数字模型在多个领域的发展和应用，具有重要的理论意义。

城市场景建模在城市规划、城市监控、智能交通、自然灾害应急处理等方面发挥着重要作用。然而，由于城市大场景的复杂性和多样性，获得的三维点云数据具有规模庞大、结构复杂、分布不均匀、噪声大且数据缺失严重等特点，给城市场景理解和自动、快速的三维重构带来了巨大的挑战。为此，本项目针对城市大场景，研究了面向大规模点云数据（SfM或LiDAR数据）的场景理解技术，提出了高效、稳定的数据处理方法，提高了点云数据质量，解决了数据缺失情况下的建筑物主体结构自动、精确重构难题，实现了建筑物的快速、高精度整体重构。在此项目支持下，发表论文45篇，在人工智能领域顶级会议CVPR、AAAI、ICCV上发表论文4篇，申请国家发明专利7项，已授权7项。同时，相关理论成果在轨道交通、航空航天、数字城市等方面进行实际应用。