融合词袋模型与深度学习结构的低空LiDAR 点云自动分类研究

The low-altitude UAV laser scanning system could obtain the mass multi-sourced data in urban, rural areas and even mountainous areas, whereas the recognition and extraction of feature information cannot be automatically identified due to the lack of the ability to describe the three-dimensional target structure. In order to address the problem, this issue explores an automatic precise classification method of the fusion data between low-level LiDAR point and images based on the deep learning. At the stage of precise classification, the high-precision registration between point clouds and images is accomplished based on the collinear equation expansion model，combing with the properties and the spatial relationship between of the conjugate road primitive. On this basis, the bag of word model (BOWM) is obtained by improving the vocabulary construction and classifier construction based on the concept of the deep learning. In the whole classification process, the training sample learning can be implemented by hierarchical search strategy, and the sample weights are automatically updated. Thus, accurate automatic extraction of similar feature boundaries and small target feature categories in discrete spaces would be effectively solved. This project is supposed to perfect the theory and methodology of the registration and precise classification between the low-level LiDAR and images, based on the viewpoint of photogrammetry and computer vision, thus providing a basis for more elaborate urban (villages and towns) morphology analysis and meeting the needs of city-based (village-based) various application requirements.

本课题针对由于三维目标结构特征描述能力不足而导致地物信息无法自动化精确提取的现状，研究基于深度学习模型的低空无人机激光点云与影像辅助下的场景目标自动精细分类。在精细分类阶段，首先，以共线方程扩展模型为基础，结合共轭道路基元自身属性及其之间的空间关系，实现低空点云到影像的高精度配准；在此基础上，借鉴深度学习思想，分别从词汇表构造与分类器构建两方面进行改进完善词袋模型，整个分类过程可以分层次实现训练样本的学习，且样本权重自动更新，有效解决点云离散空间内相似地物边界、小目标地物类别的精确自动提取问题。本课题从摄影测量和计算机视觉的角度，完善多种地形下低空LiDAR点云与影像的高精度配准和地物精细分类新理论与新方法，为更精细的城市（村镇）形态分析与规划建设提供基础，满足基于城市（村镇）形态的各种应用需求。

本项目面向低空无人机激光系统获取的海量激光点云和影像数据进行快速而精确解译的需求，针对低空无人机系统的激光点云和影像的配准难度大、分类算法效率低、点云数据分割的自动化潜力开发不充分等问题开展研究。首先，设计并实现了基于道路基元与房屋边界约束的点云与影像融合算法，探索可输入深度学习网络的点云特征数据集；然后，以提高语义分割精度为目标，基于PointNet网络提出了一种融合多尺度多特征的MSMF-PointNet地物分类分割算法，以弥补原始PointNet网络对点云特征利用不足、缺少局部特征的问题；最后，优化基于深度学习的点云语义分割效率，借鉴室内场景语义分割网络的优势，结合机载LiDAR点云的特性，探索了以最优特征融合光谱信息聚合模式下基于RandLA-Net模型的复杂场景地物精细分类方法，解决了大规模大尺度机载点云基于深度学习网络的语义分割效率与精度的平衡问题。本项目在解决低空无人机载激光点云与影像融合数据下不同场景目标精细分类算法过程中，分别采用了国际摄影测量与遥感学会三维语义标号数据集、慕尼黑技术大学（ TUM ）和同济大学于2020年发布的LASDU数据集以及自建数据集（数据为真实自采数据）进行验证，实验证明了本项目方法的有效性。本项目执行情况良好，发表学术论文4篇，均为SCI期刊论文；申请国家发明专利1项，已公示；授权软件著作权1项；荣获省部级科技进步奖一等奖2项；培养研究生4名，其中已毕业1名，已超额完成立项预期目标。