多层级特征组合约束的车载LiDAR点云路侧目标分类方法

Roadside objects such as light poles, trees and traffic signs are important components of the urban infrastructure database and classification of roadside object from mobile laser scanning data has been a popular research topic. At present, there are some problems detecting object in overlapping scenes, and the object classification accuracy is low in point cloud data with accuracy variations. This paper : (1) Proposes a clustering-based segmentation algorithm that takes into account spatial density distribution feature and spectral feature to achieve accurate extraction of roadside objects in overlapping scenes; (2) After component disassembly based on vertical region grow, we propose a unified feature representation model that combines object shape distribution feature, component distribution feature and topology feature of components, through which accurate expression of the object's shape feature and topology feature is achieved; (3) Considers the combination of different feature constraints, construct feature similarity measure function and obtain fine classification result in point cloud with accuracy variations. We are intent to build a “data segmentation - multi-level feature detection and topology representation - matching-based classification ” paradigm to achieve automatic and accurate object classification result in complex urban environment. This research will enrich the methods in the area of object classification from mobile laser scanning data, and provide an important theoretical basis for the city infrastructure database construction and update.

路灯、行道树、交通标志牌等路侧目标是城市基础设施数据库的重要组成部分，基于车载LiDAR点云的路侧目标分类已成为当前研究热点。目前研究中存在粘连重叠目标分割困难、点云精度不一致场景下目标自动精细分类精度低问题，本研究：⑴提出一种顾及空间密度分布特征和光谱特征的聚类分割算法，实现粘连重叠路侧目标的精确分割；⑵采用垂直区域增长法将目标拆解为多个部件，并提出一种目标、部件形状分布特征和部件拓扑关系特征组合的多层级特征描述方法，建立多场景下各类路侧目标的统一特征描述模型；⑶提出多层级特征组合约束的目标相似度计算方法，实现针对点云精度不一致场景的目标精细分类。形成“数据分割-多层级特征提取与拓扑描述-匹配分类”的完整技术链条，实现复杂环境下多类型路侧目标自动分类。研究成果能够丰富车载LiDAR点云目标分类研究的方法体系，为城市基础数据库建立和更新提供理论基础和技术保障。

路灯、行道树、交通标志牌等路侧目标是城市基础设施数据库的重要组成部分，而当前的研究中存在粘连重叠目标分割困难、点云精度不一致场景下目标自动精细分类精度低问题. 本研究以实现复杂环境下多类型路侧目标自动分类为目标,(1)提出了一整套路侧设施的提取与分类框架;(2)提出了一种顾及空间密度分布特征和光谱特征的聚类分割算法，实现了粘连重叠路侧目标的精确分割;(3)采用垂直区域增长法将目标拆解为多个部件，并提出了一种目标、部件形状分布特征和部件拓扑关系特征组合的多层级特征描述方法，建立多场景下各类路侧目标的统一特征描述模型;(4)提出了多层级特征组合约束的目标相似度计算方法，实现针对点云精度不一致场景的目标精细分类。研究共形成了多项成果,包括9篇论文,3个发明专利,2个软件著作权,并形成了项目转化.综上,本课题研究成果丰富了车载LiDAR点云目标分类研究的方法体系，体现了重要的科学意义和应用意义.