ABSTRACT

The automatic definition of the ground from 3D point clouds has been a common process for the last two decades, with many different approaches and applications that can be found in a vast literature. This paper presents a comparison of three different methodological concepts for ground classification, in order to establish the advantages and drawbacks of each method. First, a heuristic method, based on previous knowledge of the geometry and context of the 3D data. Secondly, a Deep Convolutional Network based on SegNet that classifies 2D images generated from the 3D point cloud. Finally, the third method applies a Deep Learning classification based on PointNet, which takes 3D points directly as inputs. To validate each method and compare them, public and labelled point clouds from the Actueel Hoogtebestand Nederland dataset are employed. Furthermore, the three methods are validated against the ISPRS 3D Semantic Labeling Contest benchmark. The results obtained show that the deep learning-based approaches outperform the heuristic method, with F-scores above 96%. The best results were obtained using a shallower version of SegNet, with F-score above 97%.

摘要

在过去的二十年中，从3D点云自动定义地面一直是一个常见的过程，在众多文献中可以找到许多不同的方法和应用。本文对三种不同的地面分类方法论概念进行了比较，以确定每种方法的优缺点。首先，一种基于3D数据的几何和上下文的先前知识的启发式方法。其次，基于SegNet的深度卷积网络对从3D点云生成的2D图像进行分类。最后，第三种方法应用了基于PointNet的深度学习分类，该分类直接将3D点作为输入。为了验证每种方法并进行比较，可以使用Actueel Hoogtebestand Nederland的公共点云和标记点云使用数据集。此外，这三种方法均已根据ISPRS 3D语义标记竞赛基准进行了验证。获得的结果表明，基于深度学习的方法优于启发式方法，F分数超过96％。使用较浅版本的SegNet（F分数高于97％）可获得最佳结果。