Synthetic railroad level crossing point clouds

The purpose of this research is to explore methods for generating synthetic point clouds that can be used to train neural networks. This is tested using level crossings between roads and railroads. The original point clouds were captured by railborne mobile laser scanning and the 3D mesh geometries come from an object library provided by the Swedish Transport Administration. The findings show that a neural network trained using point clouds created by rearranging objects into new scenes outperforms a network trained using conventional scene augmentation. Networks trained using the semi- and fully synthetic point clouds do not reach the baseline of conventional augmentation, but there could still be benefits to the approach. The main challenges of synthesizing point clouds consists of creating appropriate radiometric profiles and creating sufficient geometric variation for the object geometries.

This repository contains the code and the point cloud samples used to generate the data sets used in the study. The data consists of the original point clouds broken down into individual objects and point clouds created from 3D mesh geometries. The code is used to assemble ground truth data, create augmented copies of the original crossings, generate new crossings by semi-randomly assembling point cloud components, and creating synthetic crossings using geometries from 3D mesh models. For information about how to use the code, see readme.md.

本研究旨在探讨生成合成点云的方法，以便用于神经网络的训练。该研究通过道路与铁路交叉口的水平穿越进行验证。原始点云由车载移动激光扫描捕捉，而三维网格几何形状则来自瑞典交通管理部门提供的目标库。研究结果表明，使用重新排列成新场景的对象创建的点云训练的神经网络优于使用传统场景增强训练的神经网络。使用半合成和全合成点云训练的神经网络未达到传统增强的基线，但该方法可能仍然具有潜在优势。合成点云的主要挑战在于创建适当的辐射剖面以及为对象几何形状生成足够的几何变化。本仓库包含用于生成研究中所用数据集的代码和点云样本。数据包括分解为单个对象的原有点云以及从三维网格几何形状创建的点云。代码用于组装地面真实数据、创建原始交叉口的增强副本、通过半随机组装点云组件生成新的交叉口，以及使用三维网格模型的几何形状创建合成交叉口。有关如何使用代码的信息，请参阅readme.md。