一种可对变形建筑物点云图进行分类的深度学习算法

每天一分钟，带你读遍机器人顶级会议文章

标题：Sampled-Point Network for Classification of Deformed BuildingElement Point Clouds

作者：Jingdao Chen, Yong K. Cho and Jun Ueda

来源：2018 IEEE International Conference on Robotics and Automation (ICRA)May 21-25, 2018, Brisbane, Australia

编译：董文正

审核：颜青松，陈世浪

欢迎个人转发朋友圈；其他机构或自媒体如需转载，后台留言申请授权

摘要

在灾后城市地区运行的搜寻和救援（SAR）机器人需要精确识别物理站点信息来执行导航，建图和操作任务。这可以通过获得环境的3D点云以及从点云数据中识别物体来实现。但是，在灾难救援行动期间遇到的非结构化环境和可能变形的物体使得这项任务变得复杂。当前3D对象识别方法依赖于在合适的条件下获得点云输入，并且不考虑变形，例如异常噪声，弯曲和截断。本文工作使用了深度学习，用于从变形建筑元素的点云中识别3D物体。使用从网格模型数据库中采样来的点云训练分类网络，该网络由堆叠的卷积和直接应用于点坐标的平均池化层组成。提出的方法使用了点排序，重采样和旋转归一化实现了输入变异性的鲁棒。综合变形目标数据集的实验结果表明，该方法在分类精度和计算效率方面优于传统的深度学习方法。

图1 架构设计。

图2 门对象的各种变形。从左到右：（i）网格模型（ii）点云（iii）旋转（iv）扭曲扭曲（v）弯曲（vi）截断（vii）组合变形。

图3 不同建筑设计选择的准确性趋势。

图4 示例3D识别由变形的构建元素产生点云

表1ModelNET40和BIMNET10数据集的分类精度。

表2不同建筑之间的算法效率比较。

表3 识别任务在MODELNET10上输入预处理的灵敏度。

表4 ModelNET10识别中变形类型的敏感性。

Abstract

Search-and-rescue (SAR) robots operating in post-disaster urban areas need to accurately identify physical siteinformation to perform navigation, mapping and manipulationtasks. This can be achieved by acquiring a 3D point cloud ofthe environment and performing object recognition from thepoint cloud data. However, this task is complicated by theunstructured environments and potentially-deformed objectsencountered during disaster relief operations. Current 3Dobject recognition methods rely on point cloud input acquiredunder suitable conditions and do not consider deformationssuch as outlier noise, bending and truncation. This workintroduces a deep learning architecture for 3D class recognitionfrom point clouds of deformed building elements. The classifi-cation network, consisting of stacked convolution and averagepooling layers applied directly to point coordinates, was trainedusing point clouds sampled from a database of mesh models.The proposed method achieves robustness to input variabilityusing point sorting, resampling, and rotation normalizationtechniques. Experimental results on synthetically-deformed object datasets show that the proposed method outperforms theconventional deep learning methods in terms of classificationaccuracy and computational efficiency.