3D point cloud semantic classification is an important task in robotics as it enables a better understanding of the mapped environment. This work proposes to learn the long-term stability of the 3D objects using a neural network based on PointNet++, where the long-term stable object refers to a static object that cannot move on its own (e.g. tree, pole, building). The training data is generated in an unsupervised manner by assigning a continuous label to individual points by exploiting multiple time slices of the same environment. Instead of using discrete labels, i.e. static/dynamic, we propose to use a continuous label value indicating point temporal stability to train a regression PointNet++ network. We evaluated our approach on point cloud data of two parking lots from the NCLT dataset. The experiments' performance reveals that static vs dynamic object classification is best performed by training a regression model, followed by thresholding, compared to directly training a classification model.

翻译：三维点云语义分类是机器人领域的一项重要任务，因为它能够帮助更好地理解所建图的环境。本文提出利用基于PointNet++的神经网络学习三维物体的长期稳定性，其中长期稳定物体指无法自主移动的静态物体（如树木、电线杆、建筑物）。训练数据以无监督方式生成，通过利用同一环境的多个时间切片为单个点分配连续标签。不同于使用离散标签（即静态/动态），我们提出采用表示点时间稳定性的连续标签值来训练回归型PointNet++网络。我们在NCLT数据集中的两个停车场的点云数据上评估了该方法。实验性能表明，与直接训练分类模型相比，先训练回归模型再进行阈值分割的方法在静态与动态物体分类上效果最佳。