结合几何特征和深度神经网络的大范围点云自动分类

Automatic classification of large area point cloud combining geometric features and deep neural network

针对大范围点云自动分类存在点云噪声多以及场景复杂等挑战,本文提出一种结合几何特征的深度神经网络分类方法。在输入阶段,基于结构张量来提取单个点云邻域内的局部几何特征,从而有效保留点云的几何细节,提高特征表示能力;在特征提取阶段,设计了全局上下文聚合模块,该模块对每个位置的特征进行自适应加权,扩大网络感受野,抑制点云噪声。在大范围的室外点云Semantic3D和室内S3DIS数据集上进行了实验。结果表明,本文方法能够有效保留几何细节,减少噪声干扰,提高点云分类精度,在Semantic3D和S3DIS上总体精度分别达到93.60%、87.04%。

It is still a challenge for automatic classification of large-scale point cloud due to much noise and complex scene.In this paper,a deep neural network classification method combined with geometric features is proposed.In the network input stage,this paper extracts the local geometric features in the neighborhood of a single point cloud based on the structure tensor,so as to effectively preserve the geometric details of the point cloud and improve the feature representation ability.In the feature extraction stage,a global context aggregation module is designed,which adapts to the feature weighting of each location,expands the network receptive field and suppresses the noise of the point cloud.This method is tested on large-scale outdoor dataset Semantic3D and indoor S3DIS dataset.The results show that this method can effectively retain geometric details,reduce noise interference and improve the accuracy of point cloud classification.The overall accuracy of Semantic3D and S3DIS data is 93.60%and 87.04%respectively.