数据融合及图卷积下的深度学习点云分类研究

Research on deep learning for point cloud classification in the context of data fusion and graph convolution

针对机载LiDAR扫描点云因存在噪声、遮挡、高冗余、密度差异大等缺陷,导致深度学习方法难以充分识别局部结构的问题,本研究以三维点云深度学习算法PointNet为基础,结合图卷积模型顾及点云中的局部特征,通过引入空间金字塔池化捕获点云的局部细粒度特征,提出一种适用于机载LiDAR点云分类的深度学习算法。为提高分类精度,融合机载LiDAR点云和多光谱航空影像进行点云信息扩充,同时在点云格网化处理时应用不同尺度进行数据扩增。实验采用国际摄影测量与遥感协会(ISPRS)提供的机载LiDAR点云和多光谱航空影像,融合数据实验结果取得了83.9%的精度,比未融合光谱信息的点云提高了12.1%,同时设计对比实验验证了改进算法的有效性,为城市场景下的点云分类提供了一种思路。

Aiming at the problem that the airborne LiDAR scanning point cloud is difficult to fully recognize the local structure due to the defects such as noise,occlusion,high redundancy and large density difference,based on the deep learning algorithm for 3 D point cloud PointNet,this research proposed a classification algorithm for airborne LiDAR point cloud,which consider local features by combining graph convolution model and capture local fine-gained features by introducing spatial pyramid pooling.In order to improve the classification accuracy of the algorithm,the point cloud information was expanded by fusing Airborne LiDAR point cloud and multi spectral aerial images.At the same time,different scales were used for data expansion in the point cloud grid processing.Airborne LiDAR point cloud and multispectral aerial images provided by ISPRS were used as the experimental data in this research.The classification accuracy for the fused data was 83.9%,which was 12.1%higher than that of the original point cloud without the spectral information.At the same time,comparative experiments are designed to verify the effectiveness of the improved algorithm.This research provides an idea for point cloud classification in urban scene.