融合多特征深度学习的地面激光点云语义分割

针对当前点云语义分割研究对地面站激光点云特征利用不足、正确率较低的问题,该文提出了一种基于多尺度球形邻域特征的深度神经网络算法。该算法基于多尺度球形邻域计算的地面激光点云的粗糙度、高斯曲率,以及全方差、线性度等基于协方差的多种特征,结合XYZ坐标、RGB颜色、激光反射强度组成47维特征向量作为神经网络的输入,经过多组参数组合实验优化神经网络结构,最后通过softmax分类器输出每个点的类别。利用Semantic-3D测试集验证所提的深度神经网络模型,取得了较好的分类精度,总体正确率和平均交并比分别达到了86.6%和55.0%。实验结果表明,所提算法充分利用了地面站激光点云的特征,可有效提升语义分割的正确率。

Ground laser point cloud semantic segmentation based on multi-feature deep learning

Aiming at the problem of insufficient utilization ground point cloud feature and low accuracy of in current research on point cloud semantics segmentation,a depth neural network algorithm based on multi-scale spherical neighborhood feature is proposed.The algorithm uses multi-scale spherical neighborhood method to calculate the roughness and Gaussian curvature of ground laser point cloud,as well as various features based on covariance such as full variance and linearity.Combined with XYZ coordinates,RGB color and laser reflection intensity,47-dimensional features vector are formed as the input of the neural network.Then the neural network structure was optimized by multiple sets of parameter combination experiments,and finally outputs the category of each point by the softmax classifier.The Semantic-3 D test set was used to verify the proposed deep neural network model,and the good classification accuracy was achieved.The overall accuracy and the mean intersection over union were 86.6%and 55.0%.The experimental results show that the proposed algorithm makes full use of the characteristics of laser point cloud on the ground station,which can effectively improve the accuracy of semantic segmentation.