高速公路典型要素语义分割深度学习模型适用性分析

Applicability analysis of deep learning model for semantic segmentation of typical elements of expressway

通过对比评估了目前主流的深度学习点云语义分割的网络模型PointNet、PointNet++及RandLA-Net,在高速公路场景下的语义分割性能和效果,最后选择性能最优的RandLA-Net网络模型进行超过5 km高速公路基础要素的语义分割实验,结果表明,RandLA-Net网络模型可以较好地实现高速公路场景的激光点云语义分割,总体精度达90.53%,满足现阶段高速公路场景数字化应用的信息识别精度要求。

By comparing and evaluating the semantic segmentation performance and effect of PointNet,PointNet++and RandLA-Net,the current mainstream deep learning point cloud semantic segmentation network models,in the expressway scene,the RandLA-Net with the best performance was selected for the semantic segmentation experiment of basic elements of more than 5 km expressway.The results showed that the RandLA-Net could better realize the laser point cloud semantic segmentation of Expressway scene,The overall accuracy was 90.53%,which met the information recognition accuracy requirements of expressway scene digitization application at this stage.