1 首都师范大学资源环境与旅游学院三维信息获取与应用教育部重点实验室, 北京 100048
2 首都师范大学资源环境与旅游学院空间信息技术教育部工程研究中心, 北京 100048
针对道路场景下三维激光点云语义分割精度低的问题,提出了一种基于卷积神经网络并结合几何点云多特征的端到端的语义分割方法。首先,通过球面投影构造出点云距离、相邻夹角及表面曲率等特征图像,以便于应用卷积神经网络;接着,利用卷积神经网络对多特征图像进行语义分割,得到像素级的分割结果。所提方法将传统点云特征融入到卷积神经网络中,提升了语义分割效果。使用KITTI点云数据集进行测试,结果表明:所提三维点云多特征卷积神经网络语义分割方法的效果优于SqueezeSeg V2等没有结合点云特征的语义分割方法;与SqueezeSeg V2网络相比,所提方法对车辆、自行车和行人分割的精确率分别提高了0.3、21.4、14.5个百分点。
遥感 神经网络 激光点云 语义分割 多特征 点云投影
Author Affiliations
Abstract
1 Department of Physics, School of Applied Science, Taiyuan University of Science and Technology, Taiyuan 030024, China
2 State Key Laboratory of Precision Spectroscopy, East China Normal University, Shanghai 200062, China
We present a Herriott-type multipass laser absorption spectrometer enhanced by optical heterodyne detection. The proposal is demonstrated by measuring the spectra of water vapor molecule in the region from 12247.6873 to 12249.6954 cm-1. Compared with direct absorption spectroscopy, the signal-to-noise ratio is improved nearly one magnitude of factor by combining with the optical heterodyne spectroscopy and extra weak absorption lines are observed. The minimum detectable absorption is estimated at 4.36×10-8 cm-1and the measured line shape dominated by Doppler broadening can be precisely recovered by direct transformation of experimental optical heterodyne spectral profile.
300.1030 Absorption 300.6310 Spectroscopy, heterodyne Chinese Optics Letters
2015, 13(1): 013001