摘要
由于激光雷达点云数据有无序性、稀疏性和信息量有限的问题,提出了一种能够将点云数据与对应图像进行三维图像重建的算法,该方法首先将点云数据体素化,利用点特征直方图有效地选择深度点进行标记并消除体素中的异常点;针对传统插值方法估计精度低的缺陷,利用高斯过程回归方法强大的非线性拟合能力和小样本学习能力,提高了内插点估计精度,获得稠密点云;最后利用马尔科夫随机场对图像灰度数据和三维插值点进行融合来构建三维深度图。定性定量仿真实验结果表明,提出的算法大大提升了三维重建的鲁棒性与重构精度,可用于复杂路况中的无人驾驶应用。
Since the point-cloud data has randomness,sparsity and the limited information in laser radar,a novel and robust 3 Dimage reconstruction algorithm based on depth point-cloud and its corresponding image in this paper.The proposed algorithm firstly makes the point-cloud data voxelization,which adopts the point feature histogram(PFH)to effectively choose the depth-point and remove abnormal point in voxel;aiming at the shortcomings of the traditional interpolation method,the Gaussian process regression method(GPF)is used to improve the accuracy of the interpolation point and obtain the dense point-cloud;finally,the Markov random field is adopted to merge the gray data and points-cloud so as to build 3 Ddepth image.The results of qualitative and quantitative simulation show that proposed algorithm is superior to other existing algorithms in terms of RMSE and run time.
出处
《应用激光》
CSCD
北大核心
2017年第6期881-887,共7页
Applied Laser
基金
上海市教育发展基金会"晨光计划"资助项目(项目编号:12CGB22)
高等学校访问学者专业发展资助项目(项目编号:FX2012122)
关键词
三维重建
激光雷达
高斯过程回归
点特征直方图
马尔科夫随机场
3D reconstruction
laser radar
Gaussian process regression
point feature histogram
Markov random field