基于半球形邻域的激光雷达点云局部特征提取

Local Feature Extraction of LiDAR Point Cloud Based on Hemisphere Neighborhood

为提升激光雷达点云目标识别的效率和减少计算开销,本文通过采用改进的邻域确定方法和LRF估计方法,提出了一种新的特征描述子:半球单值形状上下文特征描述子(Hemispheric Unique Shape Context,HUSC)。首先计算关键点处的互相关矩阵,并根据邻域点附近的点密度进行加权,以此估计关键点的法向量和切平面,并建立局部参考坐标系;然后以该切平面为底面构建半球形邻域,并将其按照方位角、极角和径向划分为多个栅格;最后统计各栅格中的点数,得到关键点的局部特征描述子。HUSC特征描述子在保证描述子准确率的同时,通过减少冗余栅格数量提高目标识别的效率。在Bologna、3DMatch数据集上进行的实验表明,基于半球形邻域的HUSC特征描述子与基于球形邻域的USC描述子对噪声鲁棒、对分辨率变化稳健性相当,但HUSC特征描述子占用内存更小,运算速度更快。

In order to improve the efficiency of LIDAR point cloud target recognition and reduce the calculation cost,this paper proposed a new feature descriptor,Hemispheric Unique Shape Context(HUSC),by adopting the improved neighborhood determination method and LRF estimation method.Firstly,the covariance matrix at the key point was calculated,and weighted according to the density of the points near the neighborhood points,and then we could estimate the normal vector and tangent plane of the key points by EVD of the covariance matrix.Thus,the local reference coordinate system was established.Then,the hemispherical neighborhood was constructed based on the tangent plane and divided into several grids according to azimuth Angle,polar Angle and radial direction.Finally,the local feature descriptors of key points were obtained by counting the points in each grid.HUSC feature descriptor improved the efficiency of target recognition by reducing the number of redundant grids while ensuring the accuracy of descriptor.Experiments on the Bologna and 3DMatch dataset show that the HUSC feature descriptor based on the hemispherical neighborhood is as robust to noise and varying resolutions as the USC feature descriptor based on the spherical neighborhood,but the HUSC feature descriptor occupies less memory and has faster computing speed.