Local Feature Extraction of LiDAR Point Cloud Based on Hemisphere Neighborhood

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.