Object recognition in three-dimensional point clouds is a new research topic in the field of computer vision. Numerous nuisances, such as noise, a varying density, and occlusion greatly increase the difficulty of 3D object recognition. An improved local feature descriptor is proposed to address these problems in this paper. At each feature point, a local reference frame is established by calculating a scatter matrix based on the geometric center and the weighted point-cloud density of its neighborhood, and an improved normal vector estimation method is used to generate a new signature of histograms of orientations (SHOT) local-feature descriptor. The geometric consistency and iterative closest point method realize 3D model recognition in the point-cloud scenes. The experimental results show that the proposed SHOT feature-extraction algorithm has high robustness and descriptiveness in the object recognition of 3D local descriptors in cluttered point-cloud scenes.
Learning a Local Feature Descriptor for 3D LiDAR Scans
