Detecting keypoint sets on 3D point clouds via Histogram of Normal Orientations

We propose a method to detect keypoint sets on 3D point clouds. Contrary to the existing norm that each and every keypoint should be distinctive, most of the keypoints detected by the proposed method lie in groups (from now, we refer to 'groups' as 'sets') and these groups (sets) of keypoints are distinctive. It is reliable to have sets of keypoints at high curvature and more informative areas rather than having a single keypoint that might sometimes arise due to noise. The proposed algorithm has two well-defined steps for keypoint sets detection. Firstly, Histogram of Normal Orientations (HoNO), which is calculated at every point in the point cloud is employed to avoid planar regions and successfully detect salient regions. Secondly, the keypoint sets are detected from the salient regions by evaluating the properties of both the HoNO and the neighborhood covariance matrix. Through extensive experiments on publicly available benchmark datasets, it is shown that the detected keypoint sets offer better repeatability than those by the existing ones.