3D Descriptor of Objects in a Scene based on the Average Depth of Point Cloud Images

In the field of machine vision, 3D point cloud data contains rich and comprehensive spatial information, and local feature descriptors based on 3D point cloud data play an important role for 3D object recognition tasks. In this paper, a new local surface descriptor is designed based on a stable local reference frame. First, the key points of the point cloud data are extracted, and the CA (Component Analysis) method is applied to obtain three vectors orthogonal to each other for the key points and their local domain points, and the ambiguity of the vector direction is eliminated according to the projection of the domain points on the plane formed by the vectors, and a stable, unambiguous and repeatable local reference frame. Then, based on this reference frame, the domain of key points is specified, and the points in the domain are projected onto each of the three planes, and the depth information of the local surface is computed based on this frame, so that the average depth information of the local points is encoded into a compact histogram under multiple views. Finally, it is shown experimentally that our Point Average Depth (PAD)descriptors has descriptive and reproducible, and achieve good performance in tests on the University of Western Australia (UWA) dataset and the self-collection dataset.