Research on 3-D Laser Point Cloud Recognition Based on Depth Neural Network

Abstract

Typical convolution architectures require fairly conventional input data formats, such as image grids or three-dimensional pixels, to show shared weights and other kernel optimizations. Because point clouds and grids are not typical formats, most researchers usually convert these data into conventional three-dimensional pixel grids or picture sets before providing them to deep-net architectures. However, this data representation transformation presents unnecessary result data and introduces the natural invariance of quantified workpiece fuzzy data. For this reason, we focus on using a different simple point cloud input representation for three-dimensional geometry, and named our deep network as point network. Point cloud is a simple and unified structure, which avoids the combination of irregularity and complex grids, so it is easier to learn. This topic takes point cloud as input directly, and outputs the whole input classification label or every part label of each point input. In the basic settings, each point is represented by three coordinates (x, y, z), and additional dimensions can be added by calculating normals and other local or global characteristics.