Classification of Objects in a Point Cloud using Neural Networks

Abstract

3-dimensional scans captured in shape of point clouds are widely used in many dierent areas. Every such area use dierent kinds of sensors to ac-quire point clouds and do the analysis of the data but each of those needs some preanalysis to be done. One of the most important is segmentation and classification of points into types of objects. Such information considerably widens possibilities of usage for further purposes. There are many classifiers and many features based on which labeling can be done. In this paper few most commonly used approaches were chosen to check the influence of neighboring points acquisition on classification process. Results proof significant relation between those two steps of point cloud analysis. Visualization of analyzed point cloud also shown that precision of predictions not always comes with better visibility of certain types of objects. Additionally, color-less analysis of geometrical features seems to be promising way for further research.