Jesus Balado-Frias

Geometric feature search is a very active field in semantic segmentation and classification problems with point clouds. Most artificial intelligence algorithms assume very simple searches: number of nearest neighbors, search spheres, cubes, or cylinders. In this work we propose the use of structuring elements to obtain complex geometry to improve point cloud semantic segmentation. We define the use of a geometric form (seven points aligned with orthogonal vectors) that takes advantage of the direction of the points in the XYZ space. Twelve local geometric features (linearity, planarity, scattering, curvature, omnivariance, anisotropy, eigentropy, normals, density and average height) and two radiometric features (average and maximum reflectivity) are extracted. The proposed method was applied to street point clouds to segment by Random Forest algorithm the classes road, curbs, sidewalks, building, vegetation, cars, poles and others. The average accuracy obtained with the structuring element was 81.2%, while based on KNN of 25 neighbors, the average accuracy was 82.2% and by means of a search radius, 88.2%. Consequently, it is deduced that although the segmentation is improved in some classes, the extraction of a high number of features (fourteen for each point of the structuring element) significantly hinders the training.