Classification of LASER points cloud using the concept of principal components analysis and the non ambiguity factor

The aim of this paper is to present a method that automatic performs the classification of 3D points sampled by an airborne LASER scanning system. In this method the eigenvalues of the variance and covariance matrix (MVC) computed for a neighborhood around the points of interest are used to discriminate the predefined classes or structures. This neighborhood is dynamically obtained by using the concept of entropy and the classification is performed by comparing the estimated eigenvalues, relative to each point and its neighborhood, with the eigenvalues of the predefined structures or classes. To compute the similarity, the Euclidean distance in the eigenvalues space was considered and in order to eliminate the ambiguous points the non ambiguity factor (FNA) was incorporated. The evaluation of the proposed and implemented method was realized using a set of LASER data from the city of Presidente Prudente/SP, with an average density of 8 points/m². The results showed that even with the complexity of real environments, some structures were well defined and can be identified. With the incorporation of FNA it was possible to identify and eliminate points with a high probability of belonging to two (or more) classes (ambiguous points), generally sampled over vegetation, regions with low density points, close to regions of edges and on transmission lines near the buildings. It was also observed that the incorporation of FNA allowed to decrease the number of points incorrectly classified, mainly for big objects. Additionally, it was also possible to verify that the use of FNA is interesting for vegetation areas, since great part of ambiguous points are identified.

Keywords:
Classification of three-dimensional points; Principal component analysis; Non ambiguity factor; LASER point clouds