A new procedure supporting filtering and classification of LiDAR data based on both elevation and intensity analysis is introduced and validated. After a preliminary analysis to avoid the trivial classification of homogeneous datasets, a non-parametric estimation of the probability density function is computed for both elevation and intensity data values. Some statistical tests are used for selecting the category of data (elevation or intensity) that better satisfies a bi-or a multi-modal distribution. The iterative analysis of skewness and kurtosis is then applied to this category to obtain a first classification. At each step, the point with the highest value of elevation (or intensity) is removed. The classification is then refined by studying both statistical moments of the complementary data category, in order to look for potential sub-clusters. Remaining clusters are identified by applying the same iterative procedure to the still unclassified LiDAR points. For more complex point distribution shapes or for the classification of large scenes, a progressive analysis is proposed, which is based on the partitioning of the entire dataset into more sub-sets. Each of them is then independently classified by using the core procedure. Some numerical experiments on real LiDAR data confirmed the potentiality of the filtering/classification method.

ABSTRACT:The paper deals with a new sequential procedure to perform unsupervised LIDAR points classification by iteratively studying skewness and kurtosis for elevation and intensity point distribution values. After a preliminary local shape analysis of elevation and intensity point distributions, carried out from the original discrete frequencies by a non parametric estimation of the density functions, the procedure starts by choosing the category of data (elevation or intensity) to analyse at first: the choice falls on the category better showing by a testing procedure a bi or a multi clustering distribution. The first point cluster is identified by studying the distribution skewness and kurtosis variations, after removing at each step the largest data values. The selected cluster is furthermore analysed by studying higher order moments behaviour of the complementary data category. This makes possible to find out potential sub clusters of the original selected one, permitting, in this way, a more effective point classification. Successive clusters are identified by applying the same iterative procedure to the still unclassified LIDAR points. For complex point distribution shapes or for the classification of large areas, a progressive analysis method, based on the partition of the entire data set into regular subsets, is proposed. Some real numerical experiments confirm the capability of the method proposed. The classification total errors in the experiments range from a minimum value of 1,2% to a maximum value of 8,9%.