In this work, we show how local 3D shape descriptors can be used to efficiently detect segments of fallen trees in LiDAR point clouds. Our approach takes advantage of two types of highly expressive classes of shape features known from computer vision: Point Feature Histograms (PFH) at the single point level as well as 3D Shape Contexts (SC) at the primitive (segment) level. The low-level point information forms a basis for determining potential segment candidates, which are then classified by a binary support vector machine (SVM). Based on manual labeling of points, we tested our method on two sample plots from the Bavarian Forest National Park acquired using full waveform LiDAR data with a point density of 30 pts/m2 . The results, obtained using 5-fold validation, show that the studied shape descriptors have a high discriminative capability for the task of detecting fallen tree segments, yielding an overall classification accuracy of above 90%.     «

In this work, we show how local 3D shape descriptors can be used to efficiently detect segments of fallen trees in LiDAR point clouds. Our approach takes advantage of two types of highly expressive classes of shape features known from computer vision: Point Feature Histograms (PFH) at the single point level as well as 3D Shape Contexts (SC) at the primitive (segment) level. The low-level point information forms a basis for determining potential segment candidates, which are then classified by a...     »