Automatic DEM generation from aerial lidar data using multiscale support vector machines

Abstract

Automatic generation of DEM from LIDAR point clouds is attractive to photogrammetry community. This paper explores the possibility of using Multi-Scale SVM technique to classify untextured Lidar data into ground points and non-ground points so that DEM can be generated efficiently. First, irregular LIDAR point clouds are rasterized and a set of features including local height variation, min/max slope, plane flatness/direction and laser return intensity are generalized as well. Second, we establish Multi-Scale SVM classification levels by implementing SVM classier at different scale-space of Lidar data and one defined conditional probabilistic model is computed to make final classification. Finally, adaptive medium filter is implemented to smooth the isolated ground points mixed with little non-ground points and because the removal of non-ground points left quite a lot "blank holes", we further triangulate smoothed non-ground points to generate DEM automatically. The experimental results prove to be quite significant for real applications.