Terrain classification based on structure for autonomous navigation in complex environments

One of the main challenges for autonomous navigation in cluttered outdoor environments is to determine which obstacles can be driven over and which need to be avoided. Especially in off-road driving, the aim is not only to recognize the lethal obstacles on the vehicle's way at all costs, but also to predict the scene category thereby giving a better decision-making framework for vehicle navigation. This paper studies terrain classification based on structure relying on sparse 3-D data from LADAR mobility sensors. While most of recent methods for LADAR processing are purely found on the local point density and spatial distribution of the 3-D point cloud directly. We, on the other hand, introduce a new approach to analyze the point cloud by considering local properties and distance variation of pixels inside edgeless areas. First of all, the edgeless areas are extracted from segmenting the 3-D point cloud into homogeneous regions by Graph-Cut technique. Secondly, the neighbor distance variation inside edgeless areas (NDVIE) features are obtained by calculating the euclidean distance of neighbor distance variation inside each region. Through extensive experiments, we demonstrate that this feature has properties complementary to the conditional local point statistics features traditionally used for point cloud analysis, and show significant improvement in classification performance for tasks relevant to outdoor navigation.