Traversability analysis technics in outdoor environments: a comparative study C.Castej´on D. Blanco B.L. Boada L. Moreno Mechanical Dept. System Engineering and Automation Dept. Mechanical Dept. System Engineering and Automation Dept. Carlos III University Carlos III University Carlos III University Carlos III University Spain Spain Spain Spain castejon@ing.uc3m.es dblanco@ing.uc3m.es bboada@ing.uc3m.es moreno@ing.uc3m.es Abstract A comparative study among different methodolo- gies, to extract traversability information in three- dimensional snapshot models, is presented in this pa- per. In particular, the main goals are the sensor study, to obtain three-dimensional information in outdoor en- vironments and the terrain classification, in traversable and non traversable regions, based on the environ- ment’s and robot’s characteristics. Conclusions of these studies, based on the experimental results carried out in a real outdoor environment, with a real robot, are presented. 1 Introduction During the last years the interest in mobile robotics, operating in outdoor environments, is increasing due to the new applications such as: human service or co- operation. To achieve the commanded task (for mo- bile robotics, it consists of evolving in the environment or manipulating objects), the robot must be equipped with a suitable sensor system. Sensor allows to get information about the environment and a modelling system, that transforms the sensor information in use- ful information for the decisional system. Environment perception and modelling tasks are required for navigation, guidance or control tasks in mobile robotics. For decision-taking task (movement, manipulation or control) the robot needs to know where it is and how is the environment where it must evolve to avoid colliding. The model choice is not easy, it must be: adapted to the type of environment, to the robot’s physical re- strictions and task. Besides, the information to mem- orize in the robot’s internal model must be limited [5]. In this paper a model, representative for an out- door environment and for a large outdoor robot called GOLIAT , is obtained. Our choice is the Traversabil- ity Numerical Model (TNM). TNM is defined as the three-dimensional (3D) instantaneous model, enhanced with the traversability attribute [3]. The model ob- tained will not only depend on the terrain character- istics but on the robot’s ones. Previous to the model construction, it is necessary to obtain information about the environment. This in- formation must be obtained in real time and it must be dense enough to represent the environment. The most suitable technics used in mobile robotic, when a 3D model is presented, are based in scanner laser or stere- ovision technics. In our approach the 3D scanner laser is chosen. The system gives an environment’s image of sorted points. Studies with stereovision technics have been carried out and concluded as not interest- ing because of the lack of textures in the environment where the robot is moving. 2 Previous Work In most of the projects, related to navigation in outdoor environments, researchers try to divide the perceived terrain in regions with specific characteris- tics (segmentation process). Langer in [11] introduces the traversability concept to determine the region’s capacity to be crossable or not by a robot. Langer builds a terrain segmentation, based on depth images, and a list with the non-crossable regions is generated. This information is sorted out in a digital elevation map (DEM) before the classification, which is made up with: the cells height gradient, the normal vec- tor orientation, and the discontinuity measurements in the DEM. The list of regions is used to navigate on roads and highways, that is, simple and partially structured environments. Piat et al. in [13] use a Bayesian classifier to perform a natural terrain seg- mentation. The environment is discretized in cells, and the following attributes are calculated: typical height deviation, maximum elevation difference, mean normal vector and variances. With these attributes, the classifier labels the zones as: flat with admissible slope, uneven terrain or obstacles. This supervised al- gorithm needs a previous state of training, with a hu- 1710 Proceedings of ICAR 2003 The 11th International Conference on Advanced Robotics Coimbra, Portugal, June 30 - July 3, 2003