Outdoor Terrain Traversability Analysis for Robot Navigation using a Time-Of-Flight Camera

Autonomous robotic systems operating in unstructured outdoor environments need to estimate the traversability of the terrain in order to navigate safely. Traversability estimation is a challenging problem, as the traversability is a complex function of both the terrain characteristics, such as slopes, vegetation, rocks, etc and the robot mobility characteristics, i.e. locomotion method, wheels, etc. It is thus required to analyze in real-time the 3D characteristics of the terrain and pair this data to the robot capabilities. Stereo cameras or 3D laser range finders are generally used as input devices for traversability analysis and two main approaches can be distinguished. There are the methods as the ones advocated by Labayrade (1) and Mufti (2) who assume a (piecewise) planar ground plane. They estimate the ground plane, set a threshold and consider objects with distances to the ground plane further than this threshold as obstacles. Other methods, as the ones proposed by Birk (3) and Helmick (4) search for specific types of objects (rocks, canyons) and classify the image based on this data. To our knowledge, time-of-flight cameras have until now not been used for these kind of applications, simply because there were no sensors capable of coping with outdoor conditions, especially due to the interference of solar irradiation. This situation is changing now, with the advent of outdoor-capable sensors. Therefore, we present in this paper an approach for outdoor terrain traversability which mixes 2D and 3D information for terrain classification. The methodology towards time-of-flight-based terrain traversability analysis extends our previous work on stereobased terrain classification approaches (5). Following this strategy, the RGB data stream is segmented to group pixels belonging to the same physical objects. From the Depth data stream, the v−disparity (1) is calculated to estimate the ground plane, which leads to a first estimation of the terrain traversability. From this estimation, a number of pixels are selected which have a high probability of belonging to the ground plane (low distance to the estimated ground plane). The mean a and b color values in the Lab color space of these pixels are recorded as c. The presented methodology then classifies all image pixels as traversable or not by estimating for each pixel a traversability score which is based upon the analysis of the segmented color image and the v-disparity depth image. For each pixel i in the image, the color difference ‖ci − c‖ and the obstacle density in the region where the pixel belongs to are calculated. The obstacle density δi is here defined as: δi = 〈o∈Ai〉 〈Ai〉 , where o denotes the pixels marked as obstacles (high distance to the estimated ground plane) and Ai denotes the segment where pixel i belongs to. This allows us to define a traversability score as τi = δi‖ci− c‖, which is used for classification. This is done by setting up a dynamic threshold, as a function of the distance measured.