Page 51 - Autonomous Mobile Robots
P. 51

34                                     Autonomous Mobile Robots

                                be classified as “dangerous” or “not dangerous.” Color cameras can be used to
                                perform terrain classification. Color segmentation relies on having a complete
                                training set. As lighting changes, due to time of day or weather conditions, the
                                appearance of grass and obstacle change as well. Although color normalization
                                methods have been successfully applied to the indoor environment, they, to
                                our knowledge, fail to produce reasonable results in an outdoor environment.
                                Similarly, color segmentation can classify flat objects, such as fallen leaves, as
                                obstacles, since their color is different from grass.
                                   If dense range measurements in a scene are available (e.g., using ladar), they
                                can be used, not only to represent the scene geometry, but also to characterize
                                surface types. For example, the range measured on bare soil or rocks tends to
                                lie on a relatively smooth surface; in contrast, in the case of bushes, the range
                                is spatially scattered. While it is possible — although by no means trivial — to
                                design algorithms for terrain classification based on the local statistics of range
                                data [39–41], the confidence level of a reliable classification is low. Table 1.4
                                lists the most frequently encountered terrain types and possible classification
                                methods.


                                1.4.2 Localization and 3D Model Building from Vision
                                Structure from motion (SFM) is the recovery of camera motion and scene
                                structures — and in certain cases camera intrinsic parameters — from image



                                 TABLE 1.4
                                 Terrain Types and Methods of Classification
                                                                                     Confidence
                                 Terrain type       Sensors      Classification methods  level
                                 Vegetable       IR/Color camera  Segmentation       Medium
                                 Rocks           IR/Color camera  Segmentation       Medium
                                 Walls/fence     Camera, stereo,  Texture analysis, obstacle  High
                                                  laser         detection
                                 Road (paved, gravel,  IR/Color camera  Segmentation  Medium
                                  dirt)
                                 Slope           Stereo, ladar  Elevation analysis, surface fit  High
                                 Ditch, hole     Stereo, ladar                       Low
                                 Sand, dirt, mud,  IR/Color camera  Segmentation     Medium
                                  gravel
                                 Water           Polarized camera,  Feature detection, sensor fusion  Medium
                                                  laser scanner
                                 Moving target   Camera, stereo  Optical flow, obstacle  High
                                                                detection, pattern matching





                                 © 2006 by Taylor & Francis Group, LLC



                                 FRANKL: “dk6033_c001” — 2006/3/31 — 16:42 — page 34 — #34
   46   47   48   49   50   51   52   53   54   55   56