【正文】 above its path. This human assisted map building is the first step of our research concerning landmarks map building. We want to change it to a fully autonomous map building system. As the imageprocessing involved during the learning process is identical to the one used during the navigation, we will present the feature extraction method in sections 5 and 6. Once the teaching phase is pleted, the robot holds a map of the lights that can be used later for the autonomous navigation process. 4 Dealing with a robotmade map Odometry error’ s influence on the map Asking the robot to build a map implies dealing\ with odometry errors that will occur during the learning process itself. As the robot will be guided under new lights, because of the accumulation of odometry errors, the pose of the landmarks recorded in the map will bee more and more different from the values corresponding to the real world. Several maps of the environment represented in are given in . The odometry data recorded by the robot during the learning process has also been represented for one of the maps. Usage of the map Only one map is needed by the robot to correct its pose during the navigation process. Whenever the robot detects a light learnt previously, it corrects its absolute pose1 by using the landmark’ s information recorded in the map. Since the map contents don’ t correspond to the values of the real world, the trajectory of the robot has to be specified according to the pose of the lights in the map, and not according to the trajectory we want the robot to follow in its real environment. For example, if the mobile robot’ s task is to navigate right below a straight corridor’ s lights, the robot won’ t be requested to follow a straight line along the middle of the corridor. Instead of this simple motion mand, the robot will have to trace every segment which connects the projection on the ground of the center of two successive lights. This is illustrated in where a zoom of the trajectory specified to the robot appears in dotted line. A GUI has been developed in Tcl/Tk in order to specify easily different types of trajectories with respect to the map learnt by the robot. This GUI can also be used online in order to follow the evolution of the robot in real time on the landmarks map during the learning and navigation processes. 濟(jì)南大學(xué)泉城學(xué)院畢業(yè)設(shè)計(jì)外文資料翻譯 4 Figure 2: Several maps of the environment represented built by the same robot. Rectangles and circles represent lights of different shapes. 5 Fluorescent tube detection Fluorescent tube model It is natural to think of fluorescent tube as a natural landmark for a visionbased process aimed at improving the localization of a mobile robot in an indoor environment. Indeed, problems such as dirt, shadows, light reflection on the ground, or obstruction of the landmarks usually do not appear in this case. One advantage of fluorescent tubes pared to other possible landmarks located on the ceiling is that once they are switched on, their recognition in an image can be performed with a very simple imageprocessing algorithm since they are the only bright elements that are permanently found in such a place. If a 256 grey levels image containing a fluorescent tube is binarized with an appropriate threshold 0 ≤ T ≤ 255, the only element that remains after this operation is a rectangular shape. shows a typical camera image of the ceiling of a corridor containing a fluorescent light. The axis of the camera is perpendicular to the ceiling. Shown in (b) is the binarized image of (a). If we suppose that the distance between the camera and the ceiling remains constant and that no more than one light at a time can be seen by the camera located on the top of the robot, a fluorescent tube can be modeled by a given area S0 in a thresholded image of the ceiling. Figure 4: (a) Sample image of a fluorescent light, (b) binarized image. 濟(jì)南大學(xué)泉城學(xué)院畢業(yè)設(shè)計(jì)外文資料翻譯 5 Fluorescent light detection process Using odometry, the robot is able to know when it gets close to a light recorded in its map by paring in a close loop its actual estimated position to the different locations of the lights in the map. Once it gets close to one of them, it starts taking images of the ceiling and binarizing them with the threshold T used to memorize the corresponding light during the learning process. This step is repeated until the number N of pixels brighter than T bees close to S0. When it happens to be true and when the binarized shape does not touch any border of the image, the detection algorithm is stopped and further imageprocessing is done as explained in the next section. In order to discard too bright images, the detection algorithm increases automatically the threshold. Moreover, because the intensity of the light emitted by fluorescent tubes changes with a frequency corresponding to the cycle of electric power, the threshold has to be decreased automatically if N ≤ S0, so that the robot