Page 249 - Introduction to Autonomous Mobile Robots
P. 249

234


                              Encoder          Position Prediction  position  Estimation  Chapter 5
                                             Observation Prediction  estimate  (fusion)

                                                     predicted observations  YES
                                                                                   matched
                                                                                  and actual
                                  Map                                             predictions
                               (data base)                                       observations


                                                                             Matching




                                                                                  raw sensor data or
                                                                                  extracted features
                                                                      Perception  Actual Observations

                                                                          (on-board sensors)


                           Figure 5.28
                           Schematic for Kalman filter mobile robot localization (see [23]).



                           each relate to objects in the environment. Given a set of possible features, the Kalman filter
                           is used to fuse the distance estimate from each feature to a matching object in the map.
                           Instead of carrying out this matching process for many possible robot locations individually
                           as in the Markov approach, the Kalman filter accomplishes the same probabilistic update
                           by treating the whole, unimodal, and Gaussian belief state at once. Figure 5.28 depicts the
                           particular schematic for Kalman filter localization.
                             The first step is action update or position prediction, the straightforward application of
                           a Gaussian error motion model to the robot’s measured encoder travel. The robot then col-
                           lects actual sensor data and extracts appropriate features (e.g., lines, doors, or even the
                           value of a specific sensor) in the observation step. At the same time, based on its predicted
                           position in the map, the robot generates a measurement prediction which identifies the fea-
                           tures that the robot expects to find and the positions of those features. In matching the robot
                           identifies the best pairings between the features actually extracted during observation and
                           the expected features due to measurement prediction. Finally, the Kalman filter can fuse the
                           information provided by all of these matches to update the robot belief state in estimation.
                             In the following sections these five steps are described in greater detail. The presentation
                           is based on the work of Leonard and Durrant-Whyte [23, pp. 61–65].
   244   245   246   247   248   249   250   251   252   253   254