Page 271 - Introduction to Autonomous Mobile Robots

Chapter 5
256
cases this assumption is incorrect. In the case of wide-open spaces that are popular gather-
ing places for humans, there is rapid change in the free space and a vast majority of sensor
strikes represent detection of transient humans rather than fixed surfaces such as the perim-
eter wall. Another class of dynamic environments are spaces such as factory floors and
warehouses, where the objects being stored redefine the topology of the pathways on a day-
to-day basis as shipments are moved in and out.
In all such dynamic environments, an automatic mapping system should capture the
salient objects detected by its sensors and, furthermore, the robot should have the flexibility
to modify its map as to the positions of these salient objects changes. The subject of con-
tinuous mapping, or mapping of dynamic environments, is to some degree a direct out-
growth of successful strategies for automatic mapping of unfamiliar environments. For
example, in the case of stochastic mapping using the credibility factor c mechanism, the
t
credibility equation can continue to provide feedback regarding the probability of the exist-
ence of various mapped features after the initial map creation process is ostensibly com-
plete. Thus, a mapping system can become a map-modifying system by simply continuing
to operate. This is most effective, of course, if the mapping system is real-time and incre-
mental. If map construction requires off-line global optimization, then the desire to make
small-grained, incremental adjustments to the map is more difficult to satisfy.
Earlier we stated that a mapping system should capture only the salient objects detected
by its sensors. One common argument for handling the detection of, for instance, humans
c
in the environment is that mechanisms such as can take care of all features that did not
t
deserve to be mapped in the first place. For example, in [157] the authors develop a system
based on a set of local occupancy grids (called evidence grids) and a global occupancy grid.
Each time the robot’s most recent local evidence grid is used to update a region of the global
occupancy grid, extraneous occupied cells in the global occupancy grid are freed if the local
occupancy grid detected no objects (with high confidence) at those same positions.
The general solution to the problem of detecting salient features, however, begs a solu-
tion to the perception problem in general. When a robot’s sensor system can reliably detect
the difference between a wall and a human, using, for example, a vision system, then the
problem of mapping in dynamic environments will become significantly more straightfor-
ward.
We have discussed just two important considerations for automatic mapping. There is
still a great deal of research activity focusing on the general map-building and localization
problem [22, 23, 55, 63, 80, 134, 147, 156]. However, there are few groups working on the
general problem of probabilistic map-building (i.e., stochastic maps) and, so far, a consis-
tent and absolutely general solution is yet to be found. This field is certain to produce sig-
nificant new results in the next several years, and as the perceptual power of robots
improves we expect the payoff to be greatest here.