Mapping with mobile robots

摘要

Today, mapping is largely considered the most difficult perceptual
<br>problem in robotics. Different problems like the statistical
<br>dependence, the correspondence problem, or dynamic elements makes
<br>mapping so hard. The basic technique for mapping is incremental scan
<br>matching. We present a method, which maximizes the likelihood of a
<br>scan, given a motion model and the map built so far. This approach is
<br>the constitutive techniques for further methods which addresses
<br>special topics of the previous described problems. In this thesis we
<br>describe following methods:
<br>
<br>We present a technique which combines a people tracker in the mapping
<br>process to filter measurements caused by walking people. Additional
<br>to the fact that maps contains less spurious objects, the accuracy of
<br>the results can be increased.
<br>
<br>We present a second technique which is able to filter measurements
<br>caused by dynamic objects. This time dynamic objects are not limited
<br>to walking persons, every dynamic object can be filtered out if the
<br>place of this object was seen as both: as free space as well as
<br>occupied space. An EM-technique is used to optimize the mapping
<br>results and in several practical experiments we will show that this
<br>approach can effectively filter beams of different dynamic objects
<br>like humans, cars etc.
<br>
<br>A famous problem in mapping is the so called closing loop problem.
<br>When closing the cycle, the robot has to find out where it is relative
<br>to its previously built map. This problem is complicated by the fact
<br>that at the time of cycle closing, the robot's accumulated pose error
<br>might be unboundedly large. We describe an efficient version of the
<br>Rao-Blackwellized mapping approach. Using a Rao-Blackwellized
<br>particle filter is a good techniques for estimating the posterior. In
<br>combination with the scan matching technique we can decrease the
<br>number of particles, so that it can be executed online.
<br>
<br>The Rao-Blackwellized mapping approach suffers from the inherent
<br>problem of all particle filters: the particle depletion problem. We
<br>present a new algorithm for data association in SLAM. In essence, our
<br>approach searches the combinatorial tree of possible data association
<br>decisions. The search is lazy: only when an alternative assignment
<br>shows promise will it be evaluated. We tested our approach, and
<br>consistently found that it produces accurate maps, even if for maps
<br>with many large cycles.
<br>
<br>It can be difficult to distinguish places with data from range sensors
<br>as measurements can look the same at different positions. RFID tags
<br>have to nice property to be unique, they can report their
<br>identification numbers which are easy to distinguish. Unfortunately
<br>they don't provide any distance information, so that estimating the
<br>location of the RFID tags is a difficult problem. We present an
<br>approach to generate maps of RFID tags with mobile robots. We present
<br>a sensor model that allows us to compute the likelihood of tag
<br>detections given the relative position of the tag with respect to the
<br>robot. Additionally we describe how to compute a posterior about the
<br>position of a tag after the trajectory and the map is generated with a
<br>highly accurate Rao-Blackwellized mapping algorithm for laser range
<br>scans.
<br>
<br&g