Information Technology Reference
In-Depth Information
Test performed in a robot simulator shows a better performance in robot
navigation when using the improved method than the basic one. Robot exhibits
an improved ability in turning corners in crowded environments and tracking
goals, in a safe and ecient way. We reproduced the scenario of our real robot
experiments from [13] in the simulator. When using the basic method we can
observe many situations as those described previously in this section, a majority
of these dangerous situations were avoided using the improved method.
7 Conclusions and Future Work
In this work we have presented a novel method for line extraction from range
measures in polar coordinates. It has been designed for a noise model in which
the measure errors have a standard deviation proportional to a known function
of the expected measure, the rate constant may be unknown. This last fact allows
our method to deal with different noise levels in the same scan. The method may
be classified as a Line Regression with Clustering method and it is composed by
three elements: a line regression method using EKF, a clustering procedure to
look for adequate places in the scan that are useful to start EKF, and a merging
procedure for similar adjacent segments.
The key for designing a filter that can deal with our noise model has been filter
formulation. We have used and extended a filter formulation that is equivalent
to regression models theory, so every result, property, test, etc., from regres-
sion theory are at our disposal. This fact provides tools for: line parameters
estimation, noise variance estimation, estimators covariances matrix estimation,
outliers detection, similar segments merging, etc., grounded in a well known and
developed theory. As a future research, tools for scan matching in SLAM can be
developed with the help of regression theory. Filter has linear time complexity.
We have used scale-space techniques for clustering by developing a filter for
noise in order to get less fragmented features in the environment. Filter is based
on a coarse characterization of the set of zero-crossing level curves generated by
noise in the scale-space representation, followed by a statistical selection inside
the set.
The clustering procedure usually provides adequate places, seeds, to start EKF
estimation. Using outliers detection we can establish a segment end detection
criterion. Possible overlap between segments are solved and similar segments are
merged. Our C++ implementation provides execution times fast enough for use
in robots.
We need to test extensively the whole process, at this moment only EKF
has been tested extensively, and to make comparisons with other segmentation
methods. Test in real world has been started using laser SICK-LMS 200 and
Kinect device. Research on extending the method to curves and to 3D measures
are at initial stages of development.
The first use for our segmentation method has been to improve the center of
area method for navigation. By means of adequate split points selection, using
segments ends, it is possible to get a more reliable and safer navigation.
Search WWH ::
Custom Search