Autonomous mobile robot localization in large-scale environments using only a camera.
详细信息
文摘
Localization is a fundamental problem
of autonomous mobile robots. Localization is the determination of the position
and orientation of a robot. Most localization systems are made up of several
sensors and a map of the environment.;The PRISM/CReSIS project is developing
autonomous robots in an effort to measure ice sheets characteristics in
Greenland and Antarctica. These robots currently rely on differential GPS for
localization and navigation. In order to survive for long periods of time in
these environments, however, the robots needs to be able to return to camp sites
in order to refuel and unload the data that has been acquired. In order to
perform this task effectively and safely, a more elaborate system is required. A
localization system that can recognize the different locations of the camp sites
is the beginning of this process.;The approach of this dissertation is to use a
single camera for use in multiple types of large-scale environments: indoors,
outdoors, and in polar camp sites in Greenland and Antarctica. The camera is
selected as the sensor for several reasons. First, a single image potentially
contains a lot of information that can be used in many different ways. Second,
the size of a camera allows for the system to be used on many different
platforms, including those with limited payloads such as UAVs (uncrewed aerial
vehicles). This also allows for the system to be very portable if necessary, and
can be plugged into already existing systems more easily. Lastly, the cost of
cameras allow for the system to be used in large quantities. For example, a
potential application of this system is using teams of robots for seismic
sensing. This would require many cameras for use on numerous mobile robots.;In
order to work in large-scale environments, a hybrid map approach has been used.
The hybrid map includes both topological and geometric maps.;The system
described in this dissertation uses an appearance-based approach for recognizing
the different locations.;The results of the testing showed that 95% of the
images for non-polar regions, both indoor and outdoor, were localized correctly
with respect to the topological map. The system typically required at least two
images to solve the global localization problem, and around three images to
solve the kidnapped robot problem.;Topological testing was also performed using
images from polar camp locations, but the results are inconclusive because of
the relatively few number of images. The system is able to localize 20% of the
images.;The metrics for the geometric testing are position accuracy and
orientation accuracy. Position accuracy is the percentage of images that are
correctly localized with respect to the position. Orientation accuracy is the
percentage of images that are correctly localized with respect to the
orientation.;The geometric testing is performed in non-polar locations, both
indoors and outdoors. The experiments result in 94% of the images being
localized correctly for the position, and 90% of the images for the orientation.
These images are localized to within 1 foot and 45 degrees of the actual
position and orientation.
of autonomous mobile robots. Localization is the determination of the position
and orientation of a robot. Most localization systems are made up of several
sensors and a map of the environment.;The PRISM/CReSIS project is developing
autonomous robots in an effort to measure ice sheets characteristics in
Greenland and Antarctica. These robots currently rely on differential GPS for
localization and navigation. In order to survive for long periods of time in
these environments, however, the robots needs to be able to return to camp sites
in order to refuel and unload the data that has been acquired. In order to
perform this task effectively and safely, a more elaborate system is required. A
localization system that can recognize the different locations of the camp sites
is the beginning of this process.;The approach of this dissertation is to use a
single camera for use in multiple types of large-scale environments: indoors,
outdoors, and in polar camp sites in Greenland and Antarctica. The camera is
selected as the sensor for several reasons. First, a single image potentially
contains a lot of information that can be used in many different ways. Second,
the size of a camera allows for the system to be used on many different
platforms, including those with limited payloads such as UAVs (uncrewed aerial
vehicles). This also allows for the system to be very portable if necessary, and
can be plugged into already existing systems more easily. Lastly, the cost of
cameras allow for the system to be used in large quantities. For example, a
potential application of this system is using teams of robots for seismic
sensing. This would require many cameras for use on numerous mobile robots.;In
order to work in large-scale environments, a hybrid map approach has been used.
The hybrid map includes both topological and geometric maps.;The system
described in this dissertation uses an appearance-based approach for recognizing
the different locations.;The results of the testing showed that 95% of the
images for non-polar regions, both indoor and outdoor, were localized correctly
with respect to the topological map. The system typically required at least two
images to solve the global localization problem, and around three images to
solve the kidnapped robot problem.;Topological testing was also performed using
images from polar camp locations, but the results are inconclusive because of
the relatively few number of images. The system is able to localize 20% of the
images.;The metrics for the geometric testing are position accuracy and
orientation accuracy. Position accuracy is the percentage of images that are
correctly localized with respect to the position. Orientation accuracy is the
percentage of images that are correctly localized with respect to the
orientation.;The geometric testing is performed in non-polar locations, both
indoors and outdoors. The experiments result in 94% of the images being
localized correctly for the position, and 90% of the images for the orientation.
These images are localized to within 1 foot and 45 degrees of the actual
position and orientation.