In the context of scene building, the algorithms developed
at the IRIS Lab consider multi-modal data taken from such sensor types such
as color, thermal, and gamma radiation sensors, whose global/respective
locations can be thoroughly or partially known. To build accurate
photo-realistic models, sensor accuracy must be taken into account. This
research tends to integrate all the above research areas in a single scene
building process that drives a robot placed into an environment under
characterization.
When designing such a system, it may be necessary to
generate simulated data. Simulation allows low cost and fast design of such
complex systems involving heavy equipment and particular sensors. Algorithms
are studied and optimized to meet the needs and requirements necessary for
future implementation and use. For instance, system performance can be
tested with regard to simulated noisy data. By simulation, different methods
can also be compared before their implementation on the robot plate-form.
Data Acquisition and
Sensor Characterization
Data acquisition refers to the
process of obtaining relevant scene information. We emphasize the use of
laser range scanning since this particular sensor type provides spatial
data, thus facilitating the process of 3D model building. Other sensor
modalities, e.g. ordinary video, thermal imaging, and radiation imaging, are
to be incorporated according to specific application demands. Sensor
characterization associates with each sensor measurement the accuracy to be
used in the data fusion process to build precise models.
Sensor Placement
Sensor placement for scene
modeling is a growing area of computer vision and robotics. The objective of
a sensor placement system is to make task-directed decisions for optimal
pose selection. If no a priori information is available for the environment,
a robot must first be capable of acquiring data from which a model of the
environment can be constructed. Then, a sensor placement system must choose
a new sensor pose to capture data. For instance, a sensor pose can be chosen
so that the amount of new information captured from the new pose is
maximized. Moreover, this system ensures that a new pose is reachable by the
robot, and then computes the path to reach it by avoiding obstacles. The
robot limitations are also taken into account.
Data Registration
After acquiring data taken from
different points of view, the data is referenced to a unique world
coordinate to be fused and integrated in a unified representation. Data
registration is the process of computing the transformations required to
place the data into the same reference frame. Sensor locations, if known,
can be used to initialize the data registration process. Then, the exact
transformations, between same sensor data and different sensor data, are
computed. This process is of the utmost importance, because multi-sensory,
multi-view data can be referenced into the same world coordinate system to
build a photo-realistic model of the scene under characterization.
Data Fusion
The objective of the data
fusion process is to build a unified, accurate representation of a scene
from registered, multi-modal data sets. At this point, different approaches
can be used to integrate the data sets. A voxel-based or mesh-based approach
can be used as a framework for information gathering. In the process of
fusing data, areas common to different views have to be localized and
processed differently from the other data. Sensor accuracy plays an
important role in this process by associating a level of certainty to the
overlapping areas. For instance, only the most accurate data will be kept
when processing overlapping data. A more intelligent fusion strategy can
also be used to increase the accuracy of the data. Sensor data are also
associated with a visibility status to determine what data have to be
conserved for integration.
