The thesis deals with the registration of point clouds generated using non-
dedicated hardware and open source software solutions. Finding items in space
can be done by means of a known computer aided design (CAD) model or a
captured referential point cloud. We wanted to know if it is possible to _nd the
pose of an object, described by a point cloud in space, without using dedicated
scanning hardware.
We have constructed a system using a pair of stereo cameras and attached the
object to two linear actuators and a PT head. Actuator displacements were then
used as referential translations and rotations in space. With this system we tested
the possibility of _ne alignment using the original ICP algorithm. The calibration
of the cameras strongly a_ects the precision of point cloud generation so it has
to be carried out carefully. For useful results it is necessary to consider optimal
values for illumination, distance between the cameras, distance of the object from
the cameras, and the suitability of the sought objects.
We have established that such a system is useful for locating objects in space with
the accuracy on the order of the pixel size of the cameras; the system is better
suited for _nding the pose of objects with an appropriate level of detail. Reducing
the pixel size increases depth resolution, but can also signi_cantly increase the
computational time required for the registration of the point clouds.
|
