More and more tasks involving cameras require continuous imaging of
360 degree space around the camera head. Such true panoramic image is
useful for security systems in open landscape, in huge halls,
supermarkets, parking lots etc. Panoramic cameras are also very useful
in ski centers to show web visitors the snow conditions, on beaches
etc.
This problem is usually solved by rotation camera, which creates
panoramic view during its rotation. But this method works only for
static scenes and provides very long image refresh periods. Sometimes
is the panoramic view created by single column camera equipped with
rotating mirror. Such cameras are quite complex, expensive and offer
only very low sensitivity, because only single column of panoramic
image is exposed at once. To perform exposure of the whole panorama
then takes quite long time or each exposure of every column must be
very short. Panoramic view can be also obtained by utilizing of convex
mirror or fish-eye lens and subsequent image transformations. Still,
image of single hemisphere may not be enough.
The interesting possibility is the usage of multiple cameras with
the subsequent composition of panoramic image from individual frames.
Such task may seem simple in principle, but there are number of tough
problems, which can be effectively solved only by using of complex
image processing.
The most important problems are:
individual cameras in camera head must be mechanically
precisely aligned
different cameras often capture scene with very different
light conditions
white balance can also be quite different on images captured
by different cameras
images are distorted by wide-angle lenses and the transform
function can vary among individual lenses
the amount of both useful data and also portions of images
omitted from final image is quite high
We decided to design camera head with four cameras. All cameras are
of digital DataCam line with high dynamic range and very low noise.
What is more, the transparency of front optical windows of the camera
head varies according to the light intensity, which enables capturing
of both bright as well as dark scenes. The CCD detectors are oriented
vertically. Vertical orientation ensures maximal utilization of CCD
area and minimizes the area not used to create final panoramic image.
Another plus of this architecture is the ability of the camera head to
capture scene almost overhead and well below the head itself. Images
are digitized in parallel and transferred into host PC through four
USB 2.0 channels.
The most interesting part of the whole system is the GPGPU
processing software. High-quality results can be obtained only it the
camera head is precisely calibrated. It is almost impossible to
achieve the required precision purely by mechanical alignment. This is
why the software itself performs image transformations to eliminate
residual mechanical differences. Geometrical transformations are only
the last step in creating panoramic image, very important is also the
quality of images read from individual cameras, debayerization of raw
data and image filtering.
Image processing and camera control virtual instruments
significantly enhance capabilities of the Control Web
version 6 industrial application development system.
rc@mii.cz
|