: Covariance Computation
: Calib Tool
: Pre-Processing
目次
There are several calibration procedures available. The choice of
these will depend on how the calibration data has been obtained
and how much a-priori information needs to be taken into account.
- Tsai. This routine takes observed image locations and
known 3D measuremets and fits a model to the left and right cameras
separately using Tsai's calibration method (on grid data only). The
technique does not determine image centres or aspect ratio's and
does not deal with radial distortion effects. Though the algorithm
can often be used to provide a good initial estimate it is neither
optimal or robust and model optimisation should be finished using
alternative techniques.
- IP min. This routine performs direct image plane
minimisation of the difference between the observed and predicted
locations of selected data. Left and right calibration is performed
separately with respect to the specified "Model" camera parameters
and the six parameters of camera transformation. The method used is an
iterative minimisation technique which starts by choosing test values
around an initial estimate according to the parameter "scale init"
on the Calib Params menu. Convergence is determined by the
Calib Params "c_test1" and "c_test2". The first parameter tests for
convergence of a minimisation step, minimisation is restarted until
the overall reduction in the error function is letts than the second
parameter. The technique also excludes data which is inconsistent
with the current calibration according to the
value of the "accuracy" parameter. The technique is thus performing
a robust fit which explicitly excludes outliers from the determination
of the minimum.
- IS min. This routine performs direct image plane
minimisation of the difference between the observed and predicted
locations of selected data. Unlike the routine described above however,
left and right cameras are calibrated simultaneously. This allows
previously determined covariance estimates to be used to constrain
the new solution in an optimal manner. Parameters are the same as
for IP min.
- EPI min. This routine performs least-squares off epi-polar
minimisation. The routine thus determines the left and right camera models
which are most consistent with the observed stereo geometry. The absence
of 3D information dictates that this method can only provide information
on the ratio of the two focal lengths and the relative transformation
between the camera co-ordinate systems up to a scale factor and this only
whene there is a fair amount of perspective effects in the matched image data.
A covariance
estimate from a previous calibration must be used to constrain the
minimisation process (see init covar below) when
extra parameters are selected or the data is insufficient to uniquely
determine the parameter set.
- Model min. This routine projects the features of
a wireframe model (loaded in the Matcher Tool) from the view
direction specified by the current calibration parameters. It then
adjusts the parameters to get the maximum likelihood projection.
This process is done for separate cameras and does not yet use the
stereo camera covariance. It's main purpose is view and feature validation
for automated model construction.
: Covariance Computation
: Calib Tool
: Pre-Processing
目次
root
平成24年2月7日