International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 Inte
——
coordinates and the returned laser intensity for the centre of looking into important factors such as the resolutions and the
each pixel are generated. quality of the images, employing sub-pixel processing
> À PON en . techniques, camera calibration and possibly number of images
This 3D point is used as ground conti! pon for the automatic s
exterior orientation solution. In the second step, the matches
between the RCI and SCL, and their corresponding object
coordinates are used for the exterior orientation computation
with simultaneous bundle adjustment approach. This
computation, which is control point-point-free method, has
important applications in terrestrial photogrammetric SCI
engineering (Styliadis et all, 2003). Also, solving the camera
positions and orientations, the RCI can be reprojected into the RCI
point cloud surface to produce the photorealistic model.
Exterior Orientation Parameters
Xo Yo Zo C) p K
0 0 0 0 0 0
0.249 | 0.345 | 1.610 | -4.108 | 2.943 | -0.607
RMS Residual
5. RESULTS AND ANALYSIS Object Space Coordinates image coordinates
Figure 4 shows the results of the several hundreds "interest MX |MY MZ |MXYZ Mx My
points” detected (denoted with asterisks) automatically using
the Harris feature detector. As can be observe on both images, 0.001 | 0.001 0 0.001 0.017 | 0.010
most of the points of interest found in two images have
correspondences. The ZNCC has been implemented to match
the corners in SCI, with the ones in RCI and the results of them
are superimposed on the images (figure 5). The matches are
shown by the line linking matched points to their position in
the other images. The feature point selection found
approximately 800 points of interest and with the ZNCC
Table 1: Exterior Orientation Parameters of real scene |
(Positional unit: meter; Angular unit: degree) and
Measurement Accuracy
measure, using a matching patch size of (17 x 17) pixels, using Exterior Orientation Parameters
| Integer pixel locations, and correlation ihresheld of 0,5, there x. Yo 7 © Q x
| were 300 correspondences. As can be observed in figure 5, a
| | relatively large number of mismatches occurred. These SCI 0 0 0 0 0 nmi
D correspondences were refined with RANSAC algorithm and
1 out of the 300 correspondences, about 160 points were RCI 0.021 10.238 |1.151 -L633 (-0212]-1153
discovered as inliers. As can be seen in figure 6, there are large
number good corresponding sets of points for the orientation RMS Residuals
procedure. It should be noted that the size of the matching
window has a significant impact on the quality of the matches. Object Space Coordinates (m) | image coordinates(mm)
Also, the quality of the digital images, particularly the SCI MX MY MZ |MXYZ Mx My
influences, the accuracy and the success of the matching
process. However, the initial results demonstrate the ability of 0.001 0.001 0 | 0.001 0.014 | 0.018
the ZNCC algorithm to match automatically measured points
of interest. ; : :
Table 2: Exterior Orientation Parameters of real scene 2
Tables 1 and 2 present the values of the exterior orientation and (Positional unit: meter; Angular unit: degree) and
the accuracy of the measurements for two real scenes. The Measurement Accuracy
initial results of the first scene which includes, feature
detection and correspondence matching, are presented in 6. CONCLUSIONS
n Forkuo and King (2004). To verify the validity of the matching
IN | algorithm, the result of the second real scene is also presented The fusion of the 2D images and 3D point cloud has been
| in table 2. Both scenes were acquired with the same laser assessed, and a synthetic image has been generated by
| scanner and digital camera. It could be seen that the camera integrating information from the two sensors. Features have
position for the SCI for both scenes has zero coordinates detected, extracted and matched to develop geometric
(i.e. x, »v,-z,-0), with angular rotation parameter also equal relationship between the digital camera and laser scanner. The
M : à ; "niti PC : d ; st EU. ined
| to zero (à = = x =0). These exterior orientation parameters initial results show that we have successfully obtaine
I s corresponding points in both images. Bundle adjustment 1s
| of laser scanner do confirm the assumption already discussed in . utes and
Forkuo and King (2004). The same table also contains the used to reconstruct the 3D object space coordinates and fo
dm . ald c b 2. ~ A
: a 4 recover camera positions. The accuracy of the object
accuracy of the bundle adjustment in terms of the root mean : er a i“
gan : . : coordinate determination is with 0.001m and for the image
square error of the object point coordinates and of the image S m ARA RE
measuremen!s. Theceufacy in X (MX 0.001m). in Y My coordinate measurement; the measurement error is within two
sasurements. The accurac =0. ; AY= 3 ir Sex M inl É
0.001m) and in Z (MZ-0) for both scenes and the overall pixels. , However, future esearch Will concentre n
accuracy in the object space coordinate determination for both investigating the effect of resampling the RCI to a smaller size
/ C I. c € ~ . .
LAT : and the use of combined edge and corners approach instead ©
scenes (MXYZ) is within 0.001m. Also, the accuracy for the E i" pp
E . ; : x. only corners. Also, the impact of camera Catviauon,
image measurement for both x (Mx = 0.017) and y (My = artichiant"1 diserti ic tchis results will be
0.010) for real scene one and for x(Mx -0.014) and Partcu ame, Ia on me mauus fes
. c © € 4 ? AY . . % d ^ . .
ax RATA ; investigated. The RANSAC algorithm has been implemented to
y(My=0.018) for real scene two is within two pixels. However, filter "Asc "corres sondchees However “some Turtles
y ; CR ilter : S eS. Ver «s
the accuracy of the measurement can vary significantly by : Dons :
; developments of the algorithm are still required.
924
