International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
untransformed image. Created vectors are projected onto point
cloud using interpolation.
2. DESCRIPTION OF METHOD
2.1 Digital image
The camera used for capturing of photographic images should
be calibrated, i.e. internal orientation parameters and distortion
should be known with high enough precision. Only under this
condition good result could be achieved.
Transformation of images performed after loading for image
distortion correction. Collinearity equations are used for image
transformation:
ay (X - Xs) t a (Y - Ys) t a4(Z -Zs) Td
a4, (X - Xs) t a4, (Y - Ys) c a4(Z - Zs) s (1)
a, (X-Xs)+a,(Y - Ys) + a, (Z - Zs)
aXX) ta, Y Ys a (Z- Z9 17
» 7
yof
where
x,,y,,f - interior orientation parameters;
% J - coordinates of the image point;
q.. - elements of rotation matrix;
ij
NX spatial coordinates of focal center point;
X,Y,Z -spatial coordinates of image point;
d ,d.- image distortion;
d, E -n y, «(^ -nyl (^ * 2x? )p, +2xyp, (2)
d, = [ir -n, +r — rd Je, J+ 230m, +(r? +2y%)p,
where
k,,k, - radial distortion;
p,, p, - tangential distortion;
r-4x^4 y? - distance to the main point of image;
r, - distance to the zero distortion point.
After transformation image segmentation is performed.
Segmentation is performed using region growing algorithm.
Process of segmentation starts from initial pixel searching. A
criterion of choosing initial pixel is homogenous radiometric
intensity of neighboring pixels. Next all the neighboring pixels
are checked for homogenous intensity. If pixel is not
homogenous, then it is bounding pixel. Growing of the segment
continues until bound is locked.
Bounds of segments are extracted as radiometric edges.
2.2 Point cloud
Point cloud segmentation is performed after point cloud
loading. It is also region growing procedure. Segmentation
starts from grouping of cloud points. Curvature is calculated for
each group. For curvature calculation covariance matrix C is
calculated from point coordinates p and centroid p. (3).
pi — p à Pi p
Cs xu ug (3)
Zi Pp Pi p
Cov=A vw lo - =A +4 +4 (4)
ieN,
After that considering eigenvalue problem (4) eigenvector y
and eigenvalues à of matrix C are calculated. Curvature gis
calculated as:
A
OHS Ei
Ati +A, 6)
If curvature is below preset threshold (0.01), approximation
plane is calculated using least squares method. If deviation of
distances from points to approximating plane is lower then
14
threshold (0.01m), group of points becomes "center of
crystallization". All neighboring groups of points are iteratively
checked for planarity.
Point cloud segments are used for geometric edge extraction.
2.3 Image to point cloud referencing
Corresponding part of point cloud should be selected before
referencing. Approximating plane is calculated for selected part.
This plane is used for quasi image creation. The plane of quasi
image is coplanar to approximating plane; its center of
projection is far enough from the plane to fit all selected points
into one single image; size of image is calculated according to
scanning resolution.
Quasi image is similar to photographic, but could not be
successfully compared in automatic mode by any existing
algorithm. But edge extraction performed on point cloud and
photographic image makes possible comparing of these images.
Process of geometric and radiometric edge extraction described
above, besides of that Canny edge detection could be used for
radiometric edge extraction.
SIFT algorithm (Lowe 2004) is used for comparing of images
and tie point extraction.
Operator of PC controls mismatches, because most of facades
contain a lot of duplicated elements. Duplication of geometric
elements leads to point mismatching. Operator also able to
create matching points manually if automatic matching fails.
After tie points extraction external orientation parameters of
digital image are calculated using least squares method.
2.4 Vectorizing
After referencing digital images against point cloud geometric
edges could be projected onto plane of digital image. Geometric
edges are matched with radiometric edges. Matched edges are
saved as vectors. Operator can delete false matched vectors or
add new ones.
In practice total number of corrections could be very
considerable. It is almost impossible to achieve complete vector
model automatically regardless of used method. That is why
operator should have possibility to perform manual
vectorization and correction.
Operator creates vectors manually by photographic image. For
each node point spatial coordinates are calculated. After
calculation of two vector nodes, middle point is calculated. If
point to line distance is higher then threshold, middle point
becomes additional node point. New check of the same kind is
performed for two neighboring line segments. Check continues
until all nodes are fixed.
Spatial coordinates are calculated using interpolation of point
cloud points. Point cloud is projected onto plane of
photographic image. Distances from center of projection to
point cloud points are also calculated and stored for each point.
Two types of interpolation are available.
Dynamic interpolation is performed when operator moves
cursor: interpolation starts from searching of the nearest point in
the plane of image, then distance from the nearest point to the
center of projection is loaded, after that approximate point
coordinates are calculated according to loaded distance and
cursor 2D plane coordinates.
Another type of interpolation performed only after node point
creation, because this process needs much more time for
calculation. As the point cloud and the image are segmented,
corresponding segments could be identified. Depending on
position of the cursor and direction of move before point
fixation cursor could be identified as belonging to one or
another segment. After segment identification the group of
nearest points from segment selected. Selected group
approximated by plane. Spatial coordinates of the node point are
calc
ray
pla
3.1
For
pac
imp
Cor
savi
set,
visı
Poi
file:
ind
pro
ma)
onl
loa
The
init
ima
tie
ext:
cak
Orie
exti
mai
VIS
win
Tes
elei
Tor
dig
MI]
Fig
Fi
