Babak Ameri
Figure 1: Uncertainty buffer of homologous model edges in corresponding images
3.1. Linearity: A Local Internal Geometric Constraint. The estimation process is an orthogonal linear least
squares regression problem. It acts as a functional model with the objective to simultaneously minimize the perpen-
dicular sum of the Euclidean distances between the candidate edge-pixels and the projected 2D model edges in all the
images. A set of constraints is also integrated into the estimation model to support the minimization process. In fact, it
would be sufficient to simply solve a resection equation for modifying the coarse building model if we are able to find the
corresponding matches between the model points and their homologous points in the respective images. Thus, in order to
overcome the problem of feature corespondence, the match is actually established between the projected model edge and
candidate edge-pixels in the image. In other words, since the precise position of the endpoints of image edge is unknown
or is missing e.g., due to the occlusion, it is necessary to minimize only the perpendicular distances from the representative
points of an image edge to the projected model edge.
In order to measure the perpendicular distance between a representative edge-pixel (z ;, y;) and the projected 2D model
edge ey, it is useful to express the projected edge in image space in the following form:
€; ! T; Sin; — y: COS; — d; = 0 (1)
where d; is the distance between the origin and the line, and 0; expresses the angle between the edge and x axis. In
practice, the initial parameters of the 2D edge (85, d?) are obtained by back projecting the two endpoints of the 3D model
edge into the corresponding images using the collinearity equation (8). The computed 2D points are then plugged into the
equations (2), (3), in order to derive the initial 2D edge parameters.
Vend — Vstart
0; — arctan —— ————— (2)
Tend — Estart
dj = start SiINÔ; — Ystant COSÔ; (3)
The projected 2D edge e,, ;, in image I, serves as initial guess for finding the exact position of the edge model within the
corresponding images, An uncertainty buffer with a user specified width is generated around each edge model in image
space based on the initial position of the edge and is used as the search space to find the representative edge-pixels. Figure
1 shows the generated buffer around the homologous model edges of a reconstructed coarse building model in four images
taken from different view points.
Each pixel within the specified buffer is selected as a representative edge-pixel if it satisfies the following two conditions:
e its gradient direction is approximately perpendicular to the edge model direction, and
e the magnitude of its gray value gradient is more than a data-driven adaptive threshold. This threshold is computed
based on a cumulative histogram of the gradient magnitude of all the candidate pixels within the buffer, which satisfy
the first criteria. Thc gradient magnitude associated with each selected pixel is used as its weight in the estimation
model.
Applving these two conditions for the selection of representative edgc-pixcls which should satisfy equation (1) has the
following advantages. Firstly, pixels which are laid on the edgc image have stronger cffect during the fitting procedure
because they have more weight in the estimation process. Secondly, the gradients caused by background objects will not
interfere with the parameter estimations, as they are not in the approximate direction of the edge model. Figure 2 indicates
the selected edge-pixels of the homologous 2D model edges in corresponding images within the generated uncertainty
buffer in the first iteration of the estimation process.
At this point, after selecting the edge-pixels representative, we are ready to introduce the linearity constraints into the
estimation model. Let us represent the equation of the projected 2D edge in image 7 ,., passing through the selected edge-
pixel (z7^?, y; ?) in the following form:
img
/ A un Og. „Mg — „Mg
fir. (0,d) 2 2; sind.) — y; ? costi. —dij - eir; yi ) (4)
26 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
In th
shou
Linea
where
For e
is inse
The li
tive 2
(AG,
the ur
deterr
statist