Estimating 3D Polyhedral Building Models by Registering Aerial Images - Advanced Concepts for Intelligent Vision Systems - page 239

is obtained either manually or automatically [5]. We were inspired by the feature-
less image registration techniques where the goal is to compute the global motion
of the brightness pattern between them (e.g., ane or homography transforms)
without using matched features [6].
Unlike existing approaches for building reconstruction, our approach derives
the polyhedral building model by minimizing a global dissimilarity measure
based on the image rawbrightness. It is carried out using a genetic algorithm. To
the best of our knowledge the use of featureless and direct approaches has not
been used for extracting polyhedral models of buildings. In any feature-based
approach, the inaccuracies associated with the extracted features, in either 2D
or 3D, will inevitably affect the accuracy of the final 3D model. Most of the
feature-based approaches use sparse extracted features such as interest points
and line segments. Thus, the sparseness of data coupled with noise will definitely
affect the accuracy of the final building reconstruction.
Recently, many researchers proposed methods for extracting polyhedral mod-
els from Digital Elevation Models (DEMs) (e.g., [3, 7, 8])). Compared to these ap-
proaches, our method has the obvious advantage that the coplanarity constraints
are implicitly enforced in the model parametrization. On the other hand, the ap-
proaches based on DEMs impose the coplanarity constraint on the 3D points of
the obtained surface in the process of plane fitting. DEMs are usually computed
using local correlation scores together with a smoothing term that penalizes large
local height variation. Thus, correlation-based DEMs can be noisy. Moreover,
height discontinuities may not be located accurately. In brief, our proposed ap-
proach can give more accurate 3D reconstruction than feature-based approaches
since the process is more direct and does not involve intermediate noisy data
(e.g., the 3D points of a noisy DEM).
Although the proposed method can be used without any DEM it can be useful
for rectifying the polyhedral models that are inferred from DEMs. In this case,
our proposed method can be useful in at least two cases. The first case is when the
provided model is erroneous, e.g., a facet is not modelled. Figure 1 illustrates two
corresponding examples of erroneous estimated polyhedral models. The second
case is when the estimated shape is correct but its geometric parameters are not
accurate enough, e.g., the coordinates of some vertices are not very precise. In the
latter case our proposed approach can be used for improving the accuracy of the
model parameters. The remainder of the paper is organized as follows. Section 2
states the problem we are focusing on and describes the parametrization of the
adopted polyhedral model. Section 3 presents the proposed approach. Section 4
gives some experimental results.
2 Problem Statement
Since aerial images are used only roof models can be estimated. In this work,
we restrict our study to simple polyhedral models that are illustrated in Fig-
ure 2. The model illustrated in Figure 2. (a) can describe a building roof having
two, three, or four facets. This is made possiblesincethe3Dcoordinatesofthe