GEOMETRIC REFINEMENT OF LiDAR ROOF CONTOURS USING
PHOTOGRAMMETRIC DATA AND MARKOV RANDOM FIELD MODEL
A. P. Dal Poz
Dept, of Cartography, Sâo Paulo State University, R. Roberto Simonsen, 305, Présidente Prudente-SP,
Brazil - aluir@fct.unesp.br
Commission IV, WG IV/3
KEY WORDS: Digital Photogrammetry, Building Reconstruction, Spatial Analysis, Topographic Mapping, Optimization, LiDAR
ABSTRACT:
In this paper, a methodology is proposed for the geometric refinement of LiDAR building roof contours using high-resolution aerial
images and Markov Random Field (MRF) models. The proposed methodology assumes that the 3D description of each building roof
reconstructed from the LiDAR data (i.e., a polyhedron) is topologically correct and that it is only necessary to improve its accuracy.
Since roof ridges are accurately extracted from LiDAR data, the main objective is to use high-resolution aerial images to improve
the accuracy of roof outlines. In order to meet this goal, the available roof polyhedrons are first projected onto the image-space.
Then, the projected polygons and the straight lines extracted from the image are used to establish an MRF description, which is
based on relations (relative length, proximity, and orientation) between the two sets of straight lines. The energy function associated
with the MRF is minimized using a minimizing algorithm, resulting in the grouping of straight lines for each roof object. Finally,
each grouping of straight lines is topologically reconstructed based on the topology of the corresponding LiDAR polygon projected
onto the image-space. The preliminary results showed that the proposed methodology is promising, since most sides of the refined
polygons are geometrically better then corresponding projected LiDAR straight lines.
1. INTRODUCTION
Data acquisition for mapping and GIS using photogrammetric
techniques has traditionally been performed via the manual
extraction of cartographic features from images of the terrain
surface ranging in scale from 1:3000 to 1:90000 (Sowmya and
Trinder, 2000). Although manual extraction is adequate in
terms of accuracy and reliability, it is time-consuming and
expensive. On the other hand, due to imperfections in the image
acquisition phase and the scene complexity, feature extraction
from imagery and LiDAR data is too complex to be fully
automated.
Building extraction methodologies are very important in the
context of spatial data capture and updating for GIS
applications. These methodologies may be classified into three
categories according to the kind of input data, i.e.: LiDAR-
based methodologies, photogrammetrically-based
methodologies, and LiDAR/photogrammetrically-based
methodologies. An example of the first category is found in
Rottensteiner et al. (2005), in which an algorithm for roof line
delineation from LiDAR data is proposed. Basically, roof edges
and roof ridges are derived separately and combined to form a
consistent polyhedral model. Vosselman (1999) also described
another approach for the reconstruction of buildings by
polyhedron models from LiDAR data. Photogrammetrically-
based methodologies have been proposed for over 20 years. For
example, Fua and Hanson (1987) proposed a methodology for
locating and outlining complex rectilinear cultural objects
(buildings) in aerial images. In Shufelt (1997) is described the
PIVOT (Perspective Interpretation of Vanishing points for
Objects in Three dimension) system, which aims at
automatically extracting building from a single image. More
recently, Muller and Zaum (2005) proposed a methodology for
building detection in aerial images. First a region-growing
algorithm is used to segment the entire image and then the
buildings and vegetations are separated by a classification
procedure based on a set of geometric and photometric features
derived for each segmented region.
LiDAR/photogrammetrically-based methodologies seek to take
advantage of the synergy between LiDAR data and imagery
data. Basically, LiDAR-based techniques are superior in
deriving building heights and in extracting planar roof faces and
roof ridge lines, whereas photogrammetrically-based techniques
are superior in extracting building roof outlines (Kaartinen et al.,
2005). A few LiDAR/photogrammetrically-based
methodologies are found in the literature. Haala and Brenner
(1999) combined multispectral imagery and DEM (Digital
Elevation Model) derived from LiDAR data for separating
building from vegetation. In Sohn and Dowman (2003)
buildings are firstly extracted from both Ikonos imagery and
from DEM/LiDAR data and, then, the results obtained from
both data sources are combined to remove inconsistencies.
Vosselman (2002) combined LiDAR, plan view, and high-
resolution aerial image data to automatically reconstruct 3D
building. Basically, the plan view is used as reference to extract
polyhedral building model from LiDAR data. The high-
resolution aerial images are used to refine the roof boundaries.
In this paper, a methodology is proposed for the geometric
refinement of LiDAR building roof contours using high-
resolution aerial images and MRF models. MRF models have
been increasingly used in image analysis because they enable
the exploitation of the local statistical dependence of image
features and also allow global optimization to be accomplished
through iterative local computations. This makes sense
particularly in the context of roof building extraction because it
is not necessary that all straight lines interact with one another.
Instead, only a few straight lines that are spatially close to one