Paolo Gamba 
  
MODEL INDEPENDENT OBJECT EXTRACTION FROM DIGITAL SURFACE MODELS 
Paolo GAMBA', Vittorio CASELLA 
"University of Pavia, Italy 
Department of Electronics 
gamba@ele.unipv.it 
"Department of Building and Territory Engineeering 
casella@unipv.it 
KEY WORDS: DSM, building extraction, urban characterization. 
ABSTRACT 
This paper is devoted to analyze and improve a recently introduced technique that may be applied to Digital Surface 
Models (DSMs) produced by different sensors (Interferometric SAR, for instance, or LIDAR, or even photogrammetry) 
for object detection and extraction. The proposed approach, based on modified image analysis techniques, does not need 
to specify any surface or object model of the object to be recognized. To this aim, we apply a jointly regularization and 
segmentation algorithm to the original data, so that it is easy to partition them into significant structures, the 
background, and uninteresting areas from a 3D point of view). The criteria applied to segment the data are geometric 
ones, involving the principle of segment and plane-fitting. : 
By means of this approach we will show that it is possible to better characterize urban scenes in terms of buildings, 
trees, roads and other natural and geometrical structures. In particular, looking at their 3D structure we may be able to 
discard natural objects without referring to given object models. The original data resolution is exploited as much as 
possible, and the method provides a high robustness to noise. 
1 INTRODUCTION 
Urban three dimensional (3D) geometry and land cover are among the widely required for an efficient urban analysis. In 
turn, this analysis may be useful for a number of applications, including urban monitoring, change detection, damage 
assessment, as well as traffic and heat modeling, and finally urban scene visualization for architectural purposes. 
Currently, only a few sensors are able to produce data sufficiently accurate for such an analysis. In particular, Digital 
Surface Models (DSMs) for this complex environments may be computed starting from aerial stereo high resolution 
photographs, from Laser Ranging (LIDAR) systems and from Intereferometric Radar (IFSAR) measurements. These 
different sources provides differently characterized DSMs, so that their analysis requires the capability to cope with a 
number of different problems and noise sources. 
This paper is devoted to develop and test some recently introduced techniques that may be indifferently applied to these 
DSMs for a suitable object detection and extraction. These techniques are based on machine vision and pattern 
recognition techniques, and do not need to specify a surface or object model to which the data is compared. Indeed, to 
this kind of parametric approach may be reduced most of the systems recently proposed in literature (Gruen and Wang, 
1998, Gruen, 1998), and only a few non-parametric algorithms may be also found (see the second part of Haas and 
Vosselman, 1999, for instance). The goal of this paper is instead to characterize urban scenes in terms of buildings, 
trees, roads and other natural and geometrical structures looking at their 3D structure, but without the reference to a 
given model. 
More in detail, for extracting geometrical structures we apply regularization and segmentation algorithms to the original 
Digital Surface Models. The goal is to exploit their resolution, maintain a high robustness to noise, and partition the 
original data into background and significant objects. In particular, the criteria applied to segment the data are geometric 
ones, involving the principle of segment and plane-fitting. With this approach, objects and structures may be 
characterized as a set of planar regions whose relationships may be used, in a second time, for a model-driven 
recognition and refinement step. This approach corresponds to looking for the building roofs and walls, discarding 
noisy measurements by a clustering algorithm that aggregates points belonging to the same surface by means of a 
region growing technique. 
  
312 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
  
N