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 
LEGION SEGMENTATION FOR BUILDING EXTRACTION 
FROM LIDAR BASED DSM DATA 
Chun Liu**, Beiqi Shi^^' , Xuan Yang? and Nan Li 
‘Department of Survey and Geo-Informatics, Tongji University, Shanghai, China, 200092 liuchun@tongji.edu.cn 
"Key Laboratory of Advanced Engineering Surveying of NASMG, Shanghai, China, 200092 
‘Urban Information Research Center, Shanghai Normal University, Shanghai, China, 200234 carashi@163.com 
Commission III, WG III/4 
KEY WORDS:LiDAR DSM, LEGION segmentation, Building Extraction, Height Texture 
ABSTRACT: 
Recently, a neural oscillator network based on biologically framework named LEGION (Locally Excitatory Globally Inhibitory 
Oscillator Network),which each oscillator has excitatory lateral connections to the oscillators in its local neighbourhood as well as a 
connection with a global inhibitor, has been applied to segmentation field. The extended LEGION approach is constructed to extract 
buildings digital surface model (DSM) generated from LiDAR data. This approach is with no assumption about the underlying 
structures in DSM data and no prior knowledge regarding the number of regions. Instead of using lateral potential to find a major 
oscillator block in original way, Gray Level Co-occurrence Matrix (GLCM) homogeneity measuring DSM height texture is applied 
to distinguish buildings from trees and assist to find LEGION leaders in building targets. Alongside the DSM height texture attribute, 
extended LEGION can extract buildings close to trees automatically. Then a solution of least squares with perpendicularity 
constraints is put forward to determine regularized rectilinear building boundaries, after tracing and connecting the rough building 
boundaries. In general, the paper presents the concept, algorithms and procedures of the approach. It also gives experimental result of 
Vaihingen A2 region by then ISPRS test project and another result based on a DSM data of suburban area. The experiment result 
showed that the proposed method can effectively produce more accurate buildings boundary extraction. 
1. Introduction 
Building representations are needed in a variety of applications, 
such as cartographic analysis, urban planning, and visualization. 
And the development of building automated extraction 
algorithms is of great importance. Since LiDAR is a fast 
method for sampling the earth’s surface with a high density and 
high point accuracy, many attempts have been made on 
building extraction from a digital surface model (DSM) 
generated from LiDAR data. Wang and Schenk (2000) generate 
the triangulated irregular network (TIN) model from the 
LiDAR point clouds. Triangles are then grouped based on the 
orientation and position to form larger planar segments. The 
intersection of such planar segments results in building corners 
or edges. Al-Harthy and Bethel (2002) determine the building 
footprints by subtracting DTM from DSM obtained by initially 
filtering out the non-ground points. The building polygon 
outline is then obtained by using a rotating template to 
determine the angle of highest cross-correlation, which 
suggests the dominant directions of the building. Miliaresis and 
Kokkas (2007) presented a new method for the extraction of a 
class for buildings from LiDAR DEMs on the basis of 
geomorphometric segmentation principles. It is difficult to 
remove vegetation in urban or suburban areas. Most popular 
approaches were to detect buildings by fusing LIDAR data with 
multi-spectral images(Walter ,2004; Lu et al., 2006; Li et al., 
2010). However, fusing LiDAR data with multi-spectral data 
with different resolutions may add errors to building detection 
  
: Corresponding author: Beiqi Shi, Ph.D. candidate, research field in neural oscillator network and its application 
(Tullis and Jensen, 2003), the purpose of this paper is to 
develop an alternative automatic building extraction method 
based only on LiDAR data. 
We use an extended neural oscillator network approach for 
segmenting LiDAR DSM imagery into semantically 
meaningful entities and extracting buildings objects. This is 
based on temporal correlation theory to address the binding 
problem by using a biologically plausible representation. The 
process consists of a sequence: After generating DSM, a neural 
oscillator network based on biologically framework named 
LEGION segmentation is constructed and applied to extract 
buildings from DSM; the rough building boundaries are traced 
and connected; in the final step, all boundary points are 
integrated in a least squares solution with perpendicularity 
constraints to determine a regularized rectilinear boundary. The 
experiment on the given data provided by the "ISPRS Test 
Project on Urban Classification and 3D Building 
Reconstruction" verified that the proposed method can produce 
accurate buildings boundary extraction. 
2. Methodology 
The purpose of LIDAR DSM segmentation is to separate the 
DSM data into different classes depending on specific 
application requirements, such as building extraction. The 
region segmentation method, using height differences between 
neighboring grid points checked against a predetermined 
threshold value, has difficulty in segmenting DSM into 
semantically meaningful entities. The neural oscillator network