The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B3b. Beijing 2008
597
skewing (a v ) of the vehicle shape. Therefore, for the most
cases the vehicle will appear as deformed parallelogram in the
scanning data - combination of both motion effects. In principle
the shape deformation of the vehicle can be used to
quantitatively derive the motion status, but a prerequisite must
be fulfilled that the true vehicle length is known and its
accuracy and sensitivity depend strongly on various impact
factors, such as point density, horizontal position error, or
physical properties of surface. In practice it is firstly not easy to
access the performance of this approach; a great amount of test
data is required to prove the feasibility and robustness.
4. GENERAL APPROACHES AND ANALYSIS
In the last two decades general approaches of 3D object
representation and recognition have been widely investigated in
the computer science community (Arman & Aggarwal, 1993;
Besl & Jain, 1986). Being different from the classical object
recognition, different methods, such as graph-based shape
matching/fitting and point pattern matching were developed to
directly conduct recognition process in the 3D range data.
Formerly, the most developed methods used to deal with the
small-scale dataset of reverse engineering. Due to the high level
of detail the smoothness, or curvature-based segmentation
algorithms are adopted to facilitate the recognition process;
ALS data over urban areas characterize the large coverage, very
complex scene and low LoD concerning the shape fidelity.
Rather than generic, application-specific methods have been
widely employed for object extraction, e.g. for building, road
and tree. The almost unique standard operation to ALS data is
filtering of non-ground points, motivated by topographical
mapping applications, to obtain the DEM.
Context object
Ç Fly-over^) Ç Road fopen area or yard) C Building Tree J
i ¡¡es 1.5- 3.5m as sampled ;
i leaves gaps ,
points above
• ..>
is dose to
casts shadow i hangs over
on and occludes
Vehicle object
f Single N Vehicle
Vehicle J V queue J
Relations
Figure 4. Context-relation model
A progressive processing strategy is proposed here to tackle the
problem of traffic monitoring using ALS data: by exploring
context relations. It is a direct processing chain being in
accordance with human inference. The structure of this
algorithm is organized based on the context-relation model
(Fig.4) retaining the knowledge of various object relations in
the urban area, where the dotted arrow indicates relation
direction. It is executed in a progressive and hierarchical way
controlled by our strategy for vehicle detection (Fig.5).
The algorithm starts with raw 3D point cloud of urban area.
Some basic operations for preprocessing laser data could be run
beforehand, like outline remove and hole filling. Then, a rough
separation of ground points and non-ground points is to be
carried out, using the height histogram thresholding or building
UDAR raw data
T
Ground point
separation
Ground points/non
ground points
4
Æ
W
DGM filtering
DGM
A
Filtering
vehicle points
r
w
Building labeling
* +
Vehicle point
candidates Building outline
Elimination of ^
non-vehicle points
Vehicle points
(VHM)
*
Segmentation
+
Vehicle objects
Am
Road layer
>
GIS Map
Vegetation region.,
Æ
V*
mr
Coarse vegetation
region delineation
.Data gap left by vehicle
above
Figure 5. Flow chart of method
labeling algorithm; the objective of this step is to mask out non
ground points or man-made objects like buildings where the
vehicle is assumed not to appear. The ground points are viewed
to build ground level surface in the urban which consist of not
only road but also courtyard. A smooth and continuous surface
could be imagined to be generated from the ground points as the
reference surface for ground level, being like the terrain surface
(DTM) after filtering, which can be represented in the form of
point cloud, surface meshing or analytical function. A height
interval of 0.5 to 2.0m over this ground surface is set to slice a
laser data layer 5j including all laser points p k regarded as
vehicle hypotheses (Formula.3). Afterwards, the vehicle
candidate points are delivered to the process eliminating
disturbing objects like tree points, wire pole, parterre or some
anomaly points. In order to distinguish between different
confused objects, various extern information sources such as
GIS/map can be used to mark building and road regions.
Vegetation regions can also be first delineated, beneath which
potential vehicles are to be searched and validated with special
efforts. The remaining laser points are transformed to vehicle
height model (VHM) in regular grid - normalized DSM for
vehicle, based on which single vehicle extraction and modeling
is to be carried out. Laser point gaps (holes) on the ground
surface left by impervious vehicles provide us another clue to
the presence of vehicle. All laser points belonging to or lying
within certain buffer interval (± 0.05m) of ground surface are
projected into a 2D plane regularly gridded, where cell size is
selected according to the laser point density. Each cell is
assigned with a value indicating either whether there are points
fall into the grid representing a small neighborhood, so the gap
will be retained as small dark area in this image. Some
experimental results are illustrated in Fig.6
Si = {p k e S : z k - z Sr(Xk n) < Ah^}
(3)