2. VISUAL DATA
In recent years, attempts have been made to derive traffic data
also from aerial images, because such images belong to the
fundamental data sources in many fields of urban planning.
Therefore, an algorithm that automatically detects and counts
vehicles in aerial images would effectively support traffic-
related analyses in urban planning.
2.1 Related work
Related work on vehicle detection from optical images can be
distinguished based on the underlying type of modeling used:
Several authors propose the use of an appearance-based,
implicit model [Ruskone et al., 1996], [Rajagopalan et al.,
1999] [Schneidermann & Kanade, 2000], [Papageorgiou &
Poggio, 2000]. The model is created by example images of cars
and typically consists of grayvalue or texture features and their
statistics assembled in vectors. Detection is then performed by
computing the feature vectors from image regions and matching
them against the statistics of the model features. The other
group of approaches incorporates an explicit model that
describes a vehicle in 2D or 3D, e.g., by a filter or wire-frame
representation [Burlina et al., 1995], [Tan et al., 1998], [Haag &
Nagel, 1999], [Liu et al.; 1999], [Liu, 2000], [Michaelsen &
Stilla, 2000], [Zhao & Nevatia, 2001], [Hinz & Baumgartner,
2001], [Moon et al., 2002]. In this case, detection relies on
either matching the model "top-down" to the image or grouping
extracted image features "bottom-up" to construct structures
similar to the model. If there is sufficient support of the model
in the image, a vehicle is assumed to be detected.
2.2 Vehicle model
For detecting single vehicles an explicit model is used.
Geometrically, a car is modelled as a 3D object by a wire-frame
representation containing substructures like windshield, roof,
and hood (see Fig. 1). An accurate computation of the car's
shadow projection derived from date, daytime, and image
orientation parameters is added to the model. As a radiometric
feature, color constancy between hood color and roof color is
included.
A detailed description requires a large number of models to
cover all types of vehicles. To overcome this problem a tree-
like model hierarchy is used having a simple 3D-box model at
its root from which all models of higher level of detail can be
derived subsequently.
2.3 Detection of single vehicles
The detection of single vehicles can be summarized by the
following steps: (1) Extract edge pixels and compute gradient
direction using Deriche's filter (ii) Project the geometric model
including shadow region to edge pixel and align the model's
reference point and direction with the gradient direction. The
projection matrices are derived from the image orientation
parameters. (iii) Measure reference color and intensity at roof
region. (iv) Adapt the expected saliency of the edge features
depending on position, orientation, color, and sun direction. (v)
Measure features from the image: edge magnitude support of
each model edge, edge direction support of each model edge,
color constancy, darkness of shadow. (vi) Compute a matching
score (a likelihood) by comparing measured values with
expected values. (vii) Based on the likelihood, decide whether
the car hypothesis is accepted or not. The evaluation measures
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
a b
Figure 1. Examples. a) Aerial image (section), b) Model
involved are explained in [Hinz, 2004]. An example of detected
vehicles is given in Fig. 2a.
2.4 Exploiting context
Due to the high geometric variability of vehicles, it can hardly
be assured that the detailed model described above covers all
types of vehicles. Only the contextual information that such a
vehicle stands on a road or is part of a queue makes it clearly
distinguishable from similar structures. For these reasons the
queue model incorporates more generic and more global
knowledge. Constraints of the detailed local model are relaxed
and, in compensation for this, the global consistency of features
is emphasized. More specifically, typical local geometric and
radiometric symmetries of vehicles are exploited and, in
combination with rough dimensions and spacings of vehicles,
they are constrained to form an elongated structure ("ladder-
like" shape) of sufficient length and smoothness. According to
this model, vehicle queue detection is based on searching for
one-vehicle-wide ribbons that are characterized by (i)
significant directional symmetries of grayvalue edges with
symmetry maxima defining the queue’s center line, (ii) frequent
intersections of short and perpendicularly oriented edges with
homogeneous distribution along the center line, (ii) high
parallel edge support at both sides of the center line and (iv)
sufficient length. More details concerning the symmetry
estimation are explained in [Hinz, 2004]
The results of the independent vehicle detection and queue
detection are fused. A mutual overlap of vehicles and queues is
checked and successfully tested vehicles are further
investigated for collinearity with the queue’s medial axis. After
fusion the queues are analyzed for missing vehicles. Such
vehicles are often characterized by homogenous blobs that can
be extracted by a region-growing algorithm. In the last step, all
vehicles detected using the stringent parametric model but not
being part of a queue are added to the result.
a 7 Fa b
Figure 2. a) Result of the model match. b) Vehicles detected
using local model (white) and vehicles recovered
through fusion with global model (black)
Intern
3.1 F
Infrar
the tr
be ca]
Hz:
incluc
Proje
explo
possil
The a
alone
may
typice
pollut
for tr
that !
throu
who :
featui
can b
Spm
Figur
3.2
Stati
they
alon
grou
vehi
stati
be ir
time
givii
that
are
traff
give