Detected car hy-
queues (black).
Survey on 3D-City
1997. Automatic
space Images (II).
tomatic Extraction
Birkhäuser Verlag,
Bennetze aus Luft-
l.
Imagery by Local
lericht, PF-2004-:
nische Universität
ieues in High Res-
dung — Geoinfor-
;:xtraction Integrat-
ves of Photogram-
nces, Vol. 34(3A),
C. and Ebner, H.,
(Baltsavias et al.,
and Their Link to
otogrammetry and
rification. In: Sth
pp. 148-154.
ion by Local Con-
cartographic data.
atic Car Detection
ve 3D Model. In:
Data Fusion over
Computer Vision.
)
1999. Knowledge-
ig Semantic Nets.
(7), pp. 811-821.
Toward Automatic
of Several Struc-
] Remote Sensing
^tworks. In: Inter-
and Spatial Infor-
jletion and Evalu-
f Photogrammetry
3. Istanbul 2004
THREE LINE SCANER IMAGERY AND ON-STREET PACKED VEHICLE DETECTION
S. Punvatavungkour ', R.Shibasaki *
CSIS, The University of Tokyo, CW503, 4-6-1, Komaba, Mekuro-ku, Tokyo, Japan, 153-8505 - (sompoch,
shiba)(g)skl.iis.u-tokyo.ac.jp
Commission VI, WG VI/4
KEY WORDS: Photogrammetry, GIS Acquisition, Object Detection, Segmentation, Pattern Recognition
ABSTRACT:
In the study, we proposed new algorithm of stopped vehicle detection using Three Line Scanner Imagery or TLS briefly. A
framework of our study consists of three stages: Pre-processing
, Moving/Stopping Vehicle Discrimination and Parking/Signals
Waiting Vehicle Classification respectively. The Pre-Processing Step is fundamental information preparation such as vehicle-likely
regions etc from Forward/ Nadir images. In Stopped/ Moving Vehicle Detection algorithms, stopped vehicles are detected by
grouping vehicle candidates and verifying candidates as detected stopped vehicle as 3d objects using the Stereoscopic measurement
on forward/nadir TLS images while moving vehicles are extracted by using our new ‘expansion proceed’ method to generate
moving vehicle candidate and validating candidates as detected vehicles using spatial-temporal techniques on forward/nadir TLS
images. In our last algorithm, Parking/ Idling Vehicle Classification, on-street parked vehicle is detected by measuring the distance
from the edge of the road to the side of stopped. Parked vehicle is agreed with this distance thresholding defined by a ground
observation. The algorithm is typically useful in cities. Finally, the promising results are derived and listed.
1. INTRODUCTION
Traffic congestions in urban area worsen quickly. On-street
vehicle statistics collection is very crucial. A practical and
effective approach is the vehicle observation by Remote
Sensing techniques because image from remote sensor
distributes a top-view and wide-area observation compared with
ground-based sensors as Video Camera etc and now, remote
sensing image is higher resolution. Those images are able to
distribute their merit for object identification.
Backward-Looking
Nadir
Foward-Looking
Figure 1 Three Line Scanner
Recently, Three Line Scanner, TLS, novel airborne line-
imaging sensor, is available. Three Line Scanner imaging
systems consist of three parallel one-dimensional CCD cameras
mounted on the imaging plane with a Stabilizer for shaking
reduction during imaging flight. It obtains seamless ultrahigh-
resolution images, with three viewing directions (forward, nadir,
backward direction) simultaneously with RTK-GPS and INS.
By seamless imaging capability along the road, TLS image is
355
very suitable to detect linear feature ground object such as road
etc. Therefore, in our contribution, by using TLS, on-street
vehicle monitoring has been developed.
The merit of TLS is not only distributing 3 dimensional
imaging but also distributing positioning of Sensor during
imaging by INS/GPS mounted with sensor. Moreover, due to
TLS image with centimetre order resolution, the small object
such as car pillar is identified. Because of tropical liner
seamless image, it is simple to detect ground linear objects such
as road etc from the image. Therefore, this image is suitable
with vehicle detection from TLS imagery
2. RELATED RESEARCHES
At the present, there is not much review on vehicle detection
using aerial image. All of them are categorized into a variety of
aspects such types of sensors, target vehicle types or types of
measurement etc. In terms of data, almost all existing methods
apply frame aerial images. Several authors have presented
approaches that utilize implicit vehicle models [1] [8]. Many
successful approaches use explicit models [3] [4]. Although
both approaches have merits and weak points, the implicit
model approach, which is based on radiometry, is limited due to
local radiometric disturbances and uncertainties about the
accuracy of data training, which varies with illumination,
viewpoint, and the types of objects in the training data [4].
Therefore, the explicit model is possibly more robust than the
implicit model.
However, there are neither any approaches that mentioned
stopped/ moving vehicle detection in one system by using one
type of data nor parked/ traffic signals waiting vehicle
discrimination
Please remind the organization of the paper. Chapter 2 reviews
existing work of vehicle detection by using aerial image and
their weak points. The overall structure of our contribution is
