Masafumi Uehara
CONSTRUCTION OF URBAN SCENE MODEL
BY IMAGE SEQUENCE ANALYSIS USING DIGITAL MAP
Masafumi UEHARA, Takeshi SHIOSAKI, Heitou ZEN
Tokyo University of Mercantile Marine, Japan
{muehara, tshio, zen} @ipc.tosho-u.ac.jp
Working Group V/5
KEY WORDS: Motion Stereo, Epipolar Plane Image Analysis, Digital Map, DP Matching
ABSTRACT
Our group has researched about a method of the construction of urban model by Motion Stereo with motion images
took from a camera on a vehicle. But in the real city environment, it was difficult to construct the model accurately,
because a range of measured 3D data of buildings cluster and the slight motion of images due to the pitch, yaw and roll
from the vehicle motion cannot be removed completely. This paper proposes a method of the construct of urban scene
model with a large scale digital map and 3D measured data. A large scale map is expressed the topographical shape of
buildings accurate comparatively. Then, we established the relation between 3D measured data by Motion Stereo and a
digital map, and constructed the accurate and photorealistic city model adding height and texture data to each building.
1 INTRODUCTION
3D model of urban scene is made widely use in various fields, such as city engineering, traffic engineering and
amusement. Under present conditions, these models are made manually with CG modeling tools. Hence, huge time and
cost are wasted, and utilization of these models is restricted. Therefore, in order to construct model efficiently and
automatically, many approaches have been suggested. These approaches are roughly divided into two methods. One is
expression using stored 360 real images from observation point without 3D data[1][2]. This method is called IBR
(Image Based Rendering), it is possible to realize photorealistic urban scene. But on the other hand. it is necessary to
use a large quantity of data to utilize many images. The other is to make model using measured 3D date of buildings
with the stereo method or a laser measurement[3]. This method enables to move a viewpoint freely and construct model
with a small quantity of data. We: have constructed urban scene model based on 3D data, too. Our group proposed a
method of modeling using measured 3D data of buildings and texture data of target buildings with motion images took
from a camera on a vehicle[4]. 3D data were obtained by applying Extended EPI (Epipolar Plane Image) analysis that
had been improved over EPI analysis proposed by Bolles et al[5]. And texture data were obtained from the slit
spatiotemporal plane image from image sequence. Figure 1 shows the result of modeling with the proposed method for
our group. This method cannot separate every building because the construction result was only expressed the measured
3D data. And in the real city environment, it is difficult to measure 3D data of buildings because of the existence of
objects except buildings, such as a thick growth of trees, guardrails and poles. For reasons mentioned above, it was
difficult to construct the realistic model.
e eui ei aei
Figure 1: Previous result of modeling Figure 2: Example of digital map
912 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.