OBJECT RECONSTRUCTION FROM IMAGES OF A MOVING CAMERA 
Jussi Heikkinen 
Institute of Photogrammetry and Remote Sensing 
Helsinki University of Technology 
Finland 
e-mail: Jussi. Heikkinen@hut.fi 
Commission V, IWG V/III 
KEY WORDS: feature based photogrammetry, feature matching, edge detection, hough transformation 
ABSTRACT 
This paper depicts a proposed method to model three dimensional objects from video sequence. The method is 
based on robust 2D features extraction, feature matching, and feature based photogrammetric modeling. The 
procedure to be presented can be considered as semiautomatic as its nature. During the whole process the 
operator will assist the system, but allowing the system itself to do all heavy routine tasks. The main idea is 
that we abandon using a point to point correspondence in object reconstruction and instead use a point to feature 
correspondence. In this case, a feature denotes a three dimensional object presented with a variable number of 
parameters. The parameters indicate the shape, size and position of object in selected coordinate system. This 
means that no form fitting is required in image space. The image observations resulted from edge detection can 
be used directly to estimate three dimensional object parameters. 
1. INTRODUCTION 
The idea of this procedure is to use all possible data of 
multiple video frames to solve the feature triangu- 
lation problem. The approach requires massive compu- 
tation, that's why much of the low level feature 
extraction is executed as a background process. Using 
video imaging means an enormous amount of frames 
even though the recording session lasts only few 
minutes. This leads to a storing problem of the video 
sequence images. By using compression and pro- 
cessing data in smaller sequences this problem can be 
overcome. 
CCD cameras have developed in few years and their 
stability has improved dramatically. Still, they do not 
reach the level of close-range analog photogrammetric 
cameras both in geometry and recording the bright- 
ness and contrast differences. The lack of precision of 
the camera can although be compensated by adding 
the amount of data. When using LSQ-type estimation 
the external accuracy is depending on stochastic 
properties as well as geometry. By adding obser- 
vations the effect of noise can be diminished. The easi- 
ness and inexpensiveness of digital recording gives us 
this possibility to add the number of images and 
through this the number of observations. Naturally, 
cameras are supposed to be calibrated before the 
recording session. 
220 
Automatization of feature extraction needs robust 
algorithms to succeed. Insensitivity of algorithms 
against noise and poor contrast is required. The edge 
detection is dealt in Chapter 2. The line features are 
used as main features as they are easier to identify. 
For pixel grouping Hough transformation is implied. 
Although line extraction is executed in a batch job and 
is not as time restricted as in on-line systems, we are 
trying to speed up the process due to the amount of 
data. We are using the RHT-algorithm (Random 
Hough Transformation) for finding line features on 
image. This algorithm is based on random sampling 
and convergence mapping cycles succeeding each 
other. The method does not need as huge storage 
resources and work load as other stochastic Hough 
transformations but is still able to localize lines accu- 
rately. The RHT-algorithm is depicted more precisely 
in Chapter 3. 
In Chapter 4 problems of finding the corresponding 
lines in succeeding frames will be considered. The low 
level feature extraction is done in the most automated 
way. In this paper we are not trying to present a fully 
automated measuring system, but to apply automatic 
methods in areas where tasks are easy to automate 
and need lot of routine work. Operator is used here as 
a guidance help. Operator’s duty is to pick up all 
interesting lines on one image and add some con- 
straints like parallelism, intersection of 3D-lines, per- 
pendicularity etc. The automatic feature matching 
procedure will do the rest. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996 
The 3D-fe 
presented 
feature be 
use liner 
correspon 
correspon 
measure i 
belonging 
feature, 1 
adjustmet 
tations a: 
feature li 
formulati 
geometric 
lications. 
single pix 
no line fc 
features 2 
kind of f 
and other 
the photc 
vations ai 
the most 
Actually, 
triangula: 
circles ha 
line inte 
triangula 
The featt 
work for : 
be measu 
measurer 
the oper 
extractior 
find and « 
This pre 
measurin 
which ha 
collision 
modified 
The thre 
using pi 
points mt 
pixel cooi 
of the ob: 
results s 
coordinat 
accuratel 
how inve 
result of 
“goodnes: 
errors ha 
some rob 
transforr 
out gross 
In this r