ENHANCED MULTI-CLUSTER ANM FOR STEREO MATCHING
M. Sakamoto *”, W. Lu*, Y. Kosugi"
? Asia Air Survey Co., Ltd. R&D Department, 8-10, Tamura-cho, Atsugi-shi, Kanagawa 243-0016, Japan -
(mi.sakamoto, luwei)@ajiko.co.jp
? Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, 4259, Nagatsuta-cho,
Midori-ku, Yokohama 226-8503, Japan - kosugi@pms.titech.ac.jp
KEY WORDS: Stereo Matching, Adaptive Nonlinear Mapping, Multi Clustering, Region Division, Digital Surface Model
ABSTRACT:
For the purpose of economically reconstructing digital surface model (DSM), stereo matching technique is still one of the most
effective approaches. However since most of the stereo matching models all assume the continuity of a target, they cannot guarantee
practical results in regions such as urban area, where altitude of surface change abruptly. This paper tackles stereo matching problem
in such regions by Adaptive Nonlinear Mapping (ANM) approac
h. Originally ANM was established as a mapping technique for
modelling of self-organization process in neural network and was also applicable to forming smooth mapping in the past. It
introduces constraint of edges in the image and the concept of area-based clustering, to extend ANM for application of stereo
matching toward undulating region where pa
rallax shifts are abrupt. Preliminary verification test were executed for examine the
characteristics of proposed model. The results indicated that the proposed approach for improving mapping results in linear
structured region such as building's boundaries was effective.
1. INTRODUCTION
In the digital photogrammetric technology, it is important to
reconstruct digital surface model (DSM) of high accuracy.
Especially realization of automatic mapping techniques for
geographical feature is in great demand. On the other hand,
nowadays more economic DSM of wide area has become
available by using airborne laser profiler. However in the city
area it is difficult to acquire DSM with sharp building edges
because of the influence of spatial sampling frequency, irregular
reflection, occlusions by ground features and so on. To solve
these problems, edge information extracted from edge detecting
and matching results that reflect building’s shape or improved
stereo matching results in edge region is needed for
modification of DSM.
Most of the existing practicable approaches for stereo matching
are based on models in which continuity of topographical shifts
is the primal condition. Therefore they are effective for natural
terrain or images of low resolution, but not suitable for
drastically undulating region such as artificial area of high
resolution images.
In this study we introduce a mapping technique named
Adaptive Nonlinear Mapping (ANM) for improving stereo
matching results in artificial terrain. ANM is a technique based
on the principle of Coincidence Enhancement (CE), which is an
extended concept of Hebb's rule on self-organization process in
neural network model (Kosugi, 1993). Different from existing
stereo matching approach such as area-based (Moravec, 1981),
feature-based (Marr, 1979) or intensity-based (Horn, 1986),
ANM enables easy implementation of feedback mechanism for
matching. The original ANM creates smoothed mapping results
as well as usual approaches, therefore it is difficult to apply this
technique as it is. Thus we have developed the ANM model
with edge constraint, which allows discontinuity of shifts in
mapping and improved matching results surrounding building's
edges (Sakamoto, 2001). However, such improvement is limited
to regions where edges were detected or matched.
In this paper we propose a further enhancement of ANM model
by area-based multi-clustering. It improves mapping results by
regional clustering of shift vectors. This approach is effective
for building regions where edges are not detected satisfactorily.
2. STEREO MATCHING WITH ANM
2.1 Feature of ANM
There exist many well-known stereo matching algorithms such
as area-base ones represented by template matching, feature-
based ones which pay attention to feature points, edges and so
on, and the intensity-based one, which is also known as optical
flow estimation. In order to improve the stability of matching
results, these algorithms are usually used in combination with
approaches such as coarse-to-fine or relaxation of matching
results for smooth shifts. Another famous approach is Dynamic
Programming (DP) technique, which ensures unique solution by
minimizing the matching cost in the system. Most of these
approaches excessively depended on the assumption of smooth
matching shifts, because of lacking conceptual model for abrupt
shifts or feedback, and the modification process for matching
models.
The concept of Adaptive Nonlinear Mapping (ANM) adopted in
this study is originally developed for the control of self-
organization mechanism in neural network model, and here
applied for the stereo matching problems. This approach has a
feedback process and can realize assumption based matching
models. Different from DP matching, ANM evaluates matching
costs in each local area, and then propagates its influence to the
system.
2.2 Principle of ANM
In ANM approach, a principle called Coincidence Enhancement
(CE) is proposed as the fundamental rule for the explanation of
self-organization process, which is an extension of Hebb's rule
_R4-
(Ke
cot
act
ste]
cor
illu
are
suf
val
nur
eva
FC
whe
is a
pos
ÿ
For
vec
fun
inst
In «
for
erro
func
con
max
othe
corr
stati
com
For
map
cons
smo
maii
ofn
cons
d^
whe
an e
vect
If er
strer
