716 
LASER RADAR IMAGES ITERATIVE SEGMENTATION TOR DYNAMIC SCENES 
V.fc. fiis.i tsyn, K. V. Obrosov, N « IT. Pa ne army, V. A. Stephanov 
The Research Institute of Automatic Systems, USSR, ISPRS Commission VIT 
ABSTRACT 
The paper deals wxth Doppler optica] images (DOI) segmentation technique based on 
a posteriori probability maximum criterion. The specific feature of speckled image 
formation is an inequal accuracy of different pixels Doppler frequency measurement 
due to echo amplitude fluctuation. The dual stochastic model is used for .image des— 
cription. A posteriori probability calculation equations are derived. The paper 
presents binary segmentation algorithms for plane-parallel and arbitrary motion of 
objects being sensed when sensor moves relative to underlying surface. The mathema 
tical modeling results are presented which prove algorithm efficiency. The computa 
tional burden of segmentation algorithins implementation on mainframe digital compu 
ter is assessed. 
Key Words: Doppler speckled images, binary segmentation, dual stochastic model. 
1. INTRODUCTION 
A new class of information systems - 
Doppler optical systems with coherent 
radiation and heterodyne echo reception 
- provides aii effective means of remote 
sensing. These systems are the result 
of radar and electro-optical sensor 
technology fusion. They combine optical 
techniques and radar signal analysis 
and processing methods.(Hull,1982). 
The systems of this type offer a number 
of important advantages and are parti 
cularly suited for applied problems 
solution, including automatic search, 
detection and recognition of small-size 
mobile objects against complex back 
ground. (Harney , 1 981 i Harney,I960; Prey, 
1983). 
The solution of automatic object reco 
gnition problem based on digital image 
processing techniques requires a preli 
minary segmentation of a scene being 
sensed which is the image partitioning 
into uniform by some attribute regions 
(Sullivan,1980). DOI segmentation is 
traditionally implemented through se 
lection of pixels with -frequencies ex 
ceeding the specified threshold which 
is common in radiolocation (Pney,1983). 
This Doppler shift frequency selection 
may he complemented by amplitude selec 
ti Oil. 
The traditional approach to DOT segmen 
tation, however, often gives extremely 
unsatisfactory results as doesn't ac 
count for spatial extension of object 
images, correlation between object pix 
els frequencies and some pr ! ncipal pe 
culari ties of 1)01 formation. 
2. THE PROBLEMS OP DOPPLER 
OPTICAL IMAGES SEGMENTATION 
The so called speckle-effect is inhe 
rent in real DOI. It is related to ra 
diation coherence and manifests itself 
by exponential random DOi pixel inten 
sity distribution (Sullivan,1980;Lota- 
lick, 1986; Goodman,1985). 
p(A)Jt tv №} at 4> '° 
L 0 otherwise 
where A - is the mean intensity value 
defined by object or backgi'ound reflec 
tion factor. One-dimensional probabili 
ty distribution density provides the 
comprehensive characteristic of speckle 
structure so far as echo intensity flu 
ctuations in neighbouring pixels are 
statistically independent (Sullivan, 
1980; Wang,1984). The DOI speckle struc 
ture essentially effects echo frequency 
assessment. When sensor internal noise 
level A„ is fixed, the frequency measu 
rement error in each DOT element in a 
first approximation is defined by (Dan- 
sac, 1985) 
6 F ~ i/T/A/An (D 
where T - pixel analysis time. 
Thus the problem of statistically opti 
mal DOI segmentation procedure synthe 
sis arises. The procedure must account 
for extension of objects being analysed, 
their frequency images correlation pro 
perties, probabilistic characteristics 
of measurement results, a priori object 
shape information and correlation bet 
ween received signals frequency and in 
tensity assessments. 
There are many heuristic approaches to 
segmentation (.Bopil<Jt:HnO,I987 ), Though 
attractive regarding applied problems 
solution these approaches have obvious 
limitations in terms of quantity asses 
sment of segmentation procedures quali 
ty, stability, reliability and computa 
tional complexity. Recently some papers 
have appeared (Therrien,1986; Derin, 
1986), in which statistical evaluation 
theory is applied, to segmentation prob 
lems. 
3. DOI SEGMENTATION PROBLEM STATEMENT 
Let DOI is presented as a frequency mat- 
rix HU with dimensions and 
as an intensity matrix /4= . E sta 
tistical characteristics are given. Let 
the typos of image regions being seg 
mented are predetermined and nave num 
bers from 0 to R-1. Suppous each (x,y) 
pixel may take arbitrary state S X yequal 
to one of the region type numbers. We