Image Processing in Optical Guidance for 
Autonomous Landing of Lunar Probe 
                                 Ding Meng 1   Cao Yun-feng   Wu Qing-xian 1 and  Zhang Zhen 1 
2
1School of Automatic Engineering 
NanJing University of Aeronautics and Astronautics 
                           E-mail:nuaa_dm@hotmail.com 
2 Academy of Frontier Science 
NanJing University of Aeronautics and Astronautics 
E-mail:cyfac@nuaa.edu.cn 
 
Abstract 
Because of the communication delay between earth and 
moon, the GNC technology of lunar probe is becoming more 
important than ever. Current navigation technology is not 
able to provide precise motion estimation for probe landing 
control system Computer vision offers a new approach to 
solve this problem. In this paper, author introduces an image 
process algorithm of computer vision navigation for 
autonomous landing of lunar probe. The purpose of the 
algorithm is to detect and track feature points which are 
factors of navigation. Firstly, fixation areas are detected as 
sub-images and matched. Secondly, feature points are 
extracted from sub-images and tracked. Computer simulation 
demonstrates the result of algorithm takes less computation 
and fulfils requests of navigation algorithm. 
Key words: Fixation Area, Feature Point, Template 
Match, FPs Tracking 
I. INTRODUCTION 
n paper[1], the status of China’s deep space exploration 
plan is introduced including CE-1 lunar orbiter, the china’s 
subsequent Lunar Exploration Program. It is an important 
purpose in the second stage of china lunar exploration to land 
accurately of probe on the moon’s surface.  
The guidance-navigation-control(GNC) technology of 
moon probe is becoming more important than ever. Because 
of the communication delay induced by the large distances 
between the earth and moon, human kinds are hardly able to 
guide probe to landing safely on the moon. Probe will have to 
use on-board sensors and algorithms. However, current 
navigation technology can’t provide precision motion 
estimation for probe landing control system[2]. Therefore, 
new motion estimation technology must be developed. 
Computer vision may offer an approach to solve this 
problem.  
 
Manuscript received september 1, 2007.  
Ding Meng received B.E. and M.E degrees in School of Automatic 
Engineering of NanJing University of Aeronautics and Astronautics in 2003 
and 2006. He is currently working toward  Ph.D. in the same university. His 
interests are Computer Vision and Pattern Classification. Phone: 
86-025-84896392;  
Cao Yun-feng received Ph.D in School of Automatic Engineering of 
NanJing University of Aeronautics and Astronautics in 2005. Now, he is the 
professor in Academy of Frontier Science of NanJing University of 
Aeronautics and Astronautics.  His interests are Flight Control System and 
Intelligence Control. Phone: 86-025-84890902;  
 
 
 
Up to now, computer vision has been successfully used 
for autonomous navigation of robot and Micro Air Vehicle 
[3], autonomous landing of Unmanned Helicopter[4], and 
intelligence traffic[5].  
Some appliances are being researched in the field of 
space exploration. The Near Earth Asteroid Rendezvous 
(NEAR), rendezvoused with asteroid Eros 433 in February 
2000, used optical navigation extensively for orbit 
determination and small body 3-D modeling[6]. The Deep 
Space 1 mission as a part of the New Millennium Program is 
flying an autonomous optical navigation technology 
demonstration. The DS-1 AutoOpNav system used onboard 
centroiding of reference asteroids for autonomous navigation 
during small body fly-bys [7]. However, for probe accurate 
landing, the new, more robust and more precise vision 
algorithm should be developed.   
         Image processing is the first step and a essential part of 
computer vision navigation. Generally, image processing of 
vision navigation study is mainly focused on feature 
detection and tracking. In 1997, Kawaguchi, from Japan, 
created a method, in which several fixed reference windows 
and a 2-D fast Fourier transform (FFT) algorithm were used 
in the correlation tracking[8]. In many case, flat areas must be 
chosen 
for 
landing 
security. 
Safe 
Landing 
Area 
Determination (SLAD),which offered by A. Johnson from Jet 
Propulsion Laboratory(JPL), achieved this task successfully 
[9]. However, there are no features or distinctive features on 
those landing sites. Although tracking by feature points is not 
the best way, feature points haved to be filtered in order to 
avoid this problem. Point tracking is the usual way in vision 
navigation .In 1999, Andrew E. Johnson, from JPL, 
generated a method based on automatic feature point(FP) 
tracking between a pair of images followed by two frames[2]. 
As a part of MUSES-C, Japanese scientist Toshihiko Misu 
offered an algorithm based on fixation area, which fixates on 
the areas with local varying intensity, and employs shading 
pattern as a fixation area, and then tracked the center of the 
fixation area[10]. Real-time image processing is the essential 
request in computer vision navigation. Toshihiko’s method 
requires less computation than Johnson’s. Nevertheless, the 
Number of fixation area’s center is difficult to satisfy the 
requirement of following navigation algorithm. This paper 
introduces an image processing algorithm using feature point 
tracking in probe autonomous landing which built on the 
methods from Johnson and Toshihiko. The procedure is as 
follow (Fig1): 
I 
ICIUS 2007
Oct 24-25, 2007
Bali, Indonesia
ICIUS2007-A001
ISBN 978-979-16955-0-3
1
© 2007 ICIUS
 
 
 
Step1. Extract fixation areas as templates and sub-images. 
Step2. Match templates and detecting FPs from sub-images 
Step3. Match FPs  
In section 2, fixation area detection algorithm is proposed. 
Feature point detection and tracking in the sub-images are 
explained in section 3 and section 4. Computer simulations of 
this algorithm are illustrated to demonstrate this image 
processing algorithm in section 5. In Section 6 we state 
conclusions. 
 
Fig.1 Procedure of algorithm 
II. FIXATION AREA DETECTION 
In artificial environment, edges and vertexes are usually 
used as features in image tracking. But in natural 
environment, these shapes can’t be found easily. Paper[11] 
proposed the use of ground control points based on Gabor 
function in image registration of natural terrain. In paper[10], 
shading pattern is employed as a fixation-area. To reduce 
computation, the averaging, Laplacian filtering and variance 
are used. In this paper, improved Toshihiko Misu’s method is 
used firstly to pre-process primary image for obtaining 
sub-image. The purpose of this step is that block matching is 
more robust than point matching. We can match block firstly 
in entire image and then match points in sub-image.  
The template for matching block is extracted as 
follows[10]: 
Step1. Enhance specific spatial wavelength of the original 
image by 2-D band-pass-filter. 
Step2. Calculate local variances of filtered image to 
evaluate ”contrast”. 
Step3. Extract high local-variance areas as templates. 
Step4. Least-square block matching. The matching is 
performed so as to minimize square error of the intensity 
between templates and region to be matched.    
2.1Band-Pass-Filter 
The conception of filtering comes from the Fourier 
transform for signal processing in the frequency domain. 
Image processing is interested in filtering operations that are 
performed on the pixels of a digital image. The filter in the 
image processing is called spatial filtering. In generally, the 
image filter can be divided into smoothing spatial filter and 
sharpening spatial filter.  
Firstly, original image should be processed by 
smoothing spatial filter to reduce noise. In fact, smoothing 
spatial filter is a type of low-pass-filter(LPF), includes linear 
filter, such as averaging filters, and nonlinear filter, such as 
median filters. To reduce the consumption of computational 
power and storage, we use averaging as LPF, which is 
different from averaging filter. The averaging filter can’t 
change the size of original image and spend more time. The 
equation of averaging is: 
1
1
0
0
1
( , )
(
,
)
y
x
S
S
s
x
i
j
x
y
y
E u v
E S u
i S v
j
S S
−
−
=
=
=
+
+
∑∑
       (2.1) 
Where 
 is the intensity of averaged and 
sub-sample image at
. The size of this image is 
( , )
s
E u v
( , )
u v
1
x
y
S S
of 
size of original image. (
,
x
y
S S ) is the sub-sampling interval. 
 Secondly, the 8-neighbor Laplacian is used as 
high-pass-filter. The mask of Laplacian as follows: 
 
2.2Variance Map 
To show locally varying intensity, the variance map is 
calculated. Statistical variance within a window and the size 
of which is equal to computed template is computed. The 
formulation is: 
1
1
1
1
2
2
0
0
0
0
( , )
{
(
,
)}
{
(
,
)}
y
y
x
x
w
w
w
w
x
y
x
y
L
L
i
j
i
j
x
y
x
y
S S
S S
V u v
E u
i v
j
E u
i v
j
W W
W W
−
−
−
−
=
=
=
=
=
+
+
−
+
∑∑
∑∑
+
Where 
 is the local variance of Laplanian-filtered 
image. 
( , )
V u v
x
y
W
W
×
 is the size of template. 
is the 
intensity 
of 
Laplanian-filtered 
image 
at 
.
( , )
L
E u v
( , )
u v
,
.
y
x
x
y
x
y
W
W
w
w
S
S
=
=
 
2.3Template extraction 
Templates for tracking are extracted according to 
variance map. The high-local-variance points are selected as 
the template. We should pay attention: a.The points chosen 
as template in the variance map are not near the edge of the 
map, in order to easily track points in next frame image. 
b.Some other templates which aren’t near extracted ones 
should be extracted also. This is because the cluster or single 
template would be weak to the observation noise which 
includes tracking error and error from the range 
measurement. 
2.4Template matching 
To track the fixation area, we employ least-square 
block matching. The matching is performed so as to minimize 
the square error of the intensity between template and region 
to be matched. The number of templates should be certain. If 
any template which has been out of the image can’t be 
tracked, the new template is extracted instead.   
 
ICIUS 2007
Oct 24-25, 2007
Bali, Indonesia
ICIUS2007-A001
ISBN 978-979-16955-0-3
2
© 2007 ICIUS
 
 
 
III. FP DETECTION FROM FIXATION AREA 
To best our knowledge, some real-time FPD 
algorithms have been described in many literatures. FPD 
based on Harris Corner Finding algorithm has been employed 
by ASSET-2 motion, segmentation and shape tracking 
system[12]. After detecting the fixation area, paper[10] 
assumed the center of the fixation area as an feature point. I 
don’t think this is a good idea. In this paper, a FPD algorithm 
is selected to detect feature points from fixation area. We can 
use those points for match. 
In this paper, we select and improve a FPD algorithm 
of Benedetti and Perona[14], which is based on the work of 
Tomasi and Kanada[13]. There are two reasons to select this 
algorithm.  
a)This algorithm decreased the complexity of Tomasi’s 
method by eliminating the demand for any transcendental 
arithmetic operations and didn’t require calculation of square 
root. 
b)This algorithm is easy to track across several frames.  
       It is difficult to detect and an individual point because of 
the presence of noise and distortion. For this reason patches 
of pixels are actually considered. The method proposed in 
paper[14] for detect FP is expressed simply as follows: 
Step1.Computing the image gradient across the image at each 
pixel location 
(
1, )
(
1, )
( , , )
2
x
I x
y
I x
y
I
x y t
+
−
−
=
(3.1) 
( ,
1)
( ,
1)
( , , )
2
y
I x y
I x y
I
x y t
+
−
−
=
(3.2) 
( , , )
I x y t  denoted the continuous function defining the 
brightness values of a sequence of images captured by a 
camera. The partial derivatives of
 with respect to x, 
y, are denoted respectively by 
and
. 
( , , )
I x y t
( , , )
xI
x y t
( , , )
yI x y t
Step2.Defining matrix G 
1
1
1
1
2
2
( ( , , ))
( , , ) ( , , )
( , , ) ( , , )
( ( , , ))
(3.3)
x
x
y
W
W
x
y
y
W
W
I x y t
I x y t I
x y t
G
I x y t I
x y t
I
x y t
a
b
b
c
⎡
⎢
⎥
= ⎢
⎥
⎢
⎥
⎣
⎦
⎡
⎤
= ⎢
⎥
⎣
⎦
∑
∑
∑
∑
⎤
b
 
W1 denoted feature patch of pixels including pixel 
point(x,y). 
Step3. Find 
 
( , )
t i j
Pλ
2
( , )
(
)(
)
t i j
t
t
P
a
c
λ
λ
λ
=
−
−
−
        (3.4) 
Step4. retain pixel if
( , )
0, ( , )
t i j
t
P
a i j
λ
λ
>
>
, 
tλ is a 
threshold value. 
Step5. Perform a final non-maximum suppression step on 
1( , )
i j
λ
, yielding the actual feature locations. 
        The fixation is the area with local varying intensity, so 
good features for tracking are expected to be abundant. At 
present, there are two strategies for feature point search: 
exhaustive search and random search strategy. The first one 
is traditional feature point detection algorithm which 
searches the image for every distinction exhaustively. 
Suppose that N feature points are needed for motion 
estimation which is the purpose of searching. The random 
search strategy selects a pixel at randomly from the image. If 
the randomly selected pixel has an interest value greater than 
a predetermined threshold, it is as a feature point. This 
procedure is repeated until N feature points are detected. 
Compared with the first algorithm, the second one raised the 
speed of detection. In this paper, we use exhaustive search to 
obtain enough FPs because the size of sub-image is small.   
 
IV. FEATURE POINT TRACKING 
The next step is to locate the features detected in the first 
frame in the second frame. This procedure is called feature 
tracking. Feature tracking can be divided into two groups of 
algorithm: correlation based methods and optical flow based 
method. 
Correlation 
based 
methods 
require 
more 
computation and are appropriate when the motion of features 
in the image is large, Such as RANSAC. In this paper, we use 
Tomasi-Kanade’s feature tracking which is optical flow 
based method[15].   
      When the camera moves, the patterns of image intensities 
change in a complex way. However, images taken at late time 
instants are usually strongly related to each other, because 
they refer to the same scene taken from only slightly different 
viewpoints. 
( , , )
(
,
,
)
x
y
I x y t
I x
d
y
d
t
τ
=
+
+
+
   (4.1) 
Using Taylor-series expansion, obtaining: 
(
,
,
)
( , , )
x
y
x
x
y
y
t
I x
d
y
d t
I x y t
I d
I d
I
τ
+
+
+
≈
+
+
+
 
(4.2) 
From (4.1) and (4.2): 
0
x
x
y
y
t
I d
I d
I
+
+
=
   (4.3) 
Equation(4.3) is called optical flow constraint equation. 
If every pixel(N pixels) in the feature window has the 
same displacement, we can obtain: 
1
1
1
x
y
t
x
y
N
N
N
x
y
t
I
I
I
d
d
I
I
I
⎡
⎤
⎡
⎤
⎢
⎥
⎢
⎥
•
•
•
⎢
⎥
⎢
⎥
⎡
⎤
⎢
⎥
⎢
⎥
= −
•
•
•
⎢
⎥
⎢
⎥
⎢
⎥
⎣
⎦
•
•
•
⎢
⎥
⎢
⎥
⎢
⎥
⎢
⎥
⎣
⎦
⎣
⎦
      (4.4) 
The displacement of window can be gained by least-square. 
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡
−
=
⎥
⎦
⎤
⎢
⎣
⎡
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡
∑
∑
∑
∑
∑
∑
=
=
=
=
=
=
N
k
k
t
k
y
N
k
k
t
k
x
y
x
N
k
k
y
N
k
k
y
k
x
N
k
k
y
k
x
N
k
k
x
I
I
I
I
d
d
I
I
I
I
I
I
1
1
1
2
1
1
1
2
)
(
)
(
   (4.5) 
i.e.,
e
Gd =
              (4.6) 
Displacement is:     
     (4.7) 
e
G
d
1
−
=
 
V. COMPUTER SIMULATIONS AND RESULT 
According to the algorithm which has been expounded 
in this paper, we give the result of computer simulations in 
ICIUS 2007
Oct 24-25, 2007
Bali, Indonesia
ICIUS2007-A001
ISBN 978-979-16955-0-3
3
© 2007 ICIUS
 
 
 
this section. Fig 2(a)(b)(c) are original images of continuous 
frame. 
      
 
(a)                                                           (b) 
 
(c) 
Fig2. original images of continuous frame 
Fig2(a) is LPF image of Fig1(a), 
. Fig2(b) 
is the image by Laplacian filtering. Fig3(c) is variance map(4
×4) of Fig1(a). Fig3(d) is the extracted areas on original 
image. 
5
x
y
S
S
=
=
     
 
                         (a)                                                     (b) 
   
 
(c) 
(d) 
                                       Fig3.proceed image 
Fig4(a)(b)(c)(d)(e) is five templates(20×20) from Fig2(a) 
    
    
 
(a)                           (b)                            (c) 
Fig4.templates(sub-image) 
 
 
 
    
 
(d)                           (e) 
Fig4.templates(sub-image) 
Fig5 illustrates the result of template match. White 
Square means the position of sub-images from 3 frames 
images in the first frame. Arrow means the direction of 
template moving.   
 
                           Fig5.template match 
Fig6 illustrates the results of FPs detection. W1=3, 
threshold=1500. Fig7 illustrates the result of FPs tracking in 
the sub-image. Fig8 is the result of FPs tracking in the whole 
image. 
    
   
   
   
 
Fig6. FPs detection 
 
    
   
   
   
 
Fig7. FPs tracking of sub-image  
  
 
Fig8. FPs tracking of whole image 
ICIUS 2007
Oct 24-25, 2007
Bali, Indonesia
ICIUS2007-A001
ISBN 978-979-16955-0-3
4
© 2007 ICIUS
 
 
 
VI. CONCLUSIONS 
Computer simulation demonstrates that we can 
receive enough FPs real-timely and correctly. FPs can be 
extracted from the sub-images, as a result, FPs tracking 
algorithm cost less computation than other methods. At the 
same time, the field of FPs tracking becomes smaller, so the 
level of FPs matching becomes higher. This algorithm should 
be improved in the future. We should extend the distance 
between different feature points. The purpose is to make the 
result of motion estimation more correct  
REFERENCES 
[1]Ye Peijian, Peng Jing. Deep Space Exploration and 
Prospect in China. Engineering Science. Vol.8 No.10  
oct.2006   
[2]Larry H.Matthies and Andrew E.Johnson. Precise 
Image-Based Motion Estimation for Autonomous Small 
Body Exploration. 5th International Symposium on Artificial 
Intelligence and Automation in Space,Noordwijk,June 
1999:627-634. 
[3]Scott M.Ettinger, Michael C.Nechyba and Peter D.lfju. 
Vision-Guided Flight Stability and Control for Micro Air 
Vehicles. IEEE Int. Conf. on Intelligent Robots and Systems, 
2002:2134-2140. 
[4]Srikanth Saripalli, James F.Montgomery and Gaurav 
S.Sukhatme. Vision-Based Autonomous Landing of an 
Unmmaned Aerial Vehicle. Preceeding of IEEE International 
Conference on Robotics and Automation,2002:2799-2804. 
[5]Cucchiara R, Grana C, Piccardi M, et al. Statistic and 
Knowledge-based moving object detection in traffic scenes. 
Proceedings of IEEE Intelligent Transportation Systems, 
2000 
[6]J.K. Miller et al.“Navigation analysis For Eros rendezvous 
and orbital phases.“ Journal Astronautical Sciences, vol.43, 
No.4, pp.453-476.1995 
[7] R. Szeliski and S.B. Kang. Recovering 3-D shape and 
motion from image streams using non-linear least squares. 
Journal Visual Communication and Image Representation, 
vol. 5, No.1, pp 10-28, March 1994. 
[8] Kawaguchi, J., Hashimoto, T., Kubota, T., Sawai, S., and 
Fujii, G. (1997) Autonomous optical guidance and 
navigation strategy around a small body. AIAA Journal of 
Guidance, Control, and Dynamics, 20, 5(Sept.—Oct. 1997), 
1010—1017. 
[9] A. Johnson, A. Klump, J. Collier, and AronWolf, 
“LIDAR-Base Hazard Avoidance for Safe Landing on 
Mars," AAS/AIAA Space Flight Mechanics Meeting, Santa 
Barbara, CA,February 2001. 
[10] TOSHIHIKO MISU TATSUAKI HASHIMOTO 
KEIKEN NINOMIYA, Optical Guidance for Autonomous 
Landing of Spacecraft, IEEE Transactions on Aerospace and 
Electronic Vol. 35, No. 2 April 1999 
[11] Zheng, Q., and Chellappa, R. (1993) A computational 
approach to image registration. IEEE Transactions on Image 
Processing, 2, 3 (July 1993), 311—326. 
[12] S. Smith and J. Brady, “4sset-2: Real-time Motion 
Segmentation and Shape Tracking,” IEEE Trans. on Pattern 
Analysis and Machine Intelligence, vol. 8, no. 17, pp. 
814-820, 1995. 
[13]Jianbo Shi and Carlo Tomasi. Good Features to Track. 
IEEE 
Conf.Comput.Vision 
and 
Pattern 
Recognition, 
1994:593-600.  
[14]A.Benedetti, P.Perona. Real-time 2-D Feature Detection 
on a Reconfigurable Computer. IEEE Conf. Computer Vision 
and Pattern Recognition,1998:586-593. 
[15] Tiziano Tommasini, Andrea Fusiello. Making Good 
Feature Tracker Better. Santa Barbara, USA, IEEE 
Comput.Soc, 
Conf. 
Computer 
Vision 
and 
Pattern 
Recognition, 1998:178-182. 
 
ICIUS 2007
Oct 24-25, 2007
Bali, Indonesia
ICIUS2007-A001
ISBN 978-979-16955-0-3
5
© 2007 ICIUS