JOURNAL OF SEMICONDUCTOR TECHNOLOGY AND SCIENCE, VOL.12, NO.2, JUNE, 2012 http://dx.doi.org/10.5573/JSTS.2012.12.2.

193

A Real-time Compact Structured-light based Range

Sensing System
Byung-Joo Hong, Chan-Oh Park, Nam-Seok Seo, and Jun-Dong Cho

Abstract—In this paper, we propose a new approach using the structured-light. In this paper, we develop a
for compact range sensor system for real-time robot compact range sensing system with a CMOS image-
applications. Instead of using off-the-shelf camera sensor and a DMD (Digital Micro-mirror Device), and
and projector, we devise a compact system with a provide an FPGA solution of the entire processing of
CMOS image-sensor and a DMD (Digital Micro- range sensor system, i.e., projecting the structured-light
mirror Device) that yields smaller dimension patterns followed by computing the sensing range.
(168x50x60mm) and lighter weight (500g). We also The generic range measurement system based on
realize one chip hard-wired processing of projection structured-light is as shown in Fig. 1. First, a projector
of structured-light and computing the range by generates the structured-light patterns that contain
exploiting correspondences between CMOS image- identifiable lines. A line on the image plane of the
sensor and DMD. This application-specific chip projector creates a light plane containing the center of
processing is implemented on an FPGA in real-time. projection and the projected line. Then, a camera
Our range acquisition system performs 30 times captures the projected pattern onto a scene and identifies
faster than the same implementation in software. We the lines generated by the projector. An image point on
also devise an efficient methodology to identify a the image plane of the camera creates a light ray (camera
proper light intensity to enhance the quality of range ray) containing the center of projection and the projected
sensor and minimize the decoding error. Our point. The intersection of the plane and the ray yields the
experimental results show that the total-error is range information.
reduced by 16% compared to the average case. Generally, single-pattern structured-light encoding
method is suitable for moving scenes, yet yielding low
Index Terms—Range sensing, structured-light, FPGA accuracy range data. Whereas multiple-pattern structured-
light encoding method can acquire accurate distance data,
I. INTRODUCTION

Recently, there have been several researches based on

structured-light in order to acquire accurate range
information [1], relying on a set of bulky projector and
camera [2, 3]. One important application is smart robots
that must recognize ambient surroundings autonomously

Manuscript received Sep. 1, 2011; revised Nov. 8, 2011.

Sungkyunkwan University - Department of Electronic and Electrical
Engineering, Suwon 440-746, Korea, Republic of
E-mail : [email protected] Fig. 1. Structured-light based range sensing [4].
194 BYUNG-JOO HONG et al : A REAL-TIME COMPACT STRUCTURED-LIGHT BASED RANGE SENSING SYSTEM

yet being vulnerable to the moving object [5]. Motivated Structured-light patterns are projected by DMD, and
by these facts, our approach is to apply high-speed captured by CMOS image-sensor. 2) The captured
CMOS image-sensor (500fps) and hard-wired range images are stored in external memory and decoded. 3)
computation to overcome the vulnerability problem of The decoded images are triangulated with the ray-plane
the multiple pattern approach. Furthermore, we attain intersection. 4) The final range data is transmitted to the
real time processing of dynamic scenes without losing external main control component through USB.
accuracy.
To enhance the quality of range sensor and minimize 1. Structured Light Decoding
the decoding error, we also present a methodology to
identify a proper light intensity. There are related efforts Structured-light decoding is a process of finding
which aim to enhance the quality of range sensor with correspondences between the columns of the DMD
adaptation to the scene. In [6] the authors introduced an projected pattern and the pixels of the captured image.
adaptive algorithm with adjusting the coding patterns We use Gray code for structured-light encoding. Gray
depending on the environment. However, there is no code is well suited for binary position encoding due to its
flexibility with structured-light coding methods due to one bit difference property. More robust encoding is to
the specially designed coding patterns. In [7] the authors add an inverted Gray pattern between two Gray patterns
presented an adaptive algorithm with using the partially [8], as shown in Fig. 3.
reconstructed scene. However, this approach is not The structured-light decoding process is shown in Fig.
suitable for real-time application due to its high 4. We need one frame of non-illuminated image (i.e.,
computational complexity although it provides high reference image) and 16 frames of patterned images to
accuracy results. In this paper, we present a methodology decode the structured-light patterns. Each layer contains
to enhance the quality of range sensor in real time with two frames; original Gray coded pattern and inverted
adjusting the light intensity according to the suggested Gray coded pattern. The decoding process is as follows.
structured-light decoding error model. First, the pixel values of the patterned images are
The rest of the paper is organized as follows. In subtracted by the pixel value of the non-illuminated
Section II, we shall exhibit the overall system architecture, image (referred to ‘Ref-frame-value’). The ‘Frame value’
and describe the process of structured-light decoding and represents the pixel value at each frame. The ‘Frame
ray-plane intersection. In Section III, we present on the value*’ represents the value subtracted by the ‘Ref-
error-minimized light intensity search. In Section IV, we frame-value’. Next, we compare the ‘Frame value*’
will show our implementation and experimental results. between two ‘Frame values’ in each layer and choose the
Finally, Section V will conclude the paper. greater value (Layer value).
The ‘Max frame#’ bit is set to ‘0’ if the left frame
II. SYSTEM ARCHITECTURE value (i.e., the pixel from the gray pattern) is greater than
the right frame value (i.e., the pixel from the inverted
As shown in Fig. 2, our architecture performs: 1) gray pattern); otherwise set to ‘1’. All of the ‘Max
frame#’ bits for all layers correspond to the index of the
LUT (look-up table). The LUT yields the corresponding
column of the projection image at the specific pixel of
captured image. To remove the ambiguous decoding

Fig. 2. The overall architecture of our system. Fig. 3. The structure of sequence frames.
JOURNAL OF SEMICONDUCTOR TECHNOLOGY AND SCIENCE, VOL.12, NO.2, JUNE, 2012 195

Dec_result
Cur_x Cur_y (Col_number)

Center of DMD Planes

projection

FSM we DMD cali

Generator
Generator
Addr

Ray plane intersection

parameter

Planes

Rays
Data_in DMD distortion
Data_out parameter

Dot product Dot product

Cam rays
Cam cali

Generator
we
Addr parameter

Rays
Divider
Cam distortion
Data_in parameter
Data_out

Output reg

Range Xout Yout Zout

Fig. 6. Ray-plane intersection hardware.

intersection to yield the range. The DMD’s center of

projection and each column of DMD projected image
Fig. 4. Structured-light decoding process.
form a plane. The image point and the center of the
projection of image-sensor form a ray. We calculate
rays and planes with the pre-calibrated DMD and
image-sensor calibration parameters. The intersection
Ref frame value

point between the ray and the plane is the location of a

point in the 3D world.
The range via ray-plane intersection can be represented
as:

·
range , (1)
·

Fig. 5. Structured-light decoding hardware.

where nt is a normal vector of a plane, p is a center of
result, we reject the decoding result if ‘Layer value’ is projection of DMD, and r is a ray vector.
less than the structured-light decoding threshold value Fig. 6 shows the hardware architecture of the above-
(denoted as thsl). mentioned operation.
The above-mentioned decoding is structured as shown
in Fig. 5. The 16 images are subtracted by the reference III. ERROR MINIMIZED LIGHT INTENSITY
image. The pixel from the gray pattern and the pixel from SEARCH
the inverted gray pattern are compared. The compared
result is an index of the LUT. The LUT also provide 1. Defining Structured-light Decoding Error
flexibility in using other structured-light encoding
methods. Motivated by the fact that the quality of range sensor
requires a proper light intensity, we specify the
2. Ray-Plane Intersection structured-light decoding error which can affect the
distance results. We found that the error can be mainly
In the calibration process of the system, we determine categorized into two cases. One is the case where the
the calibration parameters such as the focal distance, decoding result is ‘0’ (referred to as “zero-error”), and
translation, and orientation of DMD and CMOS image- the other is the case where the initial order of encoded
sensor. The calibration procedure is similar to the result is altered after decoding (referred to as “order-
classical stereo-vision system with modeling a projector error”). The ‘zero-error’ represents the number of range
as an inverse-camera [9]. We triangulate with ray-plane data. And the ‘order-error’ implies the number of
196 BYUNG-JOO HONG et al : A REAL-TIME COMPACT STRUCTURED-LIGHT BASED RANGE SENSING SYSTEM

uncertain range data. The zero-error can be classified into 1 for I i, j, γ 0 and
three cases such as “zero-over-error,” “zero-under-error,” I i, j, γ I i, j, γ 255 th (5)
and “zero-occlusion-error”. 0 else
We explain on how to compute the above mentioned
Then, we define zero-occlusion-error. The zero-error
four different errors. The zero-error occurs when any
also occurs when the pixel of captured image is in
‘Layer value’ (as shown in Fig. 4) is smaller than ‘thsl’ at
shadow area or out of sight from the projection area. We
a pixel. That implies that the pixel value difference
define this case as zero-occlusion-error. We identify
between light illuminated case and light non-illuminated
zero-occlusion-error pixel when the pixel value
case is smaller than ‘thsl’. This case occurs generally
difference between all-illuminated image (denoted as
when we take the image of a dark object (e.g., black
I i, j, γ and reference image I i, j, γ is less than the
object) due to its low reflectivity.
threshold value (denoted as thocclusion ).
For this case, we define the zero-under-error. This
error can be changed depending on the intensity of the
light γ in Eq. (2). The zero-error for pixel (i,j) is E γ ∑H ∑W I i, j, γ (6)
H W
defined when the pixel value of structured-light decoding
image (denoted as , , ) is zero. The zero-under- I i, j, γ
error pixel is defined when the pixel value of reference 1 for I i, j, γ 0 and
image (denoted as , , ) is less than (255-thsl) in I i, j, γ I i, j, γ th (7)
0 else
Eq. (3). In Eq. (3), the pixel value of zero-under-error
image (denoted as , , becomes ‘1’. Then, we
Last, we define order-error. Among the valid
get zero-under-error E γ with counting all the 1’s
decoding pixels (i.e., non-zero-error pixels), the decoded
in , , and with normalizing by the size of
pixel value would contain error. The decoding result
image (H W . should be increased along the right direction since the
decoding result corresponds to the column number of
E γ ∑H ∑W I i, j, γ (2) projector’s image plane. We define the out-of-ordered
H W
decoding result error as order-error. We distinguish
1 for I i, j, γ 0 and order-error for a pixel (i,j) by checking if the pixel value
I i, j, γ I i, j, γ 255 th (3) of the decoding image , , is less than the pixel
0 else value of pre-positioned pixel 1, , . If
, , = 1, then this case is out-of-order. Then, we
Next, we define the zero-over-error. The zero-error
accumulate the number of order-error pixel while
also increases with increasing light intensity when we
I i, j, γ = 1, as shown in Eq. (8).
take image of a bright object (e.g., white object). The
pixel value in this case is almost 255 without providing
E γ ∑H ∑W I i, j, γ (8)
the projector’s light. This causes smaller difference value E γ H W

than thsl, due to the upper limited pixel value (255 for
8bit pixel resolution). We define this case as zero-over- I i, j, γ
error. In other words, we identify the zero-over-error 1 for I i, j, γ I i 1, j, γ
(9)
pixel when the pixel value of reference image 0 else
, , is larger than (255-thsl). Similar to the zero-
We normalize to get the ratio of the order-error pixels
under-error, we define zero-over-error as shown in Eqs.
to the non-zero-error pixels. Here, zero-error-total
(4, 5).
(denoted as ) is defined as in Eq. (10).

E γ ∑H ∑W I i, j, γ (4)
H W E γ E γ E γ E γ (10)
JOURNAL OF SEMICONDUCTOR TECHNOLOGY AND SCIENCE, VOL.12, NO.2, JUNE, 2012 197

Then, we finally get the total error as below.

E γ E γ ω ·E γ , (11)

where, ω is the order-error-weight.

Finally, we find the optimum light intensity value
(denoted as among γ values so as to minimize the
total-error.

Oγ arg minγ E γ (12)

Fig. 8. Order-error function (depending on the light intensity).
As shown in Fig. 7, we tested the error function
depending on the intensity of the light. Our experiments decreases as the accuracy increases. Therefore, there is a
reveal that the maximum light intensity (100%) does not tradeoff between the accuracy and the number of 3D
always yield a good result. The order-error is much points. In this paper, we determine ω to 3 after
smaller than zero-error due to the reasonable structured- experimentally analyzing the increasing rate of accuracy
light encoding method. Even though the order-error is and decreasing rate of the number of 3D points,
resided in a small value region as shown in Fig. 8, the depending on ω . When ω is more than 3, we shall
error seems to be getting higher in some special case for lose many 3D points with getting a little higher accuracy.
the weak scene (i.e., transparent object and/or a mirror)
for structured-light based range sensing system. To 2. Light Intensity Search to minimize the decoding
emphasize order-error, we assign a higher weight (ω ) error
on the as shown in Eq. (11).
We determine ω according to the relation between Light intensity search is done in an iterative manner as
zero-error and order-error functions. When ω = 1, the shown in Fig. 9. We search from 100% to 20% of light
minimal E γ light intensity is almost same as the intensity with step size of 5%. The search process is
minimal E γ light intensity, because the zero-error performed to obtain the minimum error value (i.e.,
is dominant to the total-error. However, as ω optimal light intensity) via capturing and examining a
increases, the minimal E γ light intensity varies to number of structured-light decoding images by
minimize the order-error. In this case, the minimal employing different light intensity. For simplicity, we
E γ light intensity goes far from the minimal use the exhaustive search algorithm in this paper, and are
E γ light intensity and the number of 3D points in the process of developing the sophisticated algorithm
to find the optimum light intensity with the minimum
number of scans.

Fig. 7. Zero-error function (depending on the light intensity). Fig. 9. Light Intensity Search.
198 BYUNG-JOO HONG et al : A REAL-TIME COMPACT STRUCTURED-LIGHT BASED RANGE SENSING SYSTEM

3. Hardware Implementation of Error Estimation and IV. IMPLEMENTATION AND EXPERIMENTAL

Light Intensity Search RESULTS
We implemented the error estimation and search 1. Hardware Environment and Implementation
algorithm in hardware. The block diagram is shown in
Fig. 10. First, we estimate the error defined in the Our proposed system is designed in a circuit board
previous subsection with the structured-light decoding (except for DMD interface). Overall architecture is
result. To estimate error, SL_ERROR_ESTIMATOR designed in Verilog-HDL and implemented in an FPGA
module reads the results of decoding (referred to as device (Virtex-5) with operating at 50 MHz. The
“sl_decoding_result”) stored in the external memory. Next, conventional system is configured with an off-the-shelf
OPTIMAL_LIGHT_INTENSITY_SEARCHER module camera, an off-the-shelf projector, and general PC. After
searches for the optimal light-intensity with the input implementing with hardware for entire system, we could
signals (i.e., Etotal_valid and Etotal) generated by SL_ remove several limitations from the conventional
ERROR_ESTIMATOR module. Our system mounts systems to obtain real time operation. Moreover,
DMD module that uses LED as a light source. The conventional systems are limited to the low-flame-rate
selected light intensity value is transmitted to the PWM camera (generally 30~60 fps) whose frame rate of range
generator module, and the light intensity is regulated by image is very low (max. 3 fps, required 17frames to yield
the PWM . one range image). Whereas, our solution uses high-speed
Fig. 11 exhibits SL_ERROR_ESTIMATOR module to CMOS image-sensor (500 fps) to increase the frame rate
acquire the result of Eq. (11). of range image. Furthermore, conventional systems are
limited to the low-frame-rate projector (generally 60 fps)
whose range image frame rate is very low (max. 3 fps).
To overcome the frame rate limitation of projectors, we
extract the DMD from projector. Another challenge for
the conventional system is synchronization difficulties
between camera and projector to capture projected
pattern-illuminated image. Due to the hardware and
software component synchronization problem, conventional
system requires additional hardware circuit, yielding low
frame rate. However, our hardware platform approach
could resolve the synchronization problem without area
overhead and frame rate loss. Table 1 shows the
comparison of processing time between our system and
Fig. 10. Light intensity search subsystem. conventional system.
As shown in Table 1, our approach is faster by 30

Table 1. Comparison of Processing Time & Frame Rate

Process Conventional systems Proposed system
[1, 7]
Image acquiring
60 fps 500 fps
frame rate
Structured-light
60 fps 500 fps
projection frame rate
Decoding 813 ms(1.23 fps) 56 ms(17.85 fps)
Ray-Plane 928 ms(1.07 fps) 58 ms(17.24 fps)
Intersection

Fig. 11. Error estimator hardware. Total 1741 ms(0.57 fps) 58 ms(17.24 fps)
JOURNAL OF SEMICONDUCTOR TECHNOLOGY AND SCIENCE, VOL.12, NO.2, JUNE, 2012 199

Fig. 12. System board.

times compared to the conventional system configuration. Fig. 14. Structured-light decoding image.
Fig. 12 shows the entire system board that is smaller than
the system with using off-the-shelf projector and camera. intensity represents the range (in millimeter, the larger
value is the nearer position from the acquisition system).
2. Experiments on structured light range search The distance values of the 3D points of the diagonally
positioned box was smoothly increased or decreased with
Fig. 13 shows a non-illuminated general image the front edge of the box as a center. Also note that the
captured by CMOS image-sensor. To evaluate our distances of the paper cup and the box at the bottom are
system, we stacked several objects (a book, two boxes similar.
and a paper cup from bottom to top) at a distance of Fig. 16 shows the reconstruction of the objects with
about 50cm from the acquisition system. Notice that the the 3D points without any further post processing such as
box right below the paper cup is placed diagonally so as smoothing.
to experimentally verify the smooth increase or decrease For the purpose of accuracy analysis of our system, we
in distance. placed a white plane board at a distance of about 60 cm
Fig. 14 shows the structured-light decoded image. from the acquisition system. First, we obtain 3D points of
Each pixel’s intensity value represents the column the white plane through our acquisition system. Next, we
number of the DMD projected image. The black-colored reconstructed a flat 3D plane derived from the acquired
pixel is the invalid value due to either ambiguous 3D points using a plane-fitting such as PCA (Principle
correspondences or being out of sight from the projection component analysis). The difference in distance between
area. The intensity values of the image pixel is smoothly the points in flat plane and the 3D points is used as a
increased in the right direction, implying that the column metric for the range accuracy (expressed in terms of
numbers of the DMD projected image are increased from average error and standard deviation). Table 2 shows the
left to right.
The range image is shown in Fig. 15. Each pixel’s

Fig. 13. Non-illuminated image. Fig. 15. Range image.

200 BYUNG-JOO HONG et al : A REAL-TIME COMPACT STRUCTURED-LIGHT BASED RANGE SENSING SYSTEM

Fig. 16. 3D reconstruction image.

Table 2. System accuracy result

Fig. 17. Test scene & Result images.
Average error(mm) Standard deviation(mm)
1.28 0.62
Table 4. Comparison of range quality among different light
intensity assignments
verified result (an average of 10 experiments). This result Light Intensity Error Range quality
is reasonable for robot applications. Average error(mm) in
γ(%) E γ 3D points
ROI
Aγ max(100) 0.89 21504 2.37
3. Experiments on light intensity search mid(60) 0.81 49152 1.97
min(20) 0.92 18432 1.29
We tested our algorithm using the scene shown in Fig. Oγ 45 0.72 93501 1.33

13. Table 3 compares the total-errors derived from * γ (light intensity), A (Arbitrary light intensity), O (optimal light
intensity), E γ (total error)
optimal light intensity with the ones from the arbitrary
light intensity such as maximum(100%), middle(60%),
and minimum(20%) light intensity. The total-error was area and the accuracy in ROI area at the different light
reduced by 16% compared to the average case . intensity assignment, respectively. The optimal light
The zero-error represents the number of 3D points and intensity yields five times more 3D points compared to
the order-error implies the unreliable 3D points. the worst case(γ=20) with small accuracy reduction (0.04
Therefore, we adaptively apply the optimal light intensity mm).
to maximize the number of 3D points and accuracy. To We also verified the influence of ambient-light
show the quality of the measured distance, we set the conditions on the optimal-light-intensity of the system.
ROI (Region of Interest) which is flat area in the scene as Fig. 18 shows the relationship between E γ and
shown in Fig. 17. Then, we fit a 3D plane in the ROI and structured- light intensity with varying ambient-light
compute the accuracy at the arbitrary light intensity and
optimal light intensity.
Table 4 shows the number of 3D points in entire image 1
Level4
0.8 Level3
Level2
0.6 Level1
Table 3. Comparison of decoding errors among different light
E__total

intensity assignments
0.4
Arbitrary γ O
γ(%) E γ γ(%) E γ 0.2

max(100) 0.92
50 0.78 0
mid(60) 0.87 100 80 60 40 20 0
Light Intensity(%)
min(20) 0.99
* γ (light intensity), O (optimal light intensity), E γ (total error) Fig. 18. Influence of ambient light.
JOURNAL OF SEMICONDUCTOR TECHNOLOGY AND SCIENCE, VOL.12, NO.2, JUNE, 2012 201

conditions. In Fig. 18, the higher level implies higher Imaging, Vol. 11(5-6), pp.358-369, 2005.
ambient-light conditions. With our error-model based on [4] http://en.wikipedia.org/wiki/Structured-light_3D_
E γ , our system attempts to minimize E γ scanner.
regardless of changed ambient-light condition. [5] J. Salvi, J. Pagés, and J. Batlle., “Pattern Codification
Strategies in Structured Light Systems,” Pattern
V. CONCLUSIONS Recognition, Vol.37, No.4, pp.827-849, 2004.
[6] Thomas P. Koninckx, Luc Van Gool, “Real-Time
In this paper, we proposed a new technique for the Range Acquisition by Adaptive Structured Light,”
structured-light based range sensing system that is IEEE Transactions on Pattern Analysis and
capable of processing of structured-light projection and Machine Intelligence, Vol.28, No.3, pp.432-445,
range computation. Overall processing is realized with 2006
fully hard-wired manner using an FPGA that results in [7] Yi Xu, Daniel G. Aliaga, “An Adaptive
real-time operation. The frame-rate of 17 fps (640x480) Correspondence Algorithm for Modeling Scenes
with Strong Interreflections,” IEEE Transactions
is reliable in the dynamic scene and feasible to recognize
on Visualization and Computer Graphics, Vol.15,
moving objects. Furthermore, to enhance the quality of
No.3, pp.465-480, 2009.
range sensor and minimize the decoding error, using a
[8] D. Scharstein and R. Szeliski, “High-accuracy
proper light intensity leads to desirable results.
stereo depth maps using structured light,” CVPR03,
Finally, we are posing some future works. First, the
pp.195-202, 2003.
current system has room for hardware optimization with
[9] S. Zhang, P. Huang, “Novel method for structured
pipelining optimization [10] and rescheduling the
light system calibration,” Optical Engineering,
processes. Second, we might apply various structured-
Vol.45, No.28, pp.083601, 2006.
light encoding techniques to secure more robustness with
[10] Chanho Lee, “Smart Bus Arbiter for QoS control in
the moving scenes or more accuracy the range data
H.264 decoders,” Journal of Semiconductor
depending on the different conditions.
Technology and Science, Vol.11, No.1, pp.33-39,
2011.
ACKNOWLEDGMENTS

This research is supported by the Basic Science

Research Program through the National Research
Foundation of Korea (NRF) funded by the Ministry of
Education, Science and Technology(S-2011-0251-000). Byung-Joo Hong received the B.S.
degree in the Department of Electronic
REFERENCES Engineering from Dongguk University,
Seoul, Korea, in 2005 and the M.S.
[1] J. Salvi, S. Fernandez, T. Pribanic, and X. Llado, degree in the Department of Electronic
“A state of the art in structured light patterns for and Electrical Engineering from
surface profilometry,” Pattern Recognition, Vol.43, Sungkyunkwan University, Suwon,
No.8, pp.2666-2680, 2010. Korea, in 2007. He is currently pursuing the Ph.D. degree
[2] S. Zhang, “Recent progresses on real-time 3D in the same university. His research interests include 3D
shape measurement using digital fringe projection image processing and system-on-chip design.
techniques,” Optics and Lasers in Engineering,
Vol.48, No.2, pp.149-158, 2010.
[3] F. Tsalakanidou, F. Forster, S. Malassiotis, and M.
G. Strintzis, “ Real-time acquisition of depth and
color images using structured light and its
application to 3D face recognition,” Real-Time
202 BYUNG-JOO HONG et al : A REAL-TIME COMPACT STRUCTURED-LIGHT BASED RANGE SENSING SYSTEM

Chan-Oh Park was born in Korea, Jun-Dong Cho received the B.S.
in 1986. He received the B.S. degree degree from the Department of
in department of electronic and Electronic Engineering,
information engineering from Seoul Sungkyunkwan University, Suwon,
National University of Technology Korea, in 1980, the M.S. degree
in 2010. He is currently pursuing the from the Department of Computer
M.S. degree in Sungkyunkwan Science, Polytechnic University
University. His current research interests include stereo Brooklyn, New York, in 1989, and the Ph.D. degree from
vision, computer architecture and system on chip. the Department of Computer Science, Northwestern
University, Evanston, in 1993. He was a CAD engineer
and Team Leader in Samsung Electronics Company,
from 1983 to 1987, and a Senior Technical Staff in
Nam-Seok Seo was born in Korea, Samsung Electronics Company, from 1993 to 1995. In
in 1982. He received the B.S. degree 1995, he joined the Department of Electrical and
in department of System Control Computer Engineering, Sungkyunkwan University
engineering from Hoseo University (SKKU), Suwon, Korea, where he is currently Professor.
in 2009. He is currently pursuing the His research interests include Low Power Design, 3-D
M.S. degree in Sungkyunkwan Image Processor, Embedded System Integration Design,
University. His current research and Multiprocessor on Chip for Software Defined Radio
interests include stereo vision, computer architecture and and Multimedia Applications. Prof. Cho is an IEEE
system on chip. Senior Member.