2009 Third International Conference on Multimedia and Ubiquitous Engineering

Person to Camera Distance Measurement Based on Eye-Distance

Khandaker Abir Rahman Md. Shafaeat Hossain Md. Al-Amin Bhuiyan

Department of Computer Department of Computer Science Department of Electronics &
Science & Engineering & Engineering Computer Science
University of Dhaka University of Dhaka Jahangirnagar University
Dhaka-1000, Bangladesh Dhaka-1000, Bangladesh Savar, Bangladesh
e-mail:[email protected] e-mail: [email protected]

Tao Zhang Md. Hasanuzzaman H. Ueno

Department of Automation Department of Computer Science National Institute of
Tsinghua University & Engineering Informatics (NII)
Beijing, China University of Dhaka Tokyo, Japan
e-mail: [email protected] Dhaka-1000, Bangladesh
e-mail: [email protected]

Abstract−This paper presents a novel person to camera distance of two eyes in number of pixels and camera to
distance measuring system based on eye-distance. The distance object distance (in inches), the distance of a person from the
between centers of two eyes is used for measuring the person to camera is computed.
camera distance. The variation in eye-distance (in pixels) with The proposed method in this paper is quite different
the changes in camera to person distance (in inches) is used to
from other existing image processing based person to
formulate the distance measuring system. The system starts
with computing the distance between two eyes of a person and camera distance measuring techniques which requires
then person to camera distance is measured. The proposed additional CCD cameras [8, 9], laser projectors, etc. during
distance measurement system is relatively simple and the measurements. The distance between two eyes (in
inexpensive to implement as it does not require any other pixels) of a person in an image reduces as the person moves
external distance measuring tools. Experimental results show away from the camera and vice versa. This property is used
the effectiveness of the system with an average accuracy of to measure the person to camera distance based on a certain
94.11%. eye-distance in real time.
The paper is organized as follows. In Section II, the
Keywords-Eye detection; Eye distance; Person to proposed person to camera distance measurement system
camera distance. based on eye-distance is described. Experimental results and
discussions are presented in Section III. Finally the paper is
I. INTRODUCTION concluded in Section IV.
Two widely used approaches for measuring object to
camera distance are: contact and non-contact approaches II. PERSON TO CAMERA DISTANCE MEASURING
[1]. In contact-based approach, various methods can be SYSTEM
used, such as ultrasonic distance measurement [2, 3], laser
reflection methods [4, 5]. These two methods use the theory A. Eye Distance Measurement
of reflection. If the reflection surface is not uniform, the This system forms an image pyramid of the input images
measuring system generally performs poorly or not at all. and uses a template matching approach for face and eye
On the other hand, image-based measuring systems based detection [10]. An image pyramid is a set of the original
on pattern recognition or image analysis techniques [6, 7] image at different scales. To locate the face, a mask is
generally demand huge amount of storage capacity and moved pixel-wise over each image in the pyramid, and at
high-speed processors. To overcome these problems and each position in the image the mask is passed to a function
difficulties encountered by the existing techniques, an that assesses the similarity of the image section to a face. If
image-based person to camera distance measuring system the similarity value is high enough with respect to specific
without complex calculation method is presented in this threshold, the presence of a face at that location is assumed.
paper. The system setup and configuration of the proposed From that location, the position and size of the face in the
method is very simple, consisting of only a single CCD original image is generated [10]. From the detected face eye
camera. Based on an established relationship between the is detected by forming an image pyramid of the detected

978-0-7695-3658-3/09 $25.00 © 2009 IEEE 137

DOI 10.1109/MUE.2009.34
face images and uses a template matching approach for eye noticeable that the square of eye distance versus person to
detection. The Euclidian distance between two eyes is camera distance graph is significantly identical thus it can
computed using the following (1): be generalized for persons of different physical identities.
Table I presents collected measured data of three persons on
d ep = ( E LX − E RX ) 2 + ( E LY − E RY ) 2 (1) different predefined camera to person distances (in inches).
where ( E LX , E LY ) and ( E RX , E RY ) are the center points
of left and right eyes respectively and d ep is the distance 18000

Eye Distance 2 (in pixels)

16000
14000
between two eyes in pixels. 12000
10000
8000
B. Formulation of Person to Camera Distance 6000
Measurement Equation 4000
2000
After a comprehensive study conducted over 35 people 0
of both sexes and from different height ranges, it is found 0 10 20 30 40
that a relation exists between eye distance (in pixels) and Cam era Distance (in inches)

person to camera distance (in inches). A sample square of

eye distance versus person to camera distance graph of Figure 1. Relation between eye-distance and person to camera distance
several persons is presented in Fig. 1. From the figure it is

TABLE I. SAMPLE MEASURED DATA

Square of eye distance (in pixels) Person to camera distance
(in inches)
Person 1 (Abir) Person 2 (Wahid) Person 3 (Robin)
1228 1150 1225 31
1350 1329 1370 28
1580 1450 1685 25
1900 1959 2034 22
2226 2145 2501 20
2720 2890 3000 18
4000 3986 4005 15
5800 6120 6277 12
7800 7980 8200 10
10400 10350 11211 8
14500 13500 12400 7

Equations (2) and (3) are formulated after a thorough weight. Positions of MAX ed , MIN ed , Mid G points are
study of the nature of Eye Distance2 versus Person to
shown in Fig. 2. These values are generalized considering
Camera Distance graphs of 35 people, which simulates the
graphs in real-time. the data collected of 35 people.
2 MAX ed
d ep = (2)
d c − Mid G
(1 + ) ( d c − MIN ed − 1)
Mid G
d ep
d c ' = d c ± V (2 − ) (3)
MAX ed

where d ep is the distance between two eyes, MAX ed is Figure 2. Relation between eye distance and object to camera distance

the maximum eye distance point, MIN ed is the minimum Before measuring the person to camera distance, the person
camera distance point, Mid G is the mid point of square of is trained with different predefined distances from the
camera starting from 7 inches and increased up-to 31 inches.
eye distance Vs person to camera distance graph , d c is the During the training session corresponding person to camera
primary camera to person distance (with error), d c ' is the distances (in inches) and eye distances are mapped and the
corrected camera to person distance and V is the correction
MAX ed value of that person (when the person is in the

138
highest distance from the camera) is set by the system. It is 18000 14000

Eye Distance 2 (in pixels)

Eye Distance 2 (in pixels)
16000

also found that there are generally 5 categories of MAX ed

12000
14000
10000
12000
10000 8000

values ranging from 16000 to 9500 in which the persons 8000

6000
6000
4000

tested have been categorized. Depending on the MAX ed

4000
2000
2000
0 0
0 10 20 30 40 0 10 20 30 40
value, the other parameters of (2) and (3) are set according Cam era Distance (in inches) Camera Distance (in inches)

to Table II. Fig. 3 shows the different square of eye distance

(b) (c)
versus person to camera distance graphs depending on
different MAX ed value. The values of Table II are set after 12000 8000

Eye Distance 2 (in pixels)

Eye Distance 2 (in pixels)
10000 7000
analyzing the characteristics of square of eye distance 8000
6000
5000

versus person to camera distance graphs of Fig. 3. 6000

4000
4000
3000
2000
2000
Eye Distance 2 (in pixels)

20000 1000
18000 0
0
16000
14000 0 10 20 30 0 10 20 30
12000 Cam era Distance (in inches)
Camera Distance (in inches)
10000
8000
6000
4000 (d) (e)
2000
0 Figure 3. Square of eye distance versus person to camera distance graph (a)
0 10 20 30 40

Cam era Distance (in inches) where MAX ed >16000 (b) for 13000< MAX ed <=16000, (c) for
(a) 11000< MAX ed <=13000, (d) for 9500< MAX ed <=11000 and (e)

MAX ed <=950

TABLE II. INTRINSIC PARAMETER TABLE

Value Sign
MAX ed Range MIN ed Mid G
Value Value
of V
8 23 8 +
MAX ed >16000
8 20 6 +
13000< MAX ed <=16000
8 18 4 +
11000< MAX ed <=13000
8 15 0 N/A
9500< MAX ed <=11000
7 15 4 -
MAX ed <=9500

C. Person to Camera Distance Measurement MIN ed , Mid G , V from Table II

Step 3. Set the values of
Person to camera distance measurement is accomplished
by calculating the eye distance and then mapping the according to MAX ed , where MAX ed is the maximum
corresponding person to camera distance from the eye distance point, MIN ed is the minimum camera
generalized (2) and (3) with the values of the parameters
from Table II after identifying the person along with distance point, Mid G is the mid point of Eye Distance2-
corresponding MAX ed value of that person. If the person Camera Distance graph and V is the correction weight.
is not identified then the default parameters values are Step 4. Calculate primary camera to person distance, dc
chosen. Fig. 4 shows the complete architecture of the from the (4)
proposed distance measuring system. The person to camera
d c (d c − MIN ed − 1) = ( MAX ed ×2MIDG ) 2
2
distance measurement algorithm is described bellow: (4)
d ep
Step 1. Detect the center of the two eyes and find the where d ep is the distance between two eyes.
Euclidian distance between them [10].
Step 2. If the person is identified then retrieve Step 5. Make correction to the camera to person distance by
the following equation
the MAX ed value of that person from the database.

139
d ep Accuracy of measurement results using the proposed
d c ' = d c ± V (2 − ) where d c is the primary method are shown in Table III, where real distances,
MAX ed measured distances, and accuracy (for distances from 7
camera to person distance (with error), d c ' is the corrected inches to 31 inches) of 35 persons are recorded. Fig. 6
shows the accuracy (%) of the proposed system at different
person to camera distance and V is the correction weight predefined distances. The average accuracy of 94.11% is
and return dc ' . obtained. Though other conventional measuring results
shows slight accurate where error rates range from 1 to 8%
Step 6. If the person is not identified, set the default value
[12,13], the proposed system validated its’ superiority in
as MAX ed = 11000 and goto Step 2. terms of simplicity and cost effectiveness.

TABLE III. ACCURACY OF THE DISTANCE

MEASUREMENT METHOD
Actual person to System person to Accuracy (%)
camera distance camera distance
(in inches) (in inches)
31 33.8 88.96
28 31 90.25
25 26.7 93.2
22 23 95.45
20 20.3 98.5
18 18.2 96.88
15 14.5 96.66
12 10.71 93.25
10 9.24 92.4
8 8 97.55
7 7.76 92.14

120

100
Figure 4. Person to camera distance measurement system architecture
Accuracy (%)

80

60

40

III. EXPERIMENTAL RESULTS 20

This system uses A4 Tech PK-336MB CCD camera for 0

0 10 20 30 40
image acquisition [11]. Each captured image is digitized Actual Cam era Distance (in inches)
into a 320×320 matrix with 24-bit color. The system
captures 30 image frames per second. The system considers Figure 6. Accuracy (%) of the measured distance with the actual distance
every 5th frame captured by camera for further processing.
Thus the system processes 6 image frames per second for IV. CONCLUSION
face area and eye detection [10]. Fig. 5 shows sample In this paper, a simple image-based person to camera
system output for two different scenarios. The Eye Distance distance measuring system is proposed. The proposed
and Camera Distance fields in Fig. 5 are showing square of method has significant importance because of its lower cost
eye distance (in pixels) and person to camera distance (in and simpler algorithm for real-time implementation.
inches) respectively. Because of the simplicity of the proposed approach,
hardware-intensive techniques, such as echo detection,
additional CCD cameras, laser projector [14], flash lights
etc. are no longer required for obtaining a satisfactory
person to camera distance measurement. In contrast, the
accuracy of the measured face to camera distance method
decreases as the user moves away from the camera. Though
the current system considers only frontal face view for
distance measurement, in the future, we shall consider side
face views for improving the accuracy of measurement with
a practical potential in the fields of security and robotics.

Figure 5. System output on two different scenarios

140
REFERENCES

[1] C. Chen, C. Hsu, T. Wang and C. Huang, “ Three –Dimensional

Measuremnt of a Remote Object with a Single CCD Camera”,
Proceedings of the 7th Int. Conf. on Signal Processing, Computaional
Geometry & Artificial Vision, Athens, Greece, August, 2007, Vol. 7,
No. 2, pp.141-146, 2007
[2] Alessio Carullo, Franco Ferraris, and Salvatore Graziani, “Ultrasonic
Distance Sensor Improvement Using a Two-Level Neural Network”,
IEEE Transactions on Instrumentation and Measurement, Vol. 45,
No. 2, pp.677- April 1996.
[3] A.Caarullo and M. Parvis, “An ultrasonic sensor for distance
measurement in automotive applications,” IEEE Sensors Journal,
Vol. 1, No. 2, pp.143-147, 2001.
[4] K. Osugi, K. Miyauchi, N. Furui, and H. Miyakoshi, “Development of
the scanning laser radar for ACC system,” JSAE Rev. 20, pp.579-554,
1999.
[5] H.-T. shin, “Vehicles Crashproof Laser Radar,” M.S. thesis, Opt. Sci.
Center, National Central Unive., Chung Li City, Taiwan, R.O.C.,
2000.
[6] Kanade, T., Kano, H., Kimuram S., “Development of a Video-Rate
Stereo Machine”, Proc 1995 IEEE/RSJ Int. Conf. on Intelligent
Robots and Systems, Pittsburgh, USA, August, 1995, Vol. 3, pp.95-
100, 1995.
[7] Tanaka, Y, gofuku, A ., Nagai, I., Mohamed, A.:Development of a
Compact Video-rate Range finder and its application, Proc. 3rd Int.
Conf. on Advanced Mechatronics, Okayama, Japan, August, pp.97-
102, 1998.
[8] M. A. Sid-Ahmed and M. T. Boraie, “Dual camera calibration for 3-D
machine vision metrology,” IEEE Trans. Instrum. Meas., vol. 39, no.
3, pp.512-516, Jun. 1990.
[9] C. Liguori, A. Pietrosanto, and A. Paolillo, “An on-line stereo vision
system for dimensional measurements on rubber extrusions,” in Proc.
11th IMEKO TC4 Int. Sym., Lisbon, Portugal, Sep 13/14, pp.15-19,
2001.
[10] FaceVACS-SDK, Version 1.9.9, Cognitec Systems GmbH, [Online]
URL: www.cognitec.com.
[11] www.a4tech.com, last visited on 15/12/2008
[12] Ti-Ho Wang, Ming-Chih Lu, Chen-Chien Hsu, Wei-Yen Wang,
Cheng-Pei Tsai and Cheng-chuan Chen, “A Method of Distance
Measurement by Digital Camera,” Proceedings of 2006 CACS
Automatic Control Conference, St. John's University, Tamsui,
Taiwan, Nov. 10-11, pp.1065-1069, 2006.
[13] Ming-Chih Lu, Wei-Yen Wang, and Chun-Yen Chu, “Image-Based
Distance and Area Measuring Systems,” IEEE Sensors Journal, vol.
6, no.2, pp.495-503, April 2006.
[14] T. Wang, M. Lu, Wei Wang and C. Tsai, “Distance Measurement
Using Non-metric CCD Camera”, Proceedings of the 7th Int. Conf. on
Signal Processing, Computaional Geometry & Artificial Vision,
Athens, Greece, August, 2007, Vol. 7, No. 1, pp.1-6, 2007

141