A SMART SURVEILLANCE SYSTEM FOR HUMAN FALL-DOWN

DETECTION USING DUAL HETEROGENEOUS CAMERAS

Shaou-Gang Miaou, Cheng-Yu Chien, Fu-Chiau Shih

Multimedia Computing and Telecommunications Laboratory, Department of Electronic Engineering, Chung Yuan

Christian University, Chung-Li, 32023 Taiwan, R.O.C.

Chia-Yuan Huang

Industrial Technology Research Institute, Hsinchu, 310 Taiwan, R.O.C.

Keywords: Surveillance system, omni-directional camera, PTZ camera, fall detecting.

Abstract: We propose a new surveillance system that uses both omni-directional (OD) and Pan/Tilt/Zoom (PTZ)

cameras with heterogeneous characteristics and a relatively simple image processing algorithm to achieve

the goal of real time surveillance. The system is demonstrated for detecting the occurrence of human’s

fall-down event. An OD camera has a 360º viewing angle. It is used here to replace the multiple traditional

cameras having limited viewing angles in order to reduce the system cost. A PTZ camera is also used in the

system to track the target of interest and verify the occurrence of the event. Various unique features obtained

from OD images are used for fall down detection and a multi-classifier approach is used for better

recognition performance. Experimental results show that the system is quite robust to sudden changes of

walking paths and different directions of falling. During the tracking process, a moving target is captured

and its representative coordinates is obtained based on the processing of continuous OD images. The

coordinates of the target in the OD camera space will be converted to its corresponding three dimensional

(3D) coordinates in a real-world space. This derived information is served as guidance for the automatic

control of the PTZ camera to track the moving target as closely as it can. By combining the advantages of

two heterogeneous types of cameras, our experimental results show that the proposed system can track the

moving target well without the need of a complicated method, showing the feasibility and potential of the

system.

1 INTRODUCTION

With the rapid development of technology and

medical treatment, the average life span of human

increases. As a result, the percentage of the

population group with age 65 or more becomes

higher and higher in many parts of the world. The

number of the elderly with chronic disease,

melancholia, and psychosis is increasing as well.

The elderly suffer from the fall accidents more than

young people according to the statistics. A fall

accident not only causes physical injury but also

produces emotional disturbance for the elderly.

More and more elderly need long-term care

(Chen, 2002). However, there may not be enough

manpower to take care so many elderly. One way to

alleviate the seriousness of this manpower shortage

problem is to deploy a surveillance system for the

detection of fall down events and other dangerous

situations.

A traditional camera usually has a fixed viewing

angle that limits the possible coverage of

surveillance in an environment. Multiple cameras

are usually needed to cover the entire surrounding of

the environment. A Pan/Tilt/Zoom (PTZ) camera can

extend its viewing angles by constantly panning and

tilting the camera with a motor-controlled

mechanism. However, this is not very practical for

long-term surveillance because of the considerable

waste of power and the quick wearing of

motor-driven mechanical parts. Besides, the event of

interest may not be observed if the scanning cycle of

291

Miaou S., Chien C., Shih F. and Huang C. (2008).

A SMART SURVEILLANCE SYSTEM FOR HUMAN FALL-DOWN DETECTION USING DUAL HETEROGENEOUS CAMERAS.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 291-299

DOI: 10.5220/0001933402910299

 SciTePress

the PTZ camera is much longer than the duration of

the event. Another way to extend the viewing angle

of a tradition camera is to put it on a moving carrier.

Although it can increase the cruising range of a

surveillance area, the area covered by the camera at

a time instant is still limited. And there may be some

dead spots left un-attended.

To solve the problems mentioned above,

Greiffenhagen et al. developed a new surveillance

system based on the use of an omni-directional (OD)

camera which basically consists of a traditional

camera and a convex mirror such that it has a 360º

viewing angle at any time instant (Greiffenhagen,

2001). After the OD camera was proposed, many

applications were developed based on the use of the

camera. For example, with the images taken by the

OD camera, Mituyosi et al. tried to capture and track

the features of a human face (eyes, mouths, etc)

(Mituyosi, 2003).

Although an OD camera can capture the image

from every angle at one shot, the resulting image is

distorted due to the use of the convex mirror.

Normally, the distortion level increases as it moves

from the center of the image to the border of the

image. In order to cover large surveillance areas

with reasonably good image quality, Morita used

several OD cameras and a computer network system

to connect them (Morita, 2003). Similarly, Lee also

used multiple OD cameras for outdoor surveillance

(Lee, 2002).

Again, the distortion from the convex mirror in a

PTZ camera could cause a serious problem for some

applications but it does have the advantage of the

broadest viewing angle. Thus, we use a PTZ camera

to replace the possible use of the second or more OD

cameras in our surveillance system. These two

heterogeneous types of cameras can be operated in a

complementary manner. We use an OD camera for

the preliminary tracking of a moving object in the

OD space and if the moving object presents the

features of interest, the system automatically

controls the PTZ camera to capture a sequence of the

undistorted and clear images that contain the moving

object for further visual inspection (by naked eyes).

Several advantages are associated with the system

that has such dual but heterogeneous cameras. First,

one OD camera can cover large enough surveillance

area with little dead spots. Second, the pan, tilt, and

zoom capability of the PTZ camera is fully exploited

to get clear visual information of the moving object

for better visual assessment. Third, the moving

object may be locked and stayed in locked by a PTZ

camera much more easily if the object’s location

information obtained from the OD images is

available. This dual camera system can be used in

many long-term and smart surveillance applications.

In this paper, we demonstrate its use in the detection

of human falling event in an indoor environment,

such as living room or sanatorium. Once a true

falling event is identified, emergency care may be

activated immediately to save lives. In this

application, we use an OD camera to detect the

falling event. If a suspicious event is detected by the

system, an alarm and/or message as well as PTZ

images will be sent to designated person who will

verify the event based on the PTZ images by naked

eyes. This will eliminate unnecessary false alarms

and enhance the reliability and credibility of the

system.

In the following sections, we will discuss the

issues of moving object extraction, fall-down event

recognition, and moving object tracking. The

experimental results and conclusion will be given in

the last two sections, respectively.

2 SYSTEM FRAMEWORK AND

FLOWCHART

We use the system framework shown in Figure 1.

First, we use an OD camera to capture the image of

the whole scene in a surveillance area. The resulting

OD images will be used to detect the existence of a

moving object. If a moving object is identified, its

representative coordinates on the OD image plane

will be converted to the real-world coordinates in a

three-dimensional (3D) space. The coordinate

information in 3D can be used to control a PTZ

camera such that the moving object can be tracked

and locked by the PTZ camera easily. The images

taken by the PTZ camera will be sent back to an

intended user of the surveillance system via some

network links for further visual inspection and

verification.

The system flowchart on signal processing tasks

is shown in Figure 2. Here, we use only simple and

basic image processing methods to capture a moving

object. Once the moving object is detected, we do

the coordinate conversion and fall detecting. If a

suspicious fall-down event is detected, PTZ images

will be sent to intended users to verify the event and

see if it is a true event or just a false alarm.

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

292

Figure 1: System framework.

Figure 2: System flowchart.

2.1 Object Capturing

The moving object detecting approach that we use

here is the background subtraction method which is

simple and well-known. As usual, background is

updated on a regular basis in order to overcome the

problem of light intensity change and/or chromatism.

We use Equation (1) to capture a foreground image.

),(),(),( yx

fyx

fyxD −=

(1)

where

),( yxf

denotes the pixel intensity at location

(x, y) of input image,

),( yxf

represents the pixel

intensity of background image at the same location,

and

),( yxD

denotes a difference image. Then we use

Equation (2) to eliminate the potential noise and

shadows and get a binarized image:

⎩

⎨

⎧

else0,

),( if,255

),(

TyxD

yxD

(2)

where T is an empirical threshold. When

TyxD >),( ,

the system considers that a pixel at location (x, y) is

a potential pixel of the foreground image; otherwise,

it could be just part of noises or shadows that need to

be removed. Then we use some techniques in

morphology to capture a complete foreground image.

Finally, we use the well-known connected

component labeling method to identify the moving

object. Figure 3 shows an original image and a

foreground image after performing the image

processing tasks discussed above.

After the object of interest is captured, we use

two sub-systems for further processing. One is the

fall detecting sub-system, and the other one is the

object tracking sub-system.

(a)

(b)

Figure 3: The results of foreground image capturing. (a)

An original image; (b) A binarized image after noise

removal and connected component labeling.

2.2 Fall Detecting Sub-System

2.2.1 Method

When a falling event occurs, the human body

usually has significant movement. Thus, we define

body line of human body next and use the features

associated with the body line for fall-down

recognition later. Figure 4 shows the two scanning

Input Image

Object extraction

Coordinate

conversion

Background

Removing

Color-to-Gray

Binarization

Erosion &

Dilation

Median

Filtering

Connected

Component

Labeling

Calculate d,

α, β

Calcula

e the

Position in

Real-World

Space

PTZ

Tracking

Falling

Preliminary

Warning

Get Features

Switch

Button

Fall Detecting

Verification

Rescue

Actual

Warning

Fall detecting

OD camera

Image

input

PTZ control

Image

sent

PTZ camera

A SMART SURVEILLANCE SYSTEM FOR HUMAN FALL-DOWN DETECTION USING DUAL

HETEROGENEOUS CAMERAS

293

methods to find the body line. One is horizontal

scanning, and the other is vertical scanning. In

Figure 4, the ellipse is used to represent the

foreground and the rectangle is used to represent the

smallest bounding box that encloses the foreground.

All we need is the human body, so each scan line is

dΔ pixels away from its closest sideline to avoid

the inclusion of head and feet parts. The

, a

, b

and

are the intersecting points at which the scan

lines and the boundary of the foreground object

intersect. In addition, Points

S and T are the

midpoints for

and a

, and for b

and b

respectively. Finally,

segment is the potential

body line. Figure 4 shows the results of horizontal

scans and vertical scans, resulting in two body lines,

and we choose the longer one as our final body line.

We need both horizontal and vertical scans instead

of one of them in order to avoid the situation of

wrong body line selection when the major axis of the

ellipse is parallel to the x-axis or y-axis.

(a) (b)

Figure 4: Finding a body line segment. (a) Vertical

scanning; (b) Horizontal scanning.

After the body line segment is found, we rename

the end points of the segment and call them

I and J

according to the following rule: The end point closer

to the center

M (where the OD camera is located) is

called

I and the other end point is called J, as shown

in Figure 5. Next, we define a body line vector

and a reference line vector

as follows:

OMOIMI −=

(3)

OIOJIJ −=

(4)

where

O is the origin located at the top left corner

of the image. The angle

between these two

vectors is defined as

)(

cos IJMIIJMI ⋅

−

(5)

where

IJMI ⋅

is the inner product of the vectors

and

, and

•

denotes the norm operation. Figure

6 shows a typical body line vector and a reference

line vector imposed on a real-world OD image.

Next, we define the features used for the

recognition of fall down event:

(1) The angle

defined in Equation (5).

(2) The length

L of body line vector

(3) The

x coordinate of the midpoint of body line.

(4) The

y coordinate of the midpoint of body line.

(5) The ratio

between the length and the width of

the rectangle shown in Figure 4.

One set of these five features can be obtained for

one image frame. Multiple sets of features can be

obtained for the multiple frames contained in an

observing window. Let

K be the number of frames in

the observing window. Then this sliding window is

moved such that any two adjacent windows contain

K - 1 identical image frames and one different frame.

In this study, we set

K to be 15.

Figure 5: A pictorial representation of body line and

reference line vectors.

Figure 6: A typical OD image imposed by body line and

reference line vectors and a bounding box (rectangle).

Figure 7 shows the recognition flowchart of the

proposed system. Here, a two-stage multiple

classifiers is designed to classify the feature vectors

form by the feature values discussed above. The first

(stage) classifier attempts to classify three classes:

normal (do not fall), falling down in non-radial

direction, and suspicious fall-down in radial

direction. If the last class is determined by the first

classifier, then the system uses a second (stage)

classifier to reduce the false alarm rate. The second

classifier is based on the Back-Propagation Neural

Network (BPNN).

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

294

Figure 7: System flowchart of our multiple classifiers.

2.2.2 First Classifier – Simple Decision

Criteria

We found that the angle feature defined in Equation

(5) does not change much when a target is walking

in a normal way or standing according to our

experiment. Furthermore, a large angle variation

occurs when the fall-down in non-radial direction

happens. The simple decision criteria below are used

to determine whether the non-radial directional

falling occurs.

(1) The average of angle feature values is higher

than a preset threshold (indicating the target may be

falling).

(2) The displacement of body line midpoint is

lower than another preset threshold (indicating the

target ceases moving).

If both conditions above are met, the system

determines that a non-radial directional fall-down

event occurs. However, we can not get a high

enough angle variation for radial-direction fall-down,

as shown in Figure 8. Thus, we need new

recognition features and decision criteria for this

situation. We found that when the target falls in

radial direction, the length of body line vector

changes significantly in some observing windows.

Thus, we define two other decision criteria with

different features below to deal with this situation:

(1) The variance estimate of body line lengths is

higher than a preset threshold.

(2) The displacement of body line midpoint is

lower than another preset threshold.

We also found that the body movements for

sitting and squatting also satisfy the two conditions

above. Thus, if the two conditions above are met, we

can only claim that we have a suspicious case of

radial-direction fall-down. Further differentiation

between radial-direction fall-down and sitting or

squatting is necessary. This is the objective of the

second classifier discussed next.

Figure 8: Body line vector and reference vector almost

align with each other in radial-direction fall-down.

2.2.3 Second Classifier and Multiple

Classifiers

We observed that at least 15 frames are required to

complete the process of sitting, squatting, and

radial-direction falling. Thus, the system monitors

15 continuous frames and record five parameter

values. These five parameters correspond to the five

features defined in Section 2.2.1:

(1) The angle

defined in Equation (5);

(2)

: The difference of the length L obtained

from two adjacent frames;

(3)

: The displacement in X direction for the

x coordinates obtained from two adjacent frames;

(4)

: The displacement in Y direction for the

y coordinates obtained from two adjacent frames;

(5)

: The variation between the two ratios

obtained from the bounding boxes in two adjacent

frames. For each frame in the observing window,

these five parameter values are obtained. For 15

consecutive frames, we have a total of 75 parameter

values which become the input to the BPNN for

training and testing.

2.3 Object Tracking Sub-System

Once an object is obtained, a critical point is defined

as the object point that is closest to the center of OD

image. The coordinates of this critical point in the

OD space will be converted to the real-world 3D

coordinates that could be useful for the guidance

control of PTZ cameras. Because the image taken by

an OD camera is projected from real-word space (as

First Classifie

Input features

Fall detecting

criterion met

Second

Classifier

Warning

Normal

Warning

Simple criteria

θ=0

on-radial directional falling

Suspicious radial directional falling

A SMART SURVEILLANCE SYSTEM FOR HUMAN FALL-DOWN DETECTION USING DUAL

HETEROGENEOUS CAMERAS

295

shown in Figure 9), it is possible to find the

corresponding point in real-world space for a point

in the OD space. In other words, we need to find the

point mapping from OD space to real-world space.

(a) (b)

Figure 9: Two different coordinate spaces. (a) A real-world

3D space; (b) Image taken by an OD camera in the OD

space.

We know that a point Q’ in the real-world space

as shown in Figure 9(a) can be projected to the CCD

sensor in an OD camera through the reflex of the

convex mirror, resulting in the corresponding point

Q in the OD space as shown in Fig. 9(b). This idea is

further illustrated in Figure 10. After we capture the

critical point in the OD image [marked by ◎ in

Figure 11(a)], we can calculate some useful

parameter values as follows:

QQd +=

(6)

⎟

⎠

⎞

⎜

⎝

⎛

−

tan

(7)

where

d represents the distance between the critical

point to the center of OD image and

α is the angle of

that point with respect to the

x-axis, as shown in

Figure 11.

Figure 10: Imaging principle of an OD camera (Yu, 1999).

In addition, the coordinates of 'Q can be derived

tan

sin

tan

cos

⎥

⎦

⎤

⎢

⎣

⎡

= m

(8)

where

m denotes a known constant, representing

the vertical distance (or height) from the floor to the

OD camera, and

denotes the vertical angle of the

critical point with respect to the

x-y plane in Figure

10.

(a) (b)

Figure 11: Position of a critical point in OD space. (a) an

OD image; (b) parameter calculated from the OD image

information.

We can find the corresponding position of the

critical point in real-world space by Equation (8) but

we still need one more value for the parameters

other than the

d and α calculated previously.

Specifically, we need to find the value for the

parameter tan

β. However, it can not be obtained

directly from OD images. In fact, we need to

estimate it and one way to do so is as follows. First,

design an

n-point line array pattern as shown in

Figure 12. Then, we attempt to find the position

mapping between the points in the OD image and

their corresponding points in real-world space. As

shown in Figure 12, we have

tan

(9)

= (10)

where

denotes the angle β for Point i (or P

represents the vertical distance between the plane

x-y

in real-world space and Point

i (or P

), l is the

horizontal distance between the line in the array

pattern and the center of the OD camera in

real-world space,

d is the distance from Point i to

the OD image center in OD space,

is the radius of

the largest circle containing meaningful image data

in OD images, and

k is a known ratio for Point i.

Numerical values of these parameters are shown in

Table 1 (only one third is shown). Given these data

set, we use a MATLAB curve fitting tool to calculate

the best relationship curve between tan

and k. The

result is in Equation (11):

6032.0

0442.0

226.0

1158.0

068.2929.9tan

⎟

⎠

⎞

⎜

⎝

⎛

−

⎟

⎠

⎞

⎜

⎝

⎛

−

×+×=

(11)

Therefore, given the radius

r and the distance d,

we get

k from Equation (10) and then obtain the

value for tan

from Equation (11). Finally, we get

the converted coordinates in real-world space by

using Equation (8).

After we get an estimate for the coordinates of

sensorCCD

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

296

the critical point in real-world space, we can send

proper commands to the PTZ camera via RS232 port

by properly setting panning, tilting, and zooming

such that the moving target can be centered in the

resulting PTZ images with appropriate size. For a

new critical point detected, the PTZ camera may

need a different set of commands to control it

properly. Continue in this way, the moving object

can be tracked effectively.

(a) (b)

Figure 12: Experiment trying to find the relationship

between two coordinate spaces (Hsu, 2004). (a) Data

points in real-world space; (b) data points after projection.

Table 1: Relationship between tan

and k.

i l (mm) h (mm) tanβ

d (pixel) r (pixel)

1 33.0

0.24 0.90

104

2 36.0

0.27 0.88

102

3 41.0

0.30 0.86

100

4 47.0

0.35 0.84

5 57.5

0.43 0.80

6 68.0

0.50 0.76

7 84.2

0.62 0.72

8 100.5

0.74 0.66

9 122.0

0.90 0.59

10 143.0

1.06 0.52

11 170.0

1.26 0.47

12 197.0

1.46 0.41

13 229.0

1.70 0.36

14 261.0

1.93 0.32

135

298.5

2.21 0.28

116

3 RESULTS AND DISCUSSION

3.1 Fall Detecting

Figure 13 shows a possible value variation for

Feature (5) and Feature (1) given in Section 2.2.1. It

is found that the values for ratio features may change

significantly in both walking and falling, while the

values of angel feature change significantly only in a

falling situation. Thus, the angle feature may be

more robust than the ratio feature with respect to

different walking paths.

The evaluation criteria we use for the fall

detection sub-system are accuracy, specificity, and

Kappa values (Siegel, 1988) (Lee, 2000). Table 2

shows the performance comparison between the

results of using a single (first stage) classifier and a

multiple classifiers. The multiple-classifier approach

always outperforms the single classifier method.

This demonstrates the effectiveness of the second

stage classifier in the sub-system.

Figure 13: A comparison between ratio and angle features.

Table 2: Performance comparison between single classifier

approach and multiple-classifier approach.

Accuracy Specificity Kappa

Simple

criterion

0.82 0.72 0.64

Multiple

classifier

0.87 0.88 0.73

3.2 Real-Time Tracking

We give a simple experiment to evaluate the

precision of the tracking sub-system. A target is

asked to walk along a specified path. Then the

system predicts the walking path based on the

information from OD images and the coordinate

conversion formula given in Equation (8). The result

shows that the predicted walking path is quite close

to the actual walking path, but some errors exist. The

errors may come from image processing (wrong feet

or critical point positions are obtained) and the

curving fitting formula given in Equation (11) for

tan

We check several fixed and known points to

verify the precision of the coordinate conversion

formula. Eight points shown in Figure 14 are

selected in our experiment. The results are shown in

Table 3.

As Table 3 shows, the error could be as high as

−n

Fall

happened

Walking

Ratio

Angle

Fra

A SMART SURVEILLANCE SYSTEM FOR HUMAN FALL-DOWN DETECTION USING DUAL

HETEROGENEOUS CAMERAS

297

10 pixels or more. However, the size of the image

considered here is 752*771 and the errors in

X-direction and Y-direction are about 0.8% and

0.84%, respectively. Thus, the error is relatively

small and thus the formula in Equation (11) is

acceptable. We can also see that the errors in the

inside block are smaller than those in outside block.

The reason is that the input image is distorted more

seriously in the outside block than in the inside

block. Therefore, we found that bad image

processing result (such as wrong feet detecting) is

the main source contributing the error and the curve

fitting equation has little contribution on inaccuracy.

Figure 14: Top-view diagram of experimental environment

(Size: 752*771).

Table 3: Errors for coordinate conversion.

X direction Y direction

Position

Error

(pixels)

Average

error

(pixels)

Error

(pixels)

Average

error

(pixels)

a 0.33

7.07

b 5.30

2.00

c 7.23

0.27

Inside

d 2.99

3.96

10.16

4.88

e 8.38

1.45

f 12.00

16.78

g 0.79

9.67

Outside

h 11.20

8.09

6.03

0.19

7.02

5.95

In Li’s work, a PTZ camera must be installed

directly under an OD camera and no coordinate

conversion is used (Li, 2006). Several drawbacks are

associated with the approach proposed by Li. First,

the installation constraint may prevent its system

from being practical because the installation of the

system may not be possible in some environments.

Second, the PTZ camera can cover a limited

surveillance area (with only 180viewing angle). In

other words, half of the room can not be seen. Third,

the tracking accuracy in Li’s approach may be

inferior since no coordinate conversion is used for

error compensation. In our proposed system, we do

not have the installation constraint mentioned above.

Theoretically, it can be installed anywhere in the

environment. Of course, some locations may be

more suitable than others in practical consideration.

We conducted two more experiments to further

illustrate the problems associated with the approach

proposed by Li and demonstrate the advantages with

the proposed approach in this paper. One is for the

environment where the PTZ camera is installed

directly under the OD camera (Point A in Figure 15),

and the other is for the environment where the PTZ

camera is installed at anywhere except just under the

OD camera (Point B in Figure 15). Again, we assign

a walking path for a target to follow and compare the

prediction accuracy of the path obtained by the two

approaches.

In the first experiment, since the PTZ camera is

put right under the OD camera, the accuracy results

for the approaches are quite similar and are not

shown here. In addition, a common drawback of this

installation is the limited viewing angle. For

example, suppose someone moves from Point B to

Point C along the BC segment and the PTZ camera

points to the exact south from the center in Figure 15.

In this case, half of the BC segment cannot be seen

and the tacking can not continue. In the second

experiment, we install the PTZ camera at Point B.

Note that the approach by Li uses the detected

moving target directly from OD images without any

correction. So if the target walks along the straight

line BC, the line will be distorted and becomes a

curve as shown in Figure 15. This curve rather than

straight line information is provided to the PTZ

camera for tracking. Thus, with this erroneous

information, smooth and continuous tracking

becomes harder in this case. This drawback will not

occur with the proposed approach. Since we have

corrected the distortion in some extent, the detected

path is roughly a line (as Figure 14 shows). With this

more accurate information, the smooth and

continuous tracking by the PTZ camera is easier. In

summary, the proposed approach has advantages

over the one proposed by Li, although additional

computation is required.

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

298

Figure 15: OD image showing an assigned path.

4 CONCLUSIONS

In this paper, a smart surveillance system for human

fall-down event detection based on dual

heterogeneous cameras is proposed. We attempt to

combine the advantages of both OD camera and PTZ

camera. Specifically, we can use the wide viewing

angle capability provide by the OD camera for

preliminary tracking and fall detecting, followed by

the use of a PTZ camera for fall detection

verification, face detection or other possible

applications. The precision for the coordinate

conversion between two coordinate spaces is good.

The feet detection error in image processing is the

main contribution of the inaccuracy. In our system, a

PTZ camera can be installed at the best location for

surveillance in real-world situation, which can offer

installation flexibility and convenience.

ACKNOWLEDGEMENTS

This work is supported in part by the National

Science Council of R. O. C. under Contract

NSC-95-2221-E-033-069-MY3.

REFERENCES

Chen, M. S. (2002). The Empirical Research on the

Elderly Living in Nursing Homes --- Illustrated by the

Elders the Three Public Nursing Home of Chiayi City.

Master Thesis, NanHua University, Taiwan, R.O.C.

Greiffenhagen, M., Comaniciu, D., Neimann, H., &

Ramesh, V. (2001). Design, Analysis, and Engineering

of Video Monitoring Systems: An Approach and A

Case Study, Proc. of IEEE, vol. 89, no. 10, pp.

1498-1517.

Hsu, M. Y. (2004). An Application of the Panoramic

Camera to the Positioning and the Path Planning of An

Automatic Navigated Vehicle, The 21th National Conf.

on Mechanical Eng., The Chinese Society of

Mechanical Eng. pp.1725-1730, 2004.

Lee, J. W., You, S., & Neumann, U. (2002). Tracking

with Omni-directional Vision for Outdoor AR

Systems, Proc. of the Int. Symp. on Mixed and

Augmented Reality, pp. 47-56.

Lee, T. W. (2000). Statistics, Taipei: Best-Wise Publishing

Co., Ltd.

Li, C. K. (2006). Using Omni-Directional and PTZ

Cameras to Implement Real-Time Tracking of Moving

Objects on a DSP Board, Master Thesis, Dept. of

Electronic Eng., Chung-Yuan Christian University,

Taiwan, R.O.C.

Mituyosi, T., Yagi, Y., & Yachida, M. (2003). Real-time

Human Feature Acquisition and Human Tracking by

Omni-directional Image Sensor, Proc. of IEEE Conf.

on Multisensor Fusion and Integration for Intelligent

System, pp. 258-263.

Morita, S., Yamazawa, K. & Yokoya, N. (2003).

Networked Video Surveillance Using Multiple

Omnidirectional Cameras, Proc. of IEEE Int.

Symposium on Comput. Intelligence in Robotics and

Automation, pp. 1245-1250, Kobe, Japan.

Siegel, S., & Castellan, N. (1988). Nonparametric

Statistics for the Behavioral Sciences, 2nd Ed. New

York: McGraw-Hill.

Yu, F. P. (1999). Using Panoramic Video Camera Direct

Pan-Tilt-Zoom Cameras. Master Thesis, National

Taiwan University, Taiwan, R.O.C.

A SMART SURVEILLANCE SYSTEM FOR HUMAN FALL-DOWN DETECTION USING DUAL

HETEROGENEOUS CAMERAS

299