Mobile Robot Navigation by a Distributed Vision System 1

Mobile Robot Navigation by a Distributed

Vision System

Takushi Sogo

Department of Social Informatics, Kyoto University

Sakyo-ku, Kyoto 606{8501 JAPAN

Hiroshi Ishiguro

Department of Computer and Communication Sciences,

Wakayama University

930 Sakaedani, Wakayama 640{8510 JAPAN

Toru Ishida

Department of Social Informatics, Kyoto University

Sakyo-ku, Kyoto 606{8501 JAPAN

Received 31 October 2000

Abstract This paper proposes a Distributed Vision System (DVS),

which is an intelligent infrastructure for mobile robots consisting of many

vision agents (VAs) embedded in an environment. The system monitors

the environment from various viewing points with the VAs, maintains the

dynamic environment models, and provides various information to robots.

Based on this concept, we have developed a prototype of the DVS which

consists of sixteen vision agents and simultaneously navigates two robots

in a model town.

Keywords Distributed Vision System, Vision Agents, Mobile Robot

Navigation, Perceptual Information Infrastructure.

x1 Introduction

2 Toru Ishida

Realizing autonomous robots that behave in the real world based on

visual information has been one of the goals in robotics and arti�cial intelli-

gence. For limited environments such as o�ces and factories, several types of

autonomous robots which utilize vision sensors have been developed. However,

it is still hard to realize autonomous robots behaving in dynamically changing

real worlds such as an outdoor environment.

As discussed in Active Vision 1), the main reason lies in attention control

to select viewing points according to various events relating to the robot. In order

to simultaneously execute various vision tasks such as detecting free regions and

obstacles in a complex environment, the robot needs attention control, with

which the robot can change its gazing direction and collect information according

to various events. If the robot has a single vision, it needs to change several

vision tasks in a time slicing manner in order to simultaneously execute them

(Temporal Attention Control). Furthermore, in order to execute various vision

tasks the robot needs to select the best viewing point according to the vision

tasks (Spatial Attention Control).

However, it is di�cult with current technologies to realize the attention

control, since vision systems of previous mobile robots are �xed on the mobile

platforms and it is di�cult to acquire visual information from proper viewing

points. The robots generally have a single vision sensor and a single body,

therefore the robot needs to make complex plans to execute the vision tasks

with the single vision sensor.

Another reason for the di�culty in building autonomous robots behaving

in the real world is that a single robot has to acquire a consistent model of a

wide environment with vision sensors �xed on its body. If the environment

dynamically changes, it will be very hard to maintain the model.

Our idea to solve these problems is to use many vision agents (VAs) em-

bedded in the environment as shown in Fig. 1. Each VA independently observes

events in the local environment and communicates with other VAs through the

computer network. Since the VAs do not have any constraints in the mechanism

like autonomous robots, we can install a su�cient number of VAs according to

the tasks to be performed, and the robots can acquire necessary visual informa-

tion from various viewing points.

As a new concept to generalize the idea, we have proposed Distributed

Vision 2), where multiple vision agents embedded in an environment recognize

dynamic events by communicating with each other. In the distributed vision,

Mobile Robot Navigation by a Distributed Vision System 3

Vision agents

Computer network

Mobile robots

Vision agents

Computer network

Fig. 1 Distributed Vision System

the attention control is dealt with as dynamic organization of communication

between the vision agents (the detail of the concept of the distributed vision is

discussed in 2)). Based on the concept, we have developed a Distributed Vision

System (DVS) 2), which solves the above problems and realizes robust navigation

of mobile robots in a complex and dynamic environment such as outdoor envi-

ronments. The DVS consists of VAs, solves the attention control by selecting

proper VAs, and realizes navigation of mobile robots based on visual information

in a complex environment.

This paper describes the detail implementation of the prototype of the

DVS, as well as the concept of the distributed vision. In developing the DVS,

the following are important:

1. navigation method of multiple robots using multiple VAs,

2. communication among VAs,

3. construction and management of environment models in the VA net-

work.

This paper mainly discusses the �rst point. In the following sections, the ar-

chitecture of the prototype system and the navigation method of robots are

described. Finally, experimental results of navigation are shown.

x2 Distributed Vision System

4 Toru Ishida

2.1 Design policies of the DVS

The vision agents (VAs) in the distributed vision system (DVS) are

designed based on the following idea:

Tasks of robots are closely related to local environments.

For example, when a mobile robot executes a task of approaching a target, the

task is closely related to a local area where the target locates. This idea allows

to give VAs speci�c knowledge for recognizing the local environment, therefore

each VA has a simple but robust information processing capability.

More concretely, the VAs can easily detect dynamic events since they are

�xed in the environment. A vision-guided mobile robot, the camera of which is

�xed on the body, has to move for exploring the environment, therefore there ex-

ists a di�cult problem to recognize the environment through the moving camera.

On the other hand, the VA in the DVS can easily analyze the image data and

detect moving objects by constructing a background image for a �xed viewing

point.

All of the VAs basically have the following common visual functions:

� detecting moving objects by constructing the background image and com-

paring the current image with it,

� tracking detected objects by a template matching method,

� identifying mobile robots based on given models,

� �nding relations between moving objects and static objects in the images.

For robustness and exibility, the DVS autonomously and locally cal-

ibrates the camera parameters with local coordinate systems according to de-

mand, rather than keep the precise parameters (the detail is discussed in Sec-

tion 3.4). In addition, the DVS does not use a geometrical map in robot nav-

igation. It memorizes robot tasks directly shown by a human operator, then

navigates robots based on the memorized tasks.

2.2 The architecture

Fig. 2 shows the architecture of the DVS for robot navigation. The

system consists of multiple VAs, robots, and a computer network connecting

them.

Image processor detects moving robots and tracks them by referring to

Knowledge database which stores visual features of robots. Estimator receives the

results and estimates camera parameters for establishing representation frames

Mobile Robot Navigation by a Distributed Vision System 5

Motion planner

Knowledge database

Estimator ofcamera parameters

Organizationmanager

Camera

Image processor

Memory of tasks

Communicator

Controller

Task manager

CommunicatorCommand integratorActuators

Robot

Memory oforganizations

Planner

Vision agent

Com

pute

r ne

twor

k

Fig. 2 The architecture of the DVS

for sharing robot motion plans with other VAs. Task manager memorizes the

trajectories of robots as tasks shown by a human operator, and selects proper

tasks in the memory in order to navigate robots. Planner plans robot actions

based on the memorized tasks and the estimated camera parameter. Organiza-

tion manager communicates with other VAs through Communicator and selects

proper plans. The selected plans are memorized in Memory of organizations for

planning robot tasks more properly. Controller controls the modules, according

to requests from the robots and the system state such as task teaching phase

and navigation phase.

The robot receives the plans through Communicator. Command In-

tegrator selects and integrates the plans, from which actuator commands are

generated.

x3 Development of a prototype systemWe have developed a prototype system for robot navigation based on

the architecture described in the previous section. In this section, the details of

the system are described.

3.1 Fundamental functions for robot navigation

In the DVS, robot navigation is achieved with two steps | task teach-

ing phase and navigation phase. In the task teaching phase, each VA observes

and memorizes the path of a robot which is controlled by a human operator or

6 Toru Ishida

!

Background images Detected free regions

Fig. 3 Detecting free regions

autonomously moves in the environment. In the navigation phase, VAs commu-

nicate with each other to select VAs which provide proper visual information for

navigation, then navigate robots based on the paths memorized in each sensor

image.

The system needs the following functions for robot navigation:

1. Navigate robots on free regions where the robots can move.

2. Avoid a collision with other robots and obstacles.

3. Navigate robots to their destinations.

The system realizes these functions by using visual information acquired by VAs.

Navigation on free regions is realized as follows. The knowledge of the

free regions are obtained if VAs observe moving objects in the environment for

a long time, assuming that the regions where the objects move around are free

regions. In the developed system, VAs observe a robot controlled by a human

operator in the task teaching phase (described in the next subsection), and

recognize free regions as shown in Fig. 3. Collision avoidance is realized by

generating a navigation plan so as to avoid a collision. In the developed system,

if VAs �nd that a robot is on a possible collision course with other robots or

obstacles, they temporarily modify the destination of the navigation plan in

order to avoid the collision. Navigation to a destination is realized by teaching a

knowledge of paths for navigating robots to their destinations. In the following

Mobile Robot Navigation by a Distributed Vision System 7

1

23

4

A B

C

Fig. 4 Navigation paths taught by a human operator

subsection, the task teaching phase is described.

3.2 Task teaching

The system switches into the task teaching phase with instruction of a

human operator. In this phase, VAs memorize tasks shown by a human operator.

The task consists of several subtasks, which are navigation between intersections,

in this prototype. By linking subtasks, the system navigates robots to their

destination.

First, VAs detect robots in each view. Since the VAs are �xed in an en-

vironment, they can easily detect objects by background subtraction. Then, the

VAs distinguish robots from other objects, and identify robots by their colors.

Each VA tracks the robot controlled by a human operator, and memorizes its

apparent trajectory in the view as a navigation task. When the robot passes over

a speci�c place (e.g., in front of a building), the operator noti�es the meaning

of the place to the system. The system divides the tasks shown by the opera-

tor into several subtasks of navigation between intersections. In this prototype,

the subtasks are directly shown by the operator in order to simplify the experi-

mentation, since the experimental environment is small and there are only two

intersections as shown in Fig. 4. The operator shows navigation paths 1 through

4 to the VAs as subtasks, and noti�es the places A, B and C. In this prototype,

A and C are navigation goals.

The VAs can detect robots more robustly with redundant visual infor-

mation. In Fig. 5, for example, the visual �eld of VA4 does not overlap with

8 Toru Ishida

VA1VA2

VA3 VA4

Robot 1

Robot 2

Fig. 5 Overlaps of VAs' visual �elds

Controllingthe robot

Vi

Robot

Vision agent

Camera

i

i*

camera parametersEstimating Planning

a robot motion

Selectingthe plans

Integratingthe plans

Observingthe robot motion

i

Memorizedrobot path

Observation

Fig. 6 Communication between VAs and a robot in the navigation phase

those of VA1 and VA3. Therefore, when all VAs observe a robot with the same

visual features, it is estimated that VA4 observes a di�erent robot from that

of other VAs. In the developed system, the overlaps are acquired by simulta-

neously observing a robot used in the task teaching phase from all VAs, then

they are used in the navigation phase. In addition, the VAs can robustly detect

robots using knowledge of local environments around them. In this prototype,

for example, the VAs recognize the objects as robots if their bottom is in the

free region in the view of the VAs as shown in Fig. 3.

3.3 Navigation of mobile robots

After the task teaching phase, the DVS navigates robots by iterating

the following steps (see Fig. 6):

1. A robot requires the DVS to navigate itself to a destination.

Mobile Robot Navigation by a Distributed Vision System 9

Robot

wi

(1) Nearest point

Memorized path

t(2) Navigation goal

i

View of VA i

(5) i*

Vi*

Vi(4)

Gi*

Gi(3)

Fig. 7 Generating a navigation plan

2. VAs communicate with each other to make a global navigation plan to

its destination.

3. Each VA sets an instant navigation goal in front of the robot, then

generates a navigation plan and sends it to the robot.

4. The robot receives the navigation plans from multiple VAs, then selects

proper ones, integrates them, and moves based on the integrated plan.

Each VA generates a navigation plan as follows. First, the VA estimates

the nearest point on the memorized paths from the robot in its view (see Fig. 7

(1)). Then, the VA sets an instant navigation goal at a certain distance �t

from the estimated position (Fig. 7 (2)), and computes an angle ��i between the

direction to the goal Gi = (gxi; gyi)T and the current motion direction of the

robot Vi = (vxi; vyi)T as follows (see Fig. 7 (3), (4) and (5)):

��i = arcsin

�V

�

i�G

�

i

jV�

ijjG

�

ij

�(1)

where V �

iand G�

iare:

V�

i=

1 0

01

sin�i

!Vi; G

�

i=

1 0

01

sin�i

!Gi:

�i is the tilt angle of the vision sensor of VA i, which is automatically estimated

by observation (see Section 3.4). Note that we assume orthographic projection

here. Each VA sends ��i to the robot as a navigation plan. When a VA has

10 Toru Ishida

detected an obstacle (including other robots) between the navigation goal and

the robot, the VA modi�es the goal so as to avoid the obstacle.

After the robot received the navigation plans (��i ) from VAs, it estimates

the error of ��i to eliminate navigation plans which include large errors. The error

of ��i includes an observation error of the motion direction of the robot and an

estimation error of �i (the estimated tilt angle of VA i). Here, we assume that

the former error is inversely proportional to the apparent size of the robot wi in

the view of VA i. The latter error (let this be ���

i ) is related to the error of the

estimated tilt angle �i and is computed from equation (1) as follows:

���i = arcsin

V

�

i

0

�G�

i

0

jV�

i

0

jjG�

i

0

j

!� arcsin

�V

�

i�G

�

i

jV�

ijjG

�

ij

�(2)

where V �

i

0

and G�

i

0

are:

V�

i

0

=

1 0

01

sin(�i +��i)

!Vi;

G�

i

0

=

1 0

01

sin(�i +��i)

!Gi:

��i is the estimated error of �i (see Section 3.4). Then, if the size of the

robot observed by a VA is less than 2/3 of the largest of all, the robot considers

that the navigation plan generated by the VA includes a relatively large error,

and eliminates it. Furthermore, the robot also eliminates navigation plans the

estimated errors of which (i.e., ���i ) are more than twice of the smallest of all.

Next, the robot integrates the remaining navigation plans. Since they

are represented with angles on a common coordinate system along the motion

direction of the robot, they can be integrated by computing an average angle

of them. Here, the robot computes an average angle of ��i weighted with the

estimated error ���i as follows:

�� =

Piki�

�

iPiki

; ki =wi

j���i j; (3)

where wi is an apparent size of the robot observed by VA i. Finally, the robot

generates actuator commands based on ��.

3.4 Estimating camera parameters by observation

Mobile Robot Navigation by a Distributed Vision System 11

RobotPath of the robot

x

yV

i

i

View of VA i

i

Vi

v

ui

Fig. 8 Estimating camera parameters

In general, the position of a vision sensor is represented with six param-

eters: rotation and translation parameters. Hosoda and Asada 3) proposed a

method for estimating the camera parameters by visual feedback. In the DVS,

each camera observes moving robots in an environment in order to estimate the

camera parameters in the same way. However, in the DVS, the bottom of the

robot is regarded as its position so that the robot position measured by observa-

tion is imprecise, which makes it di�cult to estimate all the six camera parame-

ters. Therefore, we estimate three parameters �i and �i as shown in Fig. 8 in an

on-line manner, which are needed for robot navigation if orthographic projection

is assumed.

[ 1 ] Estimation method

Let x and y be reference axes of rectangular coordinates, where the

direction of the x axis indicates the motion direction of the robot, and let �i,

�i and i be the tilt angle of VA i, the angle between VA i and the y axis, and

the rotation angle around the viewing direction of VA i, respectively. Assuming

orthographic projection, the velocity of the robot V is projected into the view

of VA i as follows:

V i = SiT iRiV

where the vector V i = (ui; vi)T is the velocity projected in the view of VA i, Ri

and Si represent rotation matrices of �i and � i, respectively, and T i represents

a matrix of orthographic projection:

Ri =

�cos�i � sin�i

sin�i cos�i

�;

12 Toru Ishida

Si =

�cos i sin i

� sin i cos i

�;

T i =

�1 0

0 sin�i

�:

Hence, the velocity V is represented as follows using V i = (ui; vi)T :

V = R�1

i T�1

i S�1

i V i

=

0@ cos�i

sin�i

sin�i

� sin�icos�i

sin�i

1A�u0i

v0i

�(4)

where u0i and v0i are:�u0i

v0i

�= S

�1

i V i

=

�cos i � sin i

sin i cos i

��ui

vi

�: (5)

Therefore,

V2 = u02i +

�v0i

sin�i

�2

:

If a human operator controls the robot with a constant speed, jV j is a known

value. Consequently, �i can be computed with the following equation:

sin�i =

sv02i

V2� u02i

(v0i 6= 0): (6)

Furthermore, the component y of the velocity V is always zero, so that �i can

be computed from equation (4) as follows:

u0i sin�i �v0i

sin�i

cos�i = 0: (7)

By observing two velocities of a robot (i.e., observing two di�erent V i), �i, (two

di�erent) �i and i are acquired based on equations (5), (6) and (7). In this

experimentation, in order to simplify the estimation we assume i = 0, that is,

the cameras are set up in parallel with the plane where robots move. On this

assumption, �i and �i are computed with equations (6) and (7), respectively,

by observing one motion of a robot. Note that, in practice, the velocity of the

robot V i is normalized with wi (the size of the robot in the view of VA i).

Mobile Robot Navigation by a Distributed Vision System 13

0

5

10

15

20

25

30

35

40

45

0 10 20 30 40 50 60 70 80 90

1% 5% 10%

[de

gree

]

[degree]

Fig. 9 Estimation error ��i

Table 1 �i of vision agents estimated by observation

VA VA1 VA2 VA3 VA4

Actual 30 31 9 28

Observation 1 21.7* 35.8 30.4* 6.49*

Observation 2 24.9 8.08* 16.9* 34.4

[ 2 ] Estimation error

The relation between an observation error of the robot velocity (�ui; �vi)

and an estimation error of the tilt angle ��i is represented as follows from equa-

tion (6):

��i = sin�1

(s(vi +�vi)2

V2� (ui +�ui)2

)� �i

where we assume i = 0. Fig. 9 shows ��i when �i = 30�, and �ui and �vi are

1%, 5% and 10% of jV j. In Fig. 9, the horizontal axis indicates �i since ui and

vi are determined by equations (6), (7), and �i. Thus, the estimation error ��i

becomes larger when �i approaches zero, that is, when the velocity of the robot

approaches the horizontal direction in the view of VA i. Note that ��i is used

in equations (2) and (3) for integrating navigation plans generated by multiple

VAs.

Table 1 shows an example of the tilt angles � (in degrees) of four VAs

estimated by observing two motions of a robot as shown in Fig. 10. Comparing

14 Toru Ishida

VA1 VA2 VA3 VA4

Observation 1

Observation 2

Fig. 10 Images used for estimating �i

with the actual angles (indicated as `Actual' in Table 1), the estimation error

becomes larger (values denoted with `�' in Table 1) when the velocity of the

robot approaches the horizontal direction as discussed above. The estimated

parameters are not exactly precise, however, the DVS can still navigate robots

with the estimated parameters though the robots move in a zigzag. This is

because the navigation plan is represented with a di�erential angle between the

current motion direction of the robot and the direction to an instant navigation

goal. Therefore, the direction (right or left) in which the robot is navigated is

not a�ected by the estimation error ��i. Furthermore, the DVS can successfully

navigate robots by integrating navigation plans generated by multiple VAs.

x4 ExperimentationWe have developed a prototype system based on the concept of the

distributed vision described above. Fig. 11 shows a model town and mobile

robots used in the experimentation. The model town, the scale of which is 1/12,

has been made for representing enough realities of an outdoor environment,

such as shadows, textures of trees, lawns and houses. Sixteen VAs have been

established in the model town and used for navigating two mobile robots.

Fig. 12 shows the hardware con�guration. Images taken with the VAs

are sent to image encoders which integrate sixteen images into one image, then

sent to a color frame grabber. The size of the whole image is 640�480 pixels, and

each VA's image is 160�120 pixels. The main computer, Sun Sparc Station 10,

executes sixteen VA modules, which process data from the color frame grabber

at 5 frames per second, and communicate with the two robots through serial

devices. The robots avoid a collision based on VA's commands, however, if a

collision is detected with their touch sensors, they move backward and change

Mobile Robot Navigation by a Distributed Vision System 15

Robots

Vision agents

Fig. 11 Model Town

Sun SS10

Serial

Video camerax 16

Robotx 2

4 1

4 1

4 1

4 1

4 1NTSC

NTSC

x 2Wireless unit

Encoders

Fig. 12 Hardware con�guration

the direction in order to avoid the collision.

In this prototype, a human operator shows VAs two robots and each

VA memorizes their colors (red and black) in order to identify them. Then,

the operator shows several paths using one of the robots. Finally, the system

simultaneously navigates the robots along the paths shown by the operator.

Fig. 13 shows images taken by VAs in the navigation phase. The vertical axis and

16 Toru Ishida

3 4 5 6 9 11 16

0

5

10

15

20

25

30

35

Time(sec)

VA

Fig. 13 Images processed by VAs

the horizontal axis indicate time and the ID numbers of the VAs, respectively.

The solid and the broken rectangles indicate VAs selected for navigating the red

and the black robot, respectively. As shown in Fig. 13, VAs are dynamically

selected according to navigation tasks. In other words, the system navigates

robots observing them from proper viewing points.

Fig. 14 shows robot trajectories navigated by the DVS exhibited at the

international conference IJCAI'97. The sixteen cameras were simply set up so as

to cover the whole environment. Although their locations were not measured, the

system continuously navigated the robots for three days during the exhibition.

The concept of the DVS, such as simple visual functions, exible navigation

strategies and use of redundant visual information, realizes robust navigation in

a complex environment.

x5 Discussion and ConclusionWe have developed a prototype of the DVS. In this paper, we mainly de-

Mobile Robot Navigation by a Distributed Vision System 17

Fig. 14 Robots navigated by the DVS

scribed the details of the navigation method of mobile robots using multiple VAs.

In addition, the prototype system partly deals with communication among VAs

and robots, and also deals with construction and management of environment

models. With the experimental result of robot navigation, we have con�rmed

that the DVS can robustly navigate mobile robots in a complex environment.

In distributed arti�cial intelligence, several fundamental works such as

Distributed Vehicle Monitoring Testbed (DVMT) 4) and Partial Global Planning

(PGP) 5) dealing with systems using multiple sensors have been reported. In

these systems, which are based on the blackboard model 6), agents symbolize

sensory information with a common representation, and gradually proceed with

their recognition by exchanging them. Thus, these systems deal with recogni-

tion based on symbolized information. On the other hand, the purpose of the

DVS is to navigate robots. In the task teaching phase, the VAs independently

memorize navigation paths as robot tasks from their own viewing points with-

out symbolizing them. In the navigation phase, the VAs plan a global path of a

robot by communicating with each other, generate instant navigation plans, and

the robot generates an instant actuator command based on the plans. Thus, the

DVS deals with motion recognition by multiple agents, and regeneration of the

robot tasks by cooperation of the agents.

As a future work, more detailed communication among VAs should be

18 Toru Ishida

considered. In the experimentation of Section 4, a human operator controls

a robot to show subtasks of the robot (i.e., navigation between intersections)

while directly indicating speci�c places such as intersections, and the VAs learn

the subtasks by observing the robot motion. However, they need to commu-

nicate with each other to autonomously lean the subtasks. In addition, in the

navigation phase, the VAs communicate with each other to make plans for navi-

gating robots to their destinations, however, in the prototype the communication

is simpli�ed and specialized for the small experimental environment. For real

world applications, more sophisticated communication will be needed to perform

exible planning by VAs' communications.

Furthermore, the following issues should be considered for extending the

scale of the DVS:

� more accurate identi�cation of multiple robots,

� dynamic organization for navigating many robots by a limited number of

VAs.

With respect to the �rst point, the DVS does not utilize a geometrical map,

which realizes robustness and exibility of the system. Instead, it will be

achieved by observing relations between robot commands and actual robot mo-

tions, and utilizing a qualitative map 7) which represents rough positional rela-

tions of VAs. With respect to the second point, we have to analyze behaviors of

the system in such a situation and develop more sophisticated communication

among the VAs.

Recent developments of multimedia computing environments have estab-

lished huge number of cameras and computers in o�ces and towns. The DVS is

considered as such growing infrastructures for robots consisting of sensors and

networks. They are expected to be integrated to become more intelligent sys-

tems | Perceptual Information Infrastructure (PII) 2), which provides various

information to real world agents such as robots and humans.

References

1) Ballard, D. H., \Reference frames for animate vision," Proc. IJCAI, pp.

1635{1641, 1989.

2) Ishiguro, H., \Distributed vision system: A perceptual information infrastruc-

ture for robot navigation," Proc. IJCAI, pp. 36{41, 1997.

3) Hosoda, K. and Asada, M., \Versatile visual servoing without knowledge of

True Jacobian," Proc. IROS, pp. 186{193, 1994.

4) Lesser, V. R. and Corkill, D. D., \The distributed vehicle monitoring testbed:

Mobile Robot Navigation by a Distributed Vision System 19

A tool for investigating distributed problem solving networks," AI Magazine,

pp. 15{33, 1983.

5) Durfee, E. H. and Lesser, V. R., \Partial global planning: A coordination

framework for distributed hypothesis formation," IEEE Trans. SMC, Vol. 21,

No. 5, pp. 1167{1183, 1991.

6) Erman, L. D., Hayes-Roth, F., Lesser, V. R. and Reddy, D. R., \The Hearsay-

II speech-understanding system: Integrated knowledge to resolve Uncertainty,"

Comput. Surveys, Vol. 12, pp. 213{253, 1980.

7) Sogo, T., Ishiguro, H. and Ishida, T., \Acquisition of qualitative spatial rep-

resentation by visual observation," Proc. IJCAI, pp. 1054{1060, 1999.