Robotics and Autonomous Systems 50 (2005) 1325

Differentiating walls from corners using the

amplitude of ultrasonic echoes
Gines Benet , Milagros Martnez, Francisco Blanes, Pascual Perez, Jose E. Simo
Departamento de Informatica de Sistemas y Computadores, Universidad Politecnica de Valencia, P.O. Box 22012, 46080 Valencia, Spain

Received 27 March 2003; received in revised form 16 June 2004; accepted 8 July 2004
Available online 28 August 2004

Abstract

This paper analyzes the use of ultrasonic echo amplitudes to evaluate the characteristics of the detected surfaces (such as
distinguishing between walls and corners). The shape and surface characteristics of the environment, such as roughness or
absorption coefficient, as well as the distance and the angle of incidence, have high influence on the amplitude of the echoes. As
a consequence, the amplitude of the received echoes has received little attention from robotic researchers. Instead, time-of-flight
(ToF) has been used as the main source of information of the environment. It is also well known that the shape of the echoes
of a corner is the same as that of a wall [IEEE Trans. Pattern Anal. Mach. Intell. (PAMI), 12 (1990) 560]. Several authors have
proposed special multi-transducer configurations to avoid this problem. This paper studies the amplitude of received echoes
and presents a simple model to predict the shape and amplitude of echoes received from different materials in environments
composed of walls and corners. Using this model, and analysing the amplitude of the echoes it is possible to distinguish between
walls and corners in a single scan of a single ultrasonic transducer pair. The parameters of the model were obtained from tests
performed on different materials and surfaces. The last section of the paper shows the experimental results of the wallcorner
classifications obtained in real tests during the walk of a mobile robot. The results suggest that the method proposed can be of
great interest for map building in robotics.
2004 Elsevier B.V. All rights reserved.

Keywords: Ultrasonic sensors; Amplitude response model; Map building; Mobile robots

1. Introduction

Ultrasonic sensors are widely used in robotics to

Corresponding author. Tel.: +34 96 387 7578; avoid collisions and for map building purposes. They
fax: +34 96 387 7579. have some advantages over laser or camera: they are
E-mail addresses: [email protected] (G. Benet), low-cost and low-power consuming, they are not light
[email protected] (M. Martnez), [email protected]
(F. Blanes), [email protected] (P. Perez), [email protected]
dependent and, moreover, the ultrasonic signal is low
(J.E. Simo). bandwidth and hence its data processing requirements

0921-8890/$ see front matter 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.robot.2004.07.011
14 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325

are reduced. On the other hand, there are some tion; and finally, in Section 6, the results obtained from
disadvantages: environmental temperature affects several validation tests are presented and discussed.
wave absorption; wide beam width (from 30 to 60 ) of
the ultrasonic transducer lobe yields imprecise angular
measurements, low bandwidth also means low distance 2. Considerations about the amplitude of
resolution, making close objects indistinguishable. ultrasonic echoes
Several characteristics of the received ultrasonic
echoes have been used to obtain the environment fea- The amplitude of the ultrasonic echoes depends on
tures. The most common are based on the time elapsed several factors: distance, surface characteristics of the
between transmission and reception of a pulse (time- reflector, viewing angle and shape of the reflector sur-
of-flight (ToF)) [1,23]. The duration of the echo and its face. The following sections describe these factors, in-
energy are analysed in [16], and the phase differences troducing the equations used to model the ultrasonic
are used in [8,10]. A combination of ToF and echo signal amplitude.
amplitude is used in [15,24], and a combination of fre-
quency, echo amplitude and ToF is used in [2,12]. A 2.1. Distance
significant amount of early work on sonar-based map-
ping was done by Elfes [6] and Elfes and Moravec [21]. Ultrasonic signal propagates through air and the in-
In these papers, probabilistic grid maps were obtained tensity of the signal decreases with distance due to two
using an ultrasonic ring. Each cell in the map was asso- main reasons: beam spreading and air absorption. Thus,
ciated with probability density functions for occupied, the signal amplitude obtained in a receiver placed at a
empty and unknown situations. distance x from the transmitter will decrease as x in-
Kleeman and Kuc [13] stated that two transmit- creases. In the bibliography, this decay in the amplitude
ters and two receivers are necessary and sufficient for of ultrasonic signal has been modelled using different
discriminating planes, corners, and edges in two di- expressions [4,22,24]. In [5], Cracknell indicates that
mensions. This is true if only echo ToFs are used. In ultrasound attenuation in air has two factors: one expo-
the literature, we can find several transducer configu- nential, due to air absorption, and the other, hyperbolic,
rations that exploit geometric properties of ultrasonic due to beam spreading. This can be expressed as:
signal reflection to discriminate between planes and
ex
corners with a single scan, as described in [3,9,14]. A(x) = A0 (1)
But, if only one transmitter/receiver pair is used, two x
measurements from different locations will be needed where A0 is a constant, the attenuation coefficient of
for feature extraction [17,18]. the air (dB m1 ), x the distance between emitter and
The aim of this paper is to analyse the ultrasonic receiver (m).
echo amplitude, as this information can be added to It is difficult to accurately measure the attenuation
ToF in order to obtain more information from the en- coefficient, , as it depends on several factors, the most
vironment. Echo amplitude could be added to conven- important being air density, air temperature, and the
tional information, such as range and incidence angle, square of the signal frequency (f). The following for-
to perform geometric feature differentiation. The sonar mula for in dry air at 0 C is given in [5]:
system has only an ultrasonic sensor pair (T/R), and
air (dB m1 ) = 1.61 108 f 2
only a single circular scan is needed.
The paper is structured as follows: Section 2 de- Thus, for f = 40 kHz, the following value is obtained:
scribes some considerations about the amplitudes of the
air = 0.257 dB m1 (dry air at 0 C)
ultrasonic echoes; Section 3 presents an echo amplitude
response model, used to predict echo amplitudes when Eq. (1) has been selected in the present work to
the surface reflection coefficient Cr is known; Section model ultrasound amplitude decrease in air because it
4 describes the measuring system, and its main char- agrees well with the experimental data obtained. Ex-
acteristics; Section 5 shows the application of the echo haustive tests were carried out in the laboratory using
amplitude response model for wallcorner differentia- 40 kHz Massa piezo-ceramic transducers, for distances
 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325 15

ranging from 0.6 to 3 m, at air temperatures ranging at distance x, will be expressed as:
between 23 and 25 C, relative humidity of 60%. Un-
der these conditions, the value for was obtained us- e2x
A(x) = A0 Cr . (3)
ing least squares adjustment, yielding: air = 0.275 2x
0.0018 dB m1 , with a correlation coefficient of 0.935. (Note that 2x is the total path length travelled by the
ultrasonic signal).
2.2. Surface characteristics of the reector
2.3. Transducers orientation
In robotic applications, objects are detected when
the ultrasonic beam produced in a robots transmitter The ultrasonic transducer pair has different ampli-
reaches a solid surface, and is scattered back to the tude responses depending on the incidence angle () at
receivers transducer, located also in the robot. In the which the target is viewed by the transducer, showing a
case of Polaroid transducers, the same device can be maximum value Amax of received echoes when this an-
used both as a receiver and as a transmitter. In our case, gle is zero. (In the case of a flat surface, like in a wall, the
a piezo-ceramic transmitterreceiver transducer pair is target point will be the one viewed by the transducer in
used. a direction normal to the wall). The peak amplitude of
The way these reflections occur is similar to that of a the received echoes decreases as the angle increases,
mirror, but depending on the density and surface finish, but the distance (time-of-flight) will be the same for
the reflected beam will experience additional spread- all of them. Thus, if a circular scan is performed in a
ing. In addition, only a fraction of the incident energy is scene with a single object (i.e. a flat wall, or a vertical
reflected, reducing the intensity of the reflected beam. rod), for each transducers orientation an echo will be
Thus, a polished surface will produce less beam spread- received with the same time of flight, corresponding
ing than a rough surface. Also, other physical proper- to the distance between the transducers and the target.
ties of the surface, such as relative density, modify the However, the peak amplitude of the echoes will follow
percentage of the incident energy that will be reflected a Gaussian-like shape, as can be seen in Fig. 1. In [14],
back. That is, only a reduced fraction of the energy Kuc presents a formula that models this dependence of
emitted from the transmitter will arrive at the receiv- amplitude evolution with the viewing angle :
ing transducer after reflection. This fraction of energy
A() = e4
2 / 2
will depend mainly on the physical surface properties 0 (4)
[19].
where 0 is the angle that produces a value for A
= 0.02
Taking into account these factors, a simple reflection
times Amax . This formula agrees well with the real re-
coefficient Cr is proposed to model the total intensity
sponse of the transducers used. In Fig. 1, the normalised
reduction of the ultrasonic beam reflected on a sur-
real angular response of the ultrasonic transducers used
face. Thus, a given object with a uniform surface finish
in the experiments, has been plotted, together with the
will be modelled as a surface with a quasi-constant re-
normalised plot of the values obtained from Eq. (4),
flection coefficient Cr , as expressed in the following
with 0 = 50 . As can be seen in Fig. 1 the beam width
equation:
of the angular response lobe (that is, the angle corre-
Areflected sponding to a value of 0.5 Amax ) is of 21 for the
Cr = , (2) transducers used in this work.
Aincident
Thus, from the angular shape of the echoes re-
where Cr is the surface reflection coefficient, a numer- ceived in a circular scan, the normal incidence point
ical value ranging between 0 and 1, and Areflected and (referred to in this article as target point) can be eas-
Aincident are the intensity of the reflected and incident ily located for each surface, finding the maximums
ultrasonic beams, respectively. of the scanned signal. The position of this maximum
Thus, using Eqs. (1) and (2), the amplitude of an in the time axis will give us its range, and its posi-
ultrasonic echo received after reflection in a normal tion on the angle axis will give us the angular posi-
plane of a surface with reflection coefficient Cr , placed tion where the target has its normal orientation. This
16 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325

Fig. 1. Normalised shape of the angular response of the amplitude peaks of echoes received in a circular ultrasonic scan with a wall of concrete
placed in front of the transducers. Squares: measured values; solid line: the corresponding model of Eq. (4), with 0 = 50 .

fact has been used by some authors for map building Nevertheless, it is still possible to find out differ-
and for target localisation [11,12] and constitutes an ences between echo reflections from each reflector
interesting method to establish geometric features of type. As can be seen in Fig. 2, the echo reflection from
maps. a wall and the echo reflection from a right corner are
different in a substantial aspect: wall-reflected echoes
2.4. Shape of reecting surfaces only touch the surface once, whereas corner-reflected
echoes are reflected twice during their flight towards
Commonly three types of features are extracted the receiver. Thus, signal absorption will not be the
from the environments wherein the robot usually same, and some additional dispersion will be added to
navigates: walls, right corners and edges. In our the reflected beam. This suggests an interesting con-
experiments, edges produce small amplitude echoes clusion, presented in this paper: a corner will always
of the same order of magnitude as noise, making them produce smaller echo amplitudes than a wall of the
undetectable. For this reason, this paper will focus same characteristics placed at the same distance.
only on walls and corners as the main features to be
extracted from the environment.
Some differences between edges, walls and cor- 3. Amplitude response model: parameters of
ners scanned shapes were described in [23]. However, the model
other authors have reported that the peak amplitudes of
echoes reflected from walls and from right corners, for As indicated in Section 2, the ultrasonic echo am-
the same type of material, distance and incidence angle, plitude depends on several parameters. This amplitude
have almost the same shape [24]. In fact, the numer- behaviour can be easily modelled using a simplified
ous samples taken from real environments have shown equation:
that there are not meaningful differences between the
angular responses of a corner and of a wall located at e2x (42 /o2 )
A = A0 CrN e , (5)
the same distance. 2x
 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325 17

Fig. 2. Reflection of ultrasonic echoes from different surface types: (a) Wall; (b) right angle corner. Solid lines: rays that reach the transducer;
broken lines: rays that do not reach the transducer.

where A is the peak amplitude of the echo obtained in 4. Ultrasonic signal description: signal data
the ultrasonic receiver, measured in volts. A0 is a con- processing and timedistance relationship
stant for the transducers, independent of the material or
shape of the reflecting surfaces. is the attenuation co- The robot YAIR1 has a rotary ultrasonic sensor on its
efficient of the air (dB m1 ). (In this article, a value of top [7]. The sensor has two transducers: one transmitter
0.275 dB m1 will be used). x is the distance between and one receiver, placed as shown in Fig. 3, enabling
the transducers pair and the target (m). Cr is the reflec- the two transducers to have the same rotating axis. This
tion coefficient of the reflectors surface. It is a number two-transducer array rotates driven by a stepper motor
between 0 and 1, and represents the ratio between with 1.8 per step, giving up to 200 angular samples
the intensity reflected back to the transducer and the per scan.
incident intensity of the acoustic beam. This single pa- At each angular position, the emitter sends a train
rameter includes both the absorption and the additional of 16 ultrasonic pulses, with a total time of 400 s,
dispersion effects. This is an evident simplification of because the resonant frequency of the transducers is
the complex phenomena that occur in the reflection of 40 kHz. The signal obtained in the receiver is amplified
the ultrasonic beam. Nevertheless, this simplification with programmable gain, and demodulated using the
will provide good results and agrees well with the same emission frequency. Thus, the base band signal
experimental data obtained. N can take two values, is obtained after removing the frequencies above
depending on the reflectors shape: a value of 1 in the 4 kHz from the demodulated signal. The resulting
case of a wall, and a value of 2 in the case of a right cor- signal is then sampled at a rate of 10 k samples/s and
ner. That is, the number of reflections of the ultrasonic digitized using a 12 bit A/D converter. Typically, up
beam on the targets surface before reaching the re- to 256 samples of this signal are recorded before the
ceiver. (In acute corners, N can take values higher than stepper motor advances to the next position, repeating
2, but in this paper, these targets are not considered). the process. Therefore, for each angular position
is the angle at which the target point is viewed by the a vector of 256 samples is stored and processed,
transducer. and the complete scan will produce an array of 200
Eq. (5) will be used in the following sections vectors.
to predict the echo peak amplitude of an object
placed at a distance x from the transducers, viewed 1 YAIR stands for Yet Another Intelligent Robot, and is cur-
at an angle and with a known reflection coefficient rently being developed under Spanish Government CICYT Grants
Cr . DPI2002-04434-C04-03 and DROMAIN-UPV project.
18 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325

Taking into account that the shape and duration of

any echo is always the same, the time elapsed be-
tween the beginning of an echo and its peak value
will also be the same (toffset ) and is known a pri-
ori (in our case, toffset = 400 s, equivalent to four
sampling periods).
Thus, to compute the distance to a target, it is enough
to find the peak amplitude of an echo, obtaining the
time-of-flight using the following formula:
ToF = tpeak toffset . (8)
Finally, the distance to the target will be computed
as follows:
c ToF
d= (9)
2
This method is fast, but the distance resolution
is equal to d0 (1.7 cm). Assuming random error dis-
tribution,the standard error of the readings will be:
Fig. 3. Ultrasonic rotary head of YAIR robot. Note that the layout d d0 / 12 = 0.49 cm. This is sufficient for build-
enables the two transducers to have the same rotating axis. ing maps with grid sizes over 2 cm. If better resolution
is needed, interpolation methods can be used instead.
Given that ultrasound speed in air is c = 343 m/s at Thus, the only digital signal processing needed is to
20 C, the distance corresponding to a sample period subtract the toffset value from time when each sample is
(t0 = 100 s) will be: taken in a complete scan, which is equivalent to make
ct0 a shift of 4 samples in each vector. In this way, each
d0 = = 0.01715 m, (6) peak will correspond in time with the distance from a
2
target.
and the maximum range covered by a recorded echo Fig. 4 shows the typical aspect of the data obtained
will be: in an ultrasonic scan from the YAIR sensor. As shown
dmax = 256d0 = 4.39 m. (7) in Fig. 4a, the two walls and the corner are detected, as
expected. Note that in the 3D representation of Fig. 4b,
Due to the proximity between the two transducers, the echoes are grouped forming mountains and the
while the emitter is transmitting, its vibrations are me- exact position of the target points correspond to the
chanically transmitted to the neighbouring receiver, ob- relative maximums of each mountain.
taining high signal values that do not correspond with Also, note the different amplitude of the targets: the
any target. Thus, it is necessary in practice to reject the peak corresponding to the closest wall (wall A) is the
first 25 samples (2.5 ms) of the received signal, creat- largest, wall B has a medium peak value, and the peak
ing a blind zone 25 d0 = 0.43 m around the sensor. corresponding to the corner has the smaller peak value.
This is not a problem, since the sensor is placed at the
geometric centre of YAIR, and its mean radius is 0.3 m.
Thus, objects placed at 13 cm from the periphery of the 5. Application of the model: distinguishing
robot will be detected and measured. between walls and corners
From the sampled data of an echo, the distance to
the detected targets can be calculated with an accu- In this paper, only two types of targets are assumed
racy of millimetres using any of the methods described as representative of the main part of the scene: walls
in [20,22]. In our case, a simpler but accurate-enough and corners. In fact, the target points detected in
method for map building is used: scanned scenes come mainly from targets with a
 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325 19

Fig. 4. Typical ultrasonic data after scanning a rooms corner. (a) Horizontal layout with the walls superimposed. Note the coincidence of the
signal peaks with the walls and the corner. (b) 3D view of the same map, showing the different amplitudes of the echoes and their angular
evolution.

single reflection (flat surfaces) or targets with two that produce much lower amplitude echoes. In our
reflections (right corners). The flat surface amount approach, only echoes with amplitudes large enough
needed to produce an echo is small enough to represent to be considered as walls or as corners were taken into
almost any form of room outline as if it were formed account, disregarding echoes with lower amplitudes,
by small flat pieces. Acute corners produce more considered as noise for map building purposes.
than three reflections, and thus, the final intensity Eq. (5) models the theoretical behaviour of the echo
received is considerably reduced. Edges are targets peak amplitudes. This equation can be used to predict
20 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325

the readings obtained from a given scene or, even more corner, or a surface with two ultrasonic beam reflec-
important, to identify the type of target which produces tions. Of course, the measured data will have added
a given echo. In the first case, it is necessary to know noise, some due to the measurement process, and most
the parameters of each target: distance x, reflection co- due to the non-uniform value of Cr in all the sur-
efficient Cr , exponent N (1 for walls, 2 for corners), and faces; and in practice, the values of N obtained from
the angle at which the target point is observed by the Eq. (10) will have some added noise. Grouping the
transducer. From these parameters, and using Eq. (5), experimental results of real walls and of real corners,
the predicted value of the echo peak amplitude can be each target category will present a statistical distribu-
obtained. tion grouped around the values 1 and 2, respectively.
For the second casethe classification problem af- Fig. 5 shows the experimental distribution values of N
ter a reading A has been takenonly one of the above obtained using 102 readings from corners and 206 read-
mentioned parameters is needed: the reflection coeffi- ings from concrete walls (with a calculated Cr = 0.59
cient Cr . In effect, distance x is a datum obtained from and a standard deviation of 0.09). The resulting shape
the ToF of the echo, and angle is not needed anymore agrees well with the theoretical result predicted by the
if the reading A corresponds to an angle = 0 . This model.
is possible if a complete scan has been taken, since the Each distributionwhich is supposed in this paper
peaks of the scanned mountains will always corre- to be Gaussianwill have its mean and its standard
spond to a zero sight angle with the target point. Under deviation. If the assumptions made about the model
these conditions, the value of N can be derived from are valid, the mean value of N for walls must be close
Eq. (5), as the next equation shows: to 1, and for corners must be close to 2. The stan-
dard deviations obtained in the distributions of the re-
(ln(2Ax)/A0 ) + 2x
N= (10) sults will be strongly dependent on the uniformity of
ln Cr the walls and the corners in the scene, and this fact
A value of 1 for N means that the target point is must be taken into account for the classification of the
a wall (or a flat surface) and a value of 2 indicates a targets.

Fig. 5. Experimental results show that the computed values of N are grouped around the value of 1 for walls, and around 2 for corners, as
predicted by the model. The data were obtained after 102 readings from corners and 206 readings from walls. The material was concrete, with
a calculated Cr = 0.59.
 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325 21

Fig. 7. Membership functions derived from the Gaussian distribu-

Fig. 6. Example plot of two normal distributions corresponding to
tions of Fig. 6.
values of N for walls (mean m1 = 1) and corners (mean m2 = 2);
(arbitrary standard deviations 1 = 0.3 and 2 = 0.32 have been
elected). The intersecting point (N0 = 1.49 in this example) will be 6. Model parameters estimation procedure
the limit value to classify targets as walls or as corners.

The application of the model proposed requires pre-

vious knowledge of the reflection coefficient Cr , and
Let us define m1 as the mean value of the experi- the assumption that all the walls and corners of the en-
mental distributions of N for the walls in a scene, and vironment are made of the same material, and that this
m2 for the corners (i.e., m1 will be close to 1 and m2 will is uniform in all its surfaces. If a scene fails to meet
be close to 2). And 1 and 2 will be the corresponding these restrictions, the expected results in the classifi-
standard deviations. cation will be poorer. These restrictions are somewhat
In Fig. 6, two normal distributions have been plotted strong, but applying optimal estimation procedures, the
as an example, taking m1 = 1 and m2 = 2; and 1 = 0.3 parameter Cr can be dynamically estimated, improv-
and 2 = 0.32 (arbitrary values). There is an intersection ing the classification results. When different materials
point between both distributions. Let us call N0 to the occur in the same scene, Cr can take the mean value
abscissa of this point. A target with a calculated value of of the different materials. The experiments conducted
N = N0 , will have the same probability of being a wall or showed that the variability of Cr from one material to
a corner. Thus, targets with N values above N0 will have another is of the same order of magnitude as the Cr of
more probabilities of being a corner than a wall, and N the same material in most cases. Thus, the use of Cr
values below this limit will indicate more probabilities mean values will give poorer though still acceptable
of being a wall. In Fig. 7, the membership functions of results in the final map.
each class corresponding to the distributions of Fig. 6 The procedure to obtain an estimate of Cr from a
have been calculated as a function of N, showing the given material is simple, and can be undertaken offline.
probabilities of being a wall (PW ) or a corner (PC ). Previous knowledge of A0 is assumed at sensors cali-
Of course, since only two types of targets are being bration time. In our case, the value obtained was A0 =
considered, it must always follow that: PW + PC = 1. 3.948 V m, after exhaustive tests in the laboratory. The
Also, note that for N = N0 , PW (N0 ) = PC (N0 ) = 0.5. value of air attenuation was also measured, obtaining
Using these membership functions, the final informa- = 0.275 dB m1 .
tion supplied to the local map and to the fusion process Cr can be obtained from only one ultrasonic reading
will be not only which type of target has been detected, from a wall oriented normal to the transducer, at a given
but also its probability. This latter aspect can be of great distance x (calculated from the reading, using the ToF
utility in the sensor fusion processes. method). Given that N for a wall is 1 and that = 0 ,
22 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325

Table 1
Experimental values of Cr and their standard deviation for some
common materials
Material Cr (mean) S.D.
Railite 0.76 0.03
Glass 0.71 0.10
Polished plastic 0.64 0.06
Pladur 0.62 0.07
Concrete 0.59 0.09
Cork 0.57 0.07
Natural Wood 0.51 0.04
Matt Plastic 0.47 0.06

we can use Eq. (5) to obtain Cr from the value of A

obtained from the echo, as follows:
A
Cr = 2x e2x (11) Fig. 8. Amplitude vs. distance in walls and corners. Solid lines cor-
A0 respond to model of walls (upper line) and to model of corners (lower
line); squares: readings from walls; diamonds: readings from corners.
In fact, this method is applicable in all the cases
where the target point is known to be a wall, enabling Table 2
the on-line estimation of Cr . However, as Cr shows Classification results obtained in a room with concrete walls (Cr =
significant variability in its value, it is highly recom- 0.59)
mendable to take sufficient number of readings before Obstacle type Classified as
making an estimate for Cr . Using the above method, Wall (%) Corner (%)
several materials encountered in diverse rooms, offices
Wall 98 2
and corridors of our university were measured and the Corner 9 91
results obtained for Cr and their standard deviation are
shown in Table 1.
To test other non-homogeneous environments, an-
other room was selected whose walls are made with
7. Obtained results and discussion fragments of Pladur together with glass (windows)
and metal stripes as well as Railite (door panels), as-
Fig. 8 shows the peak amplitude values obtained suming in this case, a mean value of Cr = 0.62. The
from walls and corners, corresponding to a test con- classification results are summarized in Table 3.
ducted in a room whose walls are made of concrete. As Finally, the results of the scans made to a group of
it can be observed, the points are grouped around the different rooms, of different materials, using their mean
two solid lines obtained from Eq. (5) for walls (upper value (Cr = 0.64), are summarized in Table 4, grouped
line) and for corners (lower line). Some of the readings by distance intervals, and the global results. The results
corresponding to a wall overlap with those correspond- are fairly satisfactory in most cases, providing good-
ing to a corner, and vice-versa. In these cases, the clas- enough classification results.
sification algorithm fails, but in most cases, the types
of targets can be clearly differentiated. Table 3
The model was also tested in several rooms of our Classification results obtained in a room with walls made of Pladur ,
department, made mainly of the same material (con- metal and railite (Cr MEAN = 0.62)
crete). In these experiments, the robot was guided fol- Obstacle type Classified as
lowing different trajectories, and numerous scans were
Wall (%) Corner (%)
taken to distinguish and classify detected obstacles as
walls or as corners. The classification results are shown Wall 99 1
Corner 14 86
in Table 2, and are very satisfactory.
 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325 23

Table 4
Classification results obtained in several rooms
Obstacle type All distances 0 m< D < 1 m 1 m < D < 1.5 m 1.5 m < D < 2 m

Wall (%) Corner (%) Wall (%) Corner (%) Wall (%) Corner (%) Wall (%) Corner (%)
Wall 86 14 82 18 92 8 89 11
Corner 33 67 18 82 36 64 39 61

Classified as walls and corners made of different materials: pladur, concrete, metal, glass and railite. Cr MEAN in this case was 0.64.

The use of a mean value for Cr when different averaging the probability values obtained for each grid
materials are present in a given environment simplifies cell.
the recognition algorithm, but it will logically produce Fig. 9 shows the results of this test, obtained in a
poorer classification results. However, as Table 1 room made of concrete, metal and Railite . The Cr
shows, the standard deviation of the experimental Cr used in this case was 0.62. As a simplification of the
values is about the same order of magnitude than the world, only two types of objects were considered: Walls
differences between the Cr of different materials. This or Corners. Thus, using only the probability of being a
fact makes acceptable the use of a mean Cr value of the wall (PW ) is enough to represent the world: a value of
materials present in a given scene. The obtained results 0 means that the object is a corner, whereas a value of
obtained with this simplification are good-enough, as 1 indicates a wall, and a value of 0.5 means the same
Tables 24 show us. probability of being a wall or a corner. Grid cell size is
Using the membership functions derived from 4 cm.
the Gaussian distributions of a given room material, This Fig. 9 represents the probability values of
not only the classification as a wall or as a corner each cell using grey-colour scale: the values closer
is obtained, but also their associated probability, to 1.0 are light-coloured, whereas values closer to 0.0
as previously indicated. Using this procedure, a are dark-coloured. When program starts, the value for
probabilistic grid map of a room was obtained, fusing each cell in the grid map is initialised to 0.5 (50% grey
all the results obtained in a robots walk, by simply coloured). The results obtained seem good, although

Fig. 9. Grey colour grid map showing the classification results obtained in a room made of concrete, metal and Railite . The circles show the
walk of the robot during the measurements. The cell grey level shows the probability of being a wall or a corner. White colour means high
probability of being a wall, whereas black colour means high probability of being a corner. Grey colour indicates intermediate probability.
24 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325

some false corners or walls, caused by the presence [3] B. Barshan, B. Ayrulu, S.W. Utete, Neural network-based tar-
of disturbing objects, like radiators and vertical heat get differentiation using sonar for robotics applications, IEEE
water pipes, are detected. These false detections can be Trans. Robot. Autom., 2000, pp. 435442.
[4] O. Corion, A.M. Desodt, D. Jolly, Using ultrasonic for the recog-
corrected later, using higher-level geometric reasoning nition of a real space, Proceedings of the IMEKO Sixth Interna-
algorithms. As an example, the false wall detected tional Symposium Measurement and Control in Robotics, 1996,
in the lower-left corner of Fig. 9 can be discarded, pp. 415420.
because this point is placed in the intersection of two [5] A.P. Cracknell, Ultrasonics, Wykeham Publublication Ltd.,
detected walls, and it must be a corner instead. Also, London, 1982.
[6] A. Elfes, Sonar-based real-world mapping and navigation, IEEE
a false corner detected in the middle of a wall can be J. Robot. Autom. RI-3 (3) (1987) 249265.
discarded by geometric considerations. [7] J.A. Gil, A. Pont, G. Benet, F.J. Blanes, M. Martnez, A Can Ar-
chitecture for an Intelligent Mobile Robot, Proceedings of the
Third IFAC Symposium in Intelligent Components and Instru-
ments for Control Applications (SICICA97), 1997, pp. 6570.
8. Conclusions [8] F.E. Gueuning, M. Varlan, Accurate distance measurement by
an autonomous ultrasonic system combining time-of-flight and
This paper analyses the amplitude of ultrasonic phase-shift methods, IEEE Trans. Instrum. Measure. 46 (6)
echoes, and describes a simple model to predict the (1997) 12361240.
shape and amplitude of echoes received from differ- [9] A. Heale, L. Kleeman, Fast target classification using sonar,
Proceedings of the IEEE/RSJ International Conference on In-
ent materials in environments composed of walls and telligent Robots and Systems, 2001, pp. 14461451.
corners. Using this model it is possible to distinguish [10] H. Hua, Y. Wang, D. Yan, A low-cost dynamic range-finding de-
between walls and corners in a single circular scan of a vice based on amplitude modulated continuous ultrasonic wave,
single ultrasonic transducer pair. The parameters of the IEEE Trans. Instrum. Measur. 51 (2) (2002) 362367.
model were obtained after exhaustive tests on different [11] Joong Hyup Ko, Kim Wan Joo, Chung Myung Jin, A method of
acoustic landmark extraction for mobile robot navigation, IEEE
materials and surfaces. Finally, the paper presents the Trans. Robot. Autom. 12 (3) (1996) 478485.
experimental results of wallcorner classifications ob- [12] L. Kleeman, Scanned Monocular Sonar and the Doorway Prob-
tained in real tests during the walk of a mobile robot. lem, Proceedings of the IEEE/RSJ International Conference on
The results suggest that the method proposed can be of Intelligent Robots and Systems, (IROS96), 1996, pp. 93103.
great interest for map building in robotics. [13] L. Kleeman, R. Kuc, Mobile robot sonar for target localization
and classification, Int. J. Robot. Res. 14 (4) (1995) 295318.
The application of the model requires previous [14] R. Kuc, A spatial criterion for sonar object detection, IEEE
knowledge of the reflection coefficient Cr , and the Trans. Pattern Anal. Mach. Intell. (PAMI) 12 (7) (1990)
assumption that all the walls and corners of the envi- 686690.
ronment are made of the same material. Uniformity [15] R. Kuc, B. Barshan, Differentiating sonar reflections from cor-
ners and planes by employing an intelligent sensor, IEEE Trans.
in all the surfaces is also required. If a scene fails
Pattern Anal. Mach. Intell. (PAMI) 12 (6) (1990) 560569.
to meet these restrictions, the expected results in the [16] R. Kuc, O. Bozma, A physical model-based analysis of hetero-
classification will be poorer. Parameter Cr can also be geneous environments using sonar, IEEE Trans. Pattern Anal.
dynamically estimated, improving the classification Mach. Intell. (PAMI) 16 (5) (1994) 497506.
results. For different materials in the same scene, Cr [17] R. Kuc, O. Bozma, Building a sonar map in a specular environ-
ment using a single mobile sensor, IEEE Trans. Pattern Anal.
will take the mean value of the different materials.
Mach. Intell. (PAMI) 13 (2) (1991) 12601269.
[18] J.J. Leonard, H.F. Durrant-Whyte, I.J. Cox, Dynamic map build-
ing for an autonomous mobile robot, Int. J. Robot. Res. 11 (4)
References (1992) 8996.
[19] L.C. Lynnworth, Ultrasonic Measurements for Process Control:
Theory, Techniques, and Applications, Academic Press Inc.,
[1] H. Akbarally, L. Kleeman, A sonar sensor for accurate 3D target
1989.
localisation and classification, Proceedings of the IEEE Interna-
[20] D. Marioli, E. Sardini, A. Taroni, Ultrasonic distance measure-
tional Conference on Robotics and Automation, Nagoya, Japan,
ment for linear and angular position control, IEEE Trans. In-
1995, pp. 30033008.
strum. Measure. 37 (4) (1988) 578581.
[2] C. Barat, N.A. Oufroukh, Classification of indoor environment
[21] H.P. Moravec, A. Elfes, High resolution maps from wide angle
using only one ultrasonic sensor, Proceedings of the IEEE
sonar, Proceedings of the IEEE International Conference on
Instrumentation and Measurement, Technology Conference,
Robotics and Automation, 1985, pp. 116121.
2001, pp. 17501755.
 G. Benet et al. / Robotics and Autonomous Systems 50 (2005) 1325 25

[22] M. Parrilla, J.J. Anaya, C. Frish, Digital signal processing tech- Francisco Blanes received the M.S. de-
niques for high accuracy ultrasonic range measurements, IEEE gree in 1994 and Ph.D. degree in 2000, in
Trans. Instrum. Measure. 40 (4) (1991) 759763. Computer Engineering from the Universi-
[23] H. Peremans, K. Audenaert, J. Campenhout, A high resolution dad Politecnica de Valencia, Spain. Since
sensor based on tri-aural perception, IEEE Trans. Robot. Au- 1995, he taught real time computer systems
tom. 9 (1) (1993) 3648. and he is currently an assistant professor in
[24] T. Yata, A. Ohya, S. Yuta, Use amplitude of echo for environ- the Escuela Superior de Ingeniera Indus-
ment recognition by mobile robots, Proceedings of the 2000 trial at the Universidad Politecnica de Valen-
IEEE/RSJ International Conference on Intelligent Robots and cia. His research interests include: real-time
Systems, 2000, pp. 12981303. robot control, perception and sensor fusion
in mobile robots.

Gines Benet received the M.S. and Ph.D. Pascual Perez received the M.S. degree in
degrees in Industrial Engineering from the 1998 from the Universidad Politecnica de
Universidad Politecnica de Valencia, Spain, Valencia, Spain. Since 1998, he taught com-
in 1980 and 1988, respectively. Since puter technology at Facultad de Informatica.
1984, he taught computed technology and He is currently an assistant professor in
currently he is an associate professor of the the Departamento de Informatica de Sis-
Escuela Universitaria de Informatica at the temas y Computadores at the Universidad
Universidad Politecnica de Valencia. He Politecnica de Valencia. His research inter-
has been involved in several National and ests include: real-time robot control, embed-
European research projects mainly related ded systems and field-bus networks.
to Real-Time Systems and Intelligent
Instrumentation. His research interests include: mobile robots, Jose E. Simo received the MS degree in In-
intelligent sensors, robot control and sensor data fusion. dustrial Engineering in 1990 from the Uni-
versidad Politecnica de Valencia, Spain and
Milagros Martnez received the M.S. de- Ph.D. degree in Computer Science from the
gree in 1993 and Ph.D. degree in 2004, in same university in 1997. Since 1990, he
Computer Engineering from the Universi- has been involved in several National and
dad Politecnica de Valencia, Spain. Since European research projects mainly related
1994, she taught computers structure and to Real-Time Systems and Artificial Intelli-
she is currently an assistant professor in the gence. He is currently an associate professor
Facultad de Informatica at the Universidad of Computer Engineering at the Technical
Politecnica de Valencia. Her research inter- University of Valencia and his current re-
ests include: mobile robots and the use of search is focused on the development of autonomous systems and
ultrasonic sensors. mobile robots.