Advertisement

Search Log in

Proceedings of the Seventh International Conference on Mathematics and Computing pp 1029–1041 Cite as

Sonar Sensing-Based Human Leg Localization Using Chapter EUR 27.20

Price includes VAT (Netherlands)
Gaussian Process Regression
DOI: 10.1007/978-981-16-6890-6_77

Pritam Paral, Amitava Chatterjee & Anjan Rakshit Chapter length: 13 pages
Instant PDF download
Conference paper First Online: 06 March 2022 Readable on all devices

127 Accesses Own it forever

Exclusive offer for individuals only
Part of the Advances in Intelligent Systems and Computing book series (AISC,volume 1412) Tax calculation will be finalised during checkout

Buy Chapter
Abstract
eBook EUR 181.89
Human detection and following problems are considered as one of the crucial problems in a
Softcover Book EUR 239.79
human-robot coexisting environment. The present work focuses on a significant subproblem
within this genre, where the measurements obtained with the onboard range sensing system
Learn about institutional subscriptions
of a mobile robot are used for localizing the positions of the leg pair of a walking person at a
time instance, during human tracking. This paper demonstrates how such a human leg Sections Figures References
localization problem can be formulated as a nonlinear regression problem where two distinct
prediction models are learned by using two different training datasets, and how the learned Abstract

models can be used to localize the positions of the human legs. Although sonar sensors References

produce readings fraught with uncertainty and imprecision, they are still considered as an Author information
attractive range sensing solution for robotic applications, especially in developing countries,
Editor information
where cost, size, and power consumption associated with the sensors are the major
Rights and permissions
constraints. Keeping the sonar behavior in mind, in this paper, we propose to employ Gaussian
Process Regression (GPR), a popular Bayesian nonparametric method, to solve our formulated Copyright information

nonlinear regression problem. Here, we have aimed to suitably adapt the basic GPR About this paper
framework as per our formulated problem to perform a statistical model calibration and then
localize the human leg positions based on the calibrated model. Performance evaluations
carried out for localizing the positions of human legs based on real-life data show the
superiority of the proposed GPR-based approach over the conventional position estimation
method.

Keywords
Human leg localization Sonar sensing Gaussian process regression (GPR)

Supported by the “VISVESVARAYA PHD SCHEME FOR ELECTRONICS & IT”, Ministry of
Electronics & IT, Government of India.

This is a preview of subscription content, access via your institution.

References

1. Yuan J, Chen H, Sun F, Huang Y (2015) Multisensor information fusion for people tracking
 Sonar Sensing-Based Human Leg
Author Proof

Localization Using Gaussian Process

Regression

OF
Pritam Paral , Amitava Chatterjee , and Anjan Rakshit

RO
1 Abstract Human detection and following problems are considered as one of the
2 crucial problems in a human-robot coexisting environment. The present work focuses
3 on a significant subproblem within this genre, where the measurements obtained
4 with the onboard range sensing system of a mobile robot are used for localizing
5

7
DP
the positions of the leg pair of a walking person at a time instance, during human
tracking. This paper demonstrates how such a human leg localization problem can be
formulated as a nonlinear regression problem where two distinct prediction models
8 are learned by using two different training datasets, and how the learned models can
9 be used to localize the positions of the human legs. Although sonar sensors produce
10 readings fraught with uncertainty and imprecision, they are still considered as an
TE
11 attractive range sensing solution for robotic applications, especially in developing
12 countries, where cost, size, and power consumption associated with the sensors are
13 the major constraints. Keeping the sonar behavior in mind, in this paper, we propose
14 to employ Gaussian Process Regression (GPR), a popular Bayesian nonparametric
method, to solve our formulated nonlinear regression problem. Here, we have aimed
EC

15

16 to suitably adapt the basic GPR framework as per our formulated problem to perform
17 a statistical model calibration and then localize the human leg positions based on the
18 calibrated model. Performance evaluations carried out for localizing the positions of
19 human legs based on real-life data show the superiority of the proposed GPR-based
RR

20 approach over the conventional position estimation method.

21 Keywords Human leg localization · Sonar sensing · Gaussian process regression

22 (GPR)
CO

Supported by the “VISVESVARAYA PHD SCHEME FOR ELECTRONICS & IT”, Ministry of
UN

Electronics & IT, Government of India.

P. Paral (B) · A. Chatterjee · A. Rakshit

Department of Electrical Engineering, Jadavpur University, Kolkata 700032, India

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2022 1
D. Giri et al. (eds.), Proceedings of the Seventh International Conference on Mathematics
and Computing, Advances in Intelligent Systems and Computing 1412,
https://doi.org/10.1007/978-981-16-6890-6_77

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 2 P. Paral et al.

1 Introduction
Author Proof

23

24 In recent years, detecting people in motion and subsequently tracking them has
25 emerged as one of the major problems in human-robot coexisting environments
26 [1–7]. Such people detection and tracking schemes primarily involve estimating

OF
27 the kinematics of a moving human being in a surveillance area based on a set of
28 measurements acquired by a single or multiple sensor(s). The present work proposes
29 a new approach for addressing a significant subproblem within this genre, where
30 the readings gathered by the onboard range sensing system of a mobile robot is
31 employed for localizing the positions of the leg pair of a walking person at a specific

RO
32 time instance [1–3], in the course of human tracking.
33 Such a human leg localization problem can be formulated as a nonlinear regression
34 problem where two separate prediction models are learned by using two different
35 training datasets [8], both of which contain same input data, namely, various angular
36 orientations with respect to a certain robot pose along which the human being is
37

38

39
DP
positioned and the range readings acquired at those angles by the onboard range
sensor, whereas output data of one are the corresponding x coordinates and the other
are the y-coordinates of the leg postures, measured with respect to the robot pose
40 frame [9] located at that certain robot pose. Then the learned models can be used
41 to localize the (x, y) coordinates of a given query leg posture that corresponds to a
certain input combination of angular orientation and range measurement which does
TE
42

43 not usually coincide with the pre-specified sets of input data belonging to the training
44 datasets.
45 The most popular and commonly used range sensors in different human detec-
46 tion and tracking approaches are the laser scanners. Also, ultrasonic range sensors,
EC

47 predominantly in developing countries, are considered as appealing and effective

48 options, especially from the perspectives of weight, price, and power consumption.
49 However, localizing human leg position with sonar sensor is not as reliable as it is
50 with laser scanner, primarily on account of the low angular resolution of the sonar
51 sensors and the sparseness of the sonar readings [9]. Besides, phenomena such as
RR

52 multiple reflections or cross-talking [10] regularly take place in sonar sensing, which
53 produce significant amount of spurious readings.
54 Keeping the sonar behavior in mind, in this paper, we propose to employ Gaus-
55 sian Process Regression (GPR) [8, 11] to solve our formulated nonlinear regression
56 problem. On account of their tractable posterior computation and appealing analytical
CO

57 properties, Gaussian processes (GPs) are extensively used by the statistics, machine
58 learning, and signal processing communities for solving various identification, multi-
59 class classification, and nonlinear regression problems, subjected to real-life noisy
60 data. The GP, demonstrated to be a robust Bayesian nonparametric method for find-
61 ing solutions of such problems, enables to realize a probabilistic prediction within a
UN

62 sophisticated and elegant inference framework while holding magnificent resilience

63 to over-fitting that frequently takes place in machine learning. We apply GPR to our
64 problem where the objective is to predict the x and y coordinates for a certain query
65 input when two sets of training data are provided.

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 Sonar Sensing-Based Human Leg Localization Using Gaussian Process Regression 3

However, contrary to the standard GP, where the input data are generally assumed
Author Proof

66

67 to be fixed and certain but output data are considered to be noisy or often fraught
68 with uncertainty, the input data in our approach, i.e., angular orientations and ranges
69 suffer from a certain degree of imprecision and uncertainty whereas the output data,
70 i.e., the x/ y coordinates are presumed to be fixed and stable. Unlike the majority of

OF
71 existing sonar sensor-based visual target detection and tracking approaches, where a
72 set of ultrasonic sensors at fixed poses with respect to the robot is employed, in our
73 approach, we build a cost-effective sensing system, where a single ultrasonic sensor
74 driven by a servo motor is used, which is rotated within a certain angular range to
75 acquire range readings. The servo motor performs the actuation mechanism to drive

RO
76 the sonar sensor to rotate based on a set of user-given commands. However, in reality,
77 due to different nonlinearities (e.g., friction, dead zone, saturation, etc.) commonly
78 occurring in a servo motor-based system, the sonar sensor may not rotate precisely
79 to a specific angle through the set of commands. Hence, there is every possibility of
80 uncertainty being associated with the angles. Also, the range measurements obtained
81

82

83
DP
at the corresponding angles undergo a certain degree of imprecision and uncertainty
because of the constraints of the sonar sensing discussed earlier. On the other hand, the
x/ y coordinates are measured on a two-dimensional real plane using a conventional
84 distance measurement methodology, and thus can be assumed to be fixed and reliable.
85 So, in this paper, we have aimed to suitably adapt the fundamental input-output
86 characteristics of the GPR to perform a statistical model calibration and then localize
TE
87 human leg positions based on the calibrated model.
88 The rest of this paper is organized as follows. Section 2 discusses an overview
89 of the Gaussian process regression (GPR). Section 3 describes our proposed human
90 position estimation approach. Section 4 presents the experimental results and per-
formance evaluations of our proposed research work. Then, Sect. 5 concludes the
EC

91

92 paper.

2 Gaussian Process Regression

RR

93

94 The objective of nonlinear regression is to retrieve an unknown function g : Rn →

95 R from an arbitrary dataset D = {z j , w j }uj=1 , where for the n-dimensional input
96 vector z j ∈ Rn , the observation w j ∈ R (a scalar) can be considered as related to the
underlying function value g(z j ) through a Gaussian noise model [12]:
CO

97

98 w j = g(z j ) +  j with  j ∼ N (0, h2 (z j )) (1)

99 where  j , termed as the observation error (or measurement noise), is commonly

assumed to be independently and normally distributed with a mean of 0 and a vari-
UN

100

101 ance of h2 (z j ) [12]. For ease of notation, the true function value and the noise
102 standard deviation are indicated here by g j = g(z j ) and h j = h(z j ), respectively.
103 Also, the sets of inputs, outputs, true function values and noise standard deviations

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 4 P. Paral et al.

are populated into Z = [z1 , . . . , zu ]T , w = [w1 , . . . , wu ]T , g = [g1 , . . . , gu ]T , and

Author Proof

104

105 h = [h1 , . . . , hu ]T , respectively.

106 The Gaussian process (GP) with which a nonparametric Bayesian modeling for
107 an unknown function is performed, delineates a distribution of the function values
108 g(z) on the given input space z by its mean function μ(z) and covariance function
c(z, z ). For the finite set of inputs, z1 , . . . , zu , the corresponding function values,

OF
109

110 g(z1 ), . . . , g(zu ), are then fully specified by a multivariate Gaussian distribution as
111 follows [8]: ⎡ ⎤
g(z1 )
⎢ .. ⎥
112 ⎣ . ⎦ ∼ N (μ, C), (2)

RO
g(zu )

113 where ⎡ ⎤ ⎡ ⎤
μ(z1 ) c(z1 , z1 ) · · · c(z1 , zu )
⎢ ⎥ ⎢ ⎥
114 μ = ⎣ ... ⎦ and C = ⎣ ..
.
..
.
..
. ⎦ (3)

115
μ(zu ) DP
c(zu , z1 ) · · · c(zu , zu )

A usual assumption is to place a GP with μ = 0 prior over the function value,

116 especially when no preceding information about the form of g(z) is available. This
117 is given as [12]
TE
118 p(g|Z) ∼ N (0, C) (4)

119 The covariance (kernel) function c(zk , zl ) is considered as a design parameter which
120 delineates correlations between the function values gk and gl corresponding to any
121 two given inputs zk and zl . A popular choice among the various available covariance
EC

122 functions is the squared exponential (SE) function. The infinitely differentiable SE
123 kernel is expressed as [12]:

zk − zl 2
124 c(zk , zl ) = αg2 exp − , (5)
2γ 2
RR

125 where · indicates the Euclidean distance between zk and zl , αg2 is the prior variance
126 of the signal amplitude, and γ is the characteristic length-scale. The SE kernel is
127 parameterized by φ g = {αg , γ }.
Now, the joint distribution for the observed output w and the function value gtest
CO

128

129 at a test input ztest is given as [12]

  
w 0 C+V Ctest
130 ∼N , , (6)
gtest 0 CTtest c(ztest , ztest )
UN

131 where V denotes the diagonal matrix of noise variances whose entries are [V] j j =
132 h2 (z j ) and Ctest = [c(ztest , z1 ) c(ztest , z2 ) · · · c(ztest , zu )]T . Based on the con-
133 ditional identity of a multivariate Gaussian distribution, the posterior distribution of
134 the function value gtest at the test input ztest is computed as [12]

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 Sonar Sensing-Based Human Leg Localization Using Gaussian Process Regression 5

p gtest |ztest , φ g , h, D ∼
Author Proof

135
 
136 N CTtest (C + V)−1 w, c(ztest , ztest ) − CTtest (C + V)−1 Ctest (7)
137

138 Then, we can obtain the posterior distribution over the test output wtest simply by
summing up the noise variance h2test at test input ztest with the posterior variance of

OF
139

140 gtest as [12]

141 p wtest |ztest , φ g , h, htest , D ∼

 
N CTtest (C + V)−1 w, c(ztest , ztest ) − CTtest (C + V)−1 Ctest + h2test .

RO
142

143
(8)

144 This conditional distribution presents the statistical property of the predicted test
145 output gtest , where the best estimate is generally selected as its conditional mean
146 CTtest (C + V)−1 w.
147

148

149
DP
Note that in practical modeling scenarios, neither the noise level h(z) nor kernel
parameters φ g are generally accessible, and they have to be learned from the given
data. In this paper, similar to the standard GP, the noise level is supposed to be
150 constant across the input space, and hence we define the noise variance h2 (z) ≡ αn2
151 and accordingly the noise matrix V ≡ αn2 I. The unknown parameters of the Gaussian
process regression (GPR), including φ g and αn , are then together referred to as GP
TE
152

153 hyperparameters, represented as φ w = {φ g , αn } ≡ {αg , γ , αn }.

EC

154 3 Proposed Human Position Estimation Approach

155 3.1 Framework for Data Acquirement

RR

156 During human following, we need to estimate the positions of the leg pair of the
157 target person at successive time instances. A wheeled mobile robot endowed with a
158 servo motor-driven sonar sensor is used for this purpose. In our proposed approach,
159 the onboard sonar sensor, within an angular range of 180◦ , gathers a group of range
160 readings corresponding to the leg posture of the target person at a certain time instant,
CO

161 and using that group of measurements, we aim to estimate the two-dimensional
162 human leg position at that time instant.
163 With this objective, we first generate two comprehensive datasets, both of which
164 share same input information, i.e., the angular orientations of various human leg
165 postures with respect to a definite robot pose and the range measurements acquired
at those angles, but contain different output information, i.e., one comprising the
UN

166

167 corresponding x coordinates and the other the y coordinates of the leg postures,
168 measured with respect to the robot pose frame located at that specific robot pose,
169 where the position of the robot in a two-dimensional plane is considered as the
170 reference point with coordinate (0, 0) and the heading of the robot is considered

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 6 P. Paral et al.

along the y-axis. The robot, by using its onboard servo-driven ultrasonic range finder,
Author Proof

171

172 acquires a set of range readings r̃di ≈ rd , i = 1, 2, . . . , M corresponding to the leg

173 poses of a target person, when the person is positioned around a radial distance rd
174 along a pre-specified set of equally spaced angles ψi ∈ [0◦ , 180◦ ], i = 1, 2, . . . , M,
175 and, at the same time, we measure the positions of the human legs in that two-

OF
176 dimensional real plane using a conventional distance measurement technique. This
( j)
177 mechanism is repeated for a set of P number of different radial distances rd , j =
178 1, 2, . . . , P. Therefore, we can obtain a total of N = P ∗ M number of different
179 {r̃d , ψ} combinations, and correspondingly same number of x and y coordinates,
180 which are aggregated into two datasets Dx and D y as

RO
181 Dx = {(r1 , θ1 ), (r2 , θ2 ), . . . , (r N , θ N ), x1 , x2 , . . . , x N }
182
183
D y = {(r1 , θ1 ), (r2 , θ2 ), . . . , (r N , θ N ), y1 , y2 , . . . , y N }

184 where r1 =r̃d(1) , θ1 =ψ1 , . . . , r M = r̃d(1) , θ M = ψ M ; r M+1 = r̃d(2) , θ M+1 = ψ1 , . . . ,

185

186

187
r2M =r̃d(2)
M
, θ
1

2M =ψ M
θ N = ψ M (as N = P M).
; . . . . . . ; r
M

(P−1)M+1 = r̃DP
(P)
d1 , θ(P−1)M+1 =
1

ψ (P)
1 , . . . , r N = r̃ d M ,

Note that when we acquire range data for the leg pair positioned around rd along
(q)

188 a particular angle ψk , we gather a group of range readings around the angle instead
189 of acquiring a single reading at the angle. This is primarily done to obtain a reliable
TE
190 range estimation for the above configuration and to marginalize the sonar uncertainty,
191 which may not be achieved with a single reading. However, for generating dataset, we
(q)
192 take into account a single representative reading r̃dk corresponding to that specific
193 configuration, which is defined as
 
EC

(q) 1 
194 r̃dk = range(q) (ψLk ) + range(q) (ψRk ) + range(q) (δ)
L
δ∈V
195
196
with ψLk < ψk < ψRk , (9)
RR

197 where

198 V = {δ : ψLk < δ < ψRk and range(q) (δ) = NaN}

199
200
L = cardinality(V) + 2.
CO

201 The variables ψLk and ψRk denote the angular orientations at which the left most and
202 right most sonar readings are detected, respectively, corresponding to the leg pair
(q)
203 positioned around rd along ψk , and range(q) (ψLk ) and range(q) (ψRk ) are the
204 corresponding sonar readings. Also, for this specific arrangement, we have measured
the (x, y) coordinates of the center of two feet of the target person as the real location
UN

205

206 of a human body usually lies at that position. Thus, each of such N arrangements
207 can be characterized by a single set of values for r , θ , x, and y. Figure 1 illustrates

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 Sonar Sensing-Based Human Leg Localization Using Gaussian Process Regression 7
Author Proof

OF
RO
Dead zone readings
DP
Fig. 1 Acquiring range readings corresponding to the leg pair positioned around rd along ψk . Here,
rd = 20 cm and ψk = 70◦ . The magenta points indicate the range readings
TE

208 the scheme of acquiring range readings corresponding to a specific leg posture along
209 a certain angle with a representative example.
EC

210 3.2 Proposed Measurement Model

211 As mentioned earlier, because of various nonlinearities usually occurring in a servo

RR

212 motor-based system, the sonar sensor may not rotate precisely to a specific angle
213 through a user-given set of commands. Hence, there is every possibility of uncer-
214 tainty being associated with the angle measurements. Also, it is well known that
215 the performance of the sonar sensors is significantly affected by the environmen-
216 tal phenomena, sensor design and target characteristics, and sonar readings usually
CO

217 undergo some degree of uncertainty and imprecision. However, errors involved in
218 measuring the ground truth (x, y) coordinates have been considered negligible, tak-
219 ing into account the reliability of the method adopted for measurement. Therefore,
220 the inputs (r, θ ) can be considered fraught with imprecision and uncertainty, whereas
221 the outputs (x, y) can be considered stable and reliable in our framework.
UN

222 Now, our human leg localization problem can be posed as a nonlinear regres-
223 sion problem, where two separate prediction models, best represented by the
224 unknown functions fx : R2 → R and fy : R2 → R, are learned by using two dis-
225 tinct training datasets, Dx = {s j , x j } Nj=1 and D y = {s j , y j } Nj=1 , respectively, where

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 8 P. Paral et al.

s j = [r j , θ j ] ∈ R2 indicates a two-dimensional input vector, and x j ∈ R and y j ∈ R

Author Proof

226

227 are the scalar measured outputs such that

228 x j = fx(s j ) + x j with x j ∼ N (0, h2x (s j ))

229 y j = fy(s j ) +  y j with  y j ∼ N (0, h2y (s j )), (10)

OF
230

231 where both x j and  y j signify measurement uncertainties which can be represented as
232 the equivalent modeling errors. These equivalent modeling errors which are caused
233 because of the uncertainties in input variables are random variables and are assumed
234 to be independently and normally distributed, as mentioned in (10). The noise vari-

RO
235 ances h2x (s j ) and h2y (s j ) are supposed to be constant throughout the input space, and
236 defined as h2x (s) ≡ αnx 2
and h2y (s) ≡ αny
2
, respectively. Also, we define hx j = hx (s j )
237 and hx = [hx1 , . . . , hx N ] . Then these two learned models are utilized to localize
T

238 the (xt , yt ) coordinates of a given query leg posture which corresponds to a certain
input st = [rt , θt ]. It should be noted that st = [rt , θt ] does not in general coincide
239

240

241

242
DP
with the pre-specified s j = [r j , θ j ], j = 1, 2, . . . , N , and hence xt and yt are usu-
ally not known but need to be predicted from the sets of other available variables
x j and y j , j = 1, 2, . . . , N .
243 We propose to apply GPR to our discussed problem in which the objective is to
244 predict xt = fx(st ) + xt and yt = fy(st ) +  yt for a query input st = [rt , θt ], given
TE
245 two sets of (training) data [x1 , x2 , . . . , x N ]T and [y1 , y2 , . . . , y N ]T available for the
246 corresponding inputs s1 , s2 , . . . , s N . However, we will discuss here in the context of
247 estimating xt , which can be followed in the same way to estimate yt also, except the
248 fact that the training dataset D y will be involved in the later case.
249 In this work, we have selected SE kernel as the covariance function. It is worth
EC

250 mentioning that, we have standardized the input data so that each dimension is
251 weighted on the same scale. The SE kernel, for any two given inputs sk and sl , is
252 here defined as
⎡  2 ⎤
 rˆ − rˆ
2
+ θˆ − ˆ
θ
RR

sk − sl 2
⎢ k l k l
⎥
253 c(sk , sl ) = αfx
2
exp − 2
= αfx2
exp ⎣− 2 ⎦,
2γfx 2γfx

254 where rˆj and θˆj are the updated versions of r j and θ j , respectively, after standardiza-
CO

255 tion. Following the discussions of the last section, the SE kernel is here parameterized
256 by φ fx = {αfx , γfx }, and the GP hyperparameters are given by φ x = {αfx , γfx , αnx }. If
257 we define X = [x1 , x2 , . . . , x N ]T , a multivariate Gaussian distribution can be formed
258 similar to (6) as   
X 0 C+V Cst
259 ∼N , (11)
CTst c(st , st ),
UN

fxt 0

260 where fxt denotes the value of the function fx at the query input st = [rt , θt ],

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 Sonar Sensing-Based Human Leg Localization Using Gaussian Process Regression 9
⎡ ⎤
c(s1 , s1 ) · · · c(s1 , sv )
Author Proof

⎢ .. .. .. ⎥
261 C=⎣ . . . ⎦
c(sv , s1 ) · · · c(sv , sv )
262

Cst = [c(st , s1 ) c(st , s2 ) · · · c(st , sv )]T (12)

OF
263

264 and V is the diagonal matrix of noise variances whose entries are [V] j j = h2x (st ) =
265 αnx
2
.
266 Similar to (8), the conditional distribution p xt |st , φ x , hx , hxt , Dx is here
obtained as

RO
267

268 p xt |st , φ x , hx , hxt , Dx ∼

 
269 N CTst (C + V)−1 X, c(st , st ) − CTst (C + V)−1 Cst + αnx
2
. (13)
270

271

272

273
distribution:
xt = CTst (C + V)−1 X
DP
As mentioned in the previous section, our best estimate for xt is the mean of this

(14)

274 and the uncertainty associated with our estimate is captured in its variance:
TE

275 var(xt ) = c(st , st ) − CTst (C + V)−1 Cst + αnx

2
(15)

4 Experimental Results
EC

276

277 4.1 Experimental Setup

The proposed human leg localization approach is tested using real-world range data
RR

278

279 acquired by the onboard HC-SR04 ultrasonic range sensor of an inexpensive Rasp-
280 berry Pi 3 model B+ based differentially driven wheeled mobile robot [7]. The robot,
281 depicted in Fig. 2, has been indigenously developed in our Electrical Measurement
282 and Instrumentation Laboratory of Electrical Engineering Department, Jadavpur Uni-
versity, India [7]. In this scheme, Python 2.7 (along with Numpy, Scipy and
CO

283

284 Matplotlib) has been used as the experimental platform.

285 4.2 Dataset Generation

UN

286 Based on the framework discussed in section, we generate two real-world datasets
287 Dx = {(r j θ j ), x j } Nj=1 and D y = {(r j θ j ), y j } Nj=1 corresponding to various human
288 leg positions. Considering an effective sonar range of 5 − 55 cm in this work, the

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 10 P. Paral et al.

Fig. 2 The wheeled mobile

Author Proof

robot used in this paper for

testing the proposed
GPR-based approach [7]

OF
RO
( j)
set of radial distances rd , j = 1, 2, . . . , P is adopted as {5 cm, 7 cm, 9 cm, . . . ,
289

290

291

292
DP
51 cm, 53 cm, 55 cm}, where the distances are separated by an interval of 2 cm. On
the other hand, because of the presence of dead zone nonlinearity in the sensing sys-
tem, we consider an effective angular range of [10◦ , 170◦ ], and taking into account the
293 shape of the object and the lower angular resolution of sonar sensor, the set of angular
294 orientations ψi , i = 1, 2, . . . , M are selected as {10◦ , 25◦ , . . . , 160◦ , 170◦ }, with
TE
295 the consecutive angles separated by a span of 15◦ (except the last angle which is
296 separated from the preceding one by 10◦ ). Thus, the parameters P and M are set to
297 26 and 12 respectively. Accordingly, we obtain the parameter N = 26 ∗ 12 = 312.
298 In this paper, we have used approximately 80% of the observations for the purpose
of training and the remaining 20% for the purpose of testing.
EC

299

300 4.3 Performance Evaluations

RR

301 The performance of our proposed human leg localization approach has been evaluated
302 quantitatively by means of two metrics, namely mean absolute percentage error
303 (MAPE) and root mean square error (RMSE), and compared with traditional method
304 of computing (x, y) for a given (r, θ ) combination.
CO

305 1. Mean Absolute Percentage Error (MAPE): The MAPE is a measure of the
306 prediction accuracy achieved by a forecasting method, such as regression. In this
307 paper, the MAPE is defined as

S  
1   Gi − Ei 
λ= ,  × 100%
UN

(16)
S i=1  Gi
308

309 where S is the testing dataset size, Gi and Ei are the ground truth observation
310 and estimated output corresponding to the ith test input, respectively. Our pro-

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 Sonar Sensing-Based Human Leg Localization Using Gaussian Process Regression 11

Table 1 Comparison of performance metrics between proposed and traditional approach in esti-
Author Proof

mating x and y
Method Estimation of x Estimation of y
λ (in %) ξ (in m) λ (in %) ξ (in m)
Traditional 20.14 0.0357 20.61 0.0396

OF
approach
Proposed 06.22 0.0155 06.51 0.0137
approach

RO
DP
(a) (b)
TE
Fig. 3 APEs for the estimated outputs corresponding to the testing inputs: a in the case of estimating
x-coordinates, b in the case of estimating y-coordinates

311 posed approach outperforms the traditional approach in localizing both x and
EC

312 y coordinates, in terms of this quantitative performance metric, as presented in

313 Table 1. Also, Fig. 3a, b show the absolute percentage errors (APEs) for the esti-
314 mated outputs corresponding to the testing inputs, in the cases of estimating x
315 and y coordinates respectively. As can be observed from Fig. 3, our GPR-based
proposed approach and traditional method show better results for approximately
RR

316

317 85% and 15% of the test instances, respectively, in the case of estimating x, and
318 approximately 88% and 12% of the test instances, respectively, in the case of
319 estimating y, which reaffirms the overall usefulness of our proposed GPR-based
320 approach. In our experimentation, the proposed approach is found to improve the
MAPE performance over the traditional method by 69.11% and 68.41% in the
CO

321

322 case of estimating x− and y-coordinates, respectively.

323 2. Root Mean Square Error (RMSE): The RMSE is a standard statistical parameter
324 to measure the error of a model in predicting quantitative data. In this paper, the
325 RMSE is defined as follows:

 S
UN


 (Ei − Gi )2
 i=1
326 ξ= (17)
S

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 12 P. Paral et al.

As revealed by the experimental results in Table 1, our proposed approach is

Author Proof

327

328 superior to the traditional approach, also in terms of this metric. In our experi-
329 mentation, the proposed approach is found to improve the RMSE performance
330 over the traditional method by 56.58% and 65.40% in the case of estimating x−
331 and y-coordinates, respectively.

OF
332 We have evaluated the performance of our proposed GPR-based approach with the
333 parameter values αfx = αfy = 1, and γfx = γfy = 0.7.

RO
334 5 Conclusions

335 This paper demonstrates how a human leg localization problem can be formulated
336 as a nonlinear regression problem where two separate prediction models are learned
337 by using two different training datasets, and how the models can be used to localize
338

339

340
DP
the x− and y-coordinates of a leg posture with respect to a specific robot pose frame.
Also, in this paper, we have adopted sonar sensing-based system for acquiring range
data and considering the uncertainty and imprecision associated with the sonar read-
341 ings, we have proposed to employ a popular Bayesian nonparametric method, called
342 Gaussian Process Regression (GPR), to solve our formulated nonlinear regression
TE
343 problem. We have suitably adapted the basic I/O characteristics of the GPR accord-
344 ing to our formulated problem, performed a statistical model calibration, and then
345 predict human positions based on the calibrated model. Experimental results and per-
346 formance evaluations carried out for estimating the positions of human legs based
347 on real-life data aptly demonstrate the superiority of the proposed approach over the
EC

348 traditional method.

349 References
RR

350 1. Yuan J, Chen H, Sun F, Huang Y (2015) Multisensor information fusion for people tracking
351 with a mobile robot: a particle filtering approach. IEEE Trans Instrum Meas 64(9):2427–2442
352 2. Bellotto, N, Hu H (2009) Multisensor-based human detection and tracking for mobile service
353 robots. IEEE Trans Syst Man Cybern Part B (Cybern) 39(1):167–181
3. Li D, Li L, Li Y, Yang F, Zuo X (2018) A multi-type features method for leg detection in 2-D
CO

354

355 laser range data. IEEE Sens J 18(4):1675–1684

356 4. Wang M, Liu Y, Su D, Liao Y, Shi L, Xu J, Miro JV (2018) Accurate and real-time 3-D
357 tracking for the following robots by fusing vision and ultrasonar information. IEEE/ASME
358 Trans Mechatron 23(3):997–1006
359 5. Paral P, Chatterjee A, Rakshit A (2019) Vision sensor-based shoe detection for human tracking
360 in a human-robot coexisting environment: a photometric invariant approach using DBSCAN
UN

361 algorithm. IEEE Sens J 19(12):4549–4559

362 6. Paral P, Chatterjee A, Rakshit A (2019) OPTICS-based template matching for vision sensor-
363 based shoe detection in human-robot coexisting environments. IEEE Trans Instrum Meas
364 68(11):4276–4284

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard

 Sonar Sensing-Based Human Leg Localization Using Gaussian Process Regression 13

7. Paral P, Chatterjee A, Rakshit A (2020) Chaos-based random sampling for photometric invari-
Author Proof

365

366 ant shoe detection with vision sensor in human-robot coexisting environments. IEEE Sens J
367 20(10):5424–5434
368 8. Lee S, McBride J (2019) Extended object tracking via positive and negative information fusion.
369 IEEE Trans Signal Process 67(7):1812–1823
370 9. Burguera A, Gonzàlez Y, Oliver G (2008) A probabilistic framework for sonar scan matching

OF
371 localization. Adv Robot 22(11):1223–1241
372 10. Burguera A, Gonzàlez Y, Oliver G (2009) On the use of likelihood fields to perform sonar scan
373 matching localization. Auton Robot 26(4):203–222
374 11. Bishop CM (2006) Pattern recognition and machine learning, 1st edn. Springer, Berlin
375 12. Zhang Q-H, Ni Y-Q (2020) Improved most likely heteroscedastic Gaussian process regression
376 via Bayesian residual moment estimator. IEEE Trans Signal Process 68:3450–3460

RO
DP
TE
EC
RR
CO
UN

510918_1_En_77_Chapter  TYPESET DISK LE  CP Disp.:12/11/2021 Pages: 13 Layout: T1-Standard