VOL. 7, NO.

1, JANUARY 2023 6000504

Sensor applications

Automatic Relevance Determination Kernel-Embedded Gaussian Process

Regression for Sonar-Based Human Leg Localization with a Mobile Robot
Pritam Paral1∗ , Saibal Ghosh2 , Amitava Chatterjee2∗∗ , and Sankar K. Pal1∗∗∗
1 Center for Soft Computing Research, Indian Statistical Institute (ISI), Kolkata 700108, India
2 Department of Electrical Engineering, Jadavpur University, Kolkata 700032, India
∗ Member, IEEE
∗∗ Senior Member, IEEE
∗ ∗ ∗ Life Fellow, IEEE

Manuscript received 25 October 2022; revised 21 November 2022; accepted 26 December 2022. Date of publication 29 December 2022; date of current
version 19 January 2023.

Abstract—Human leg localization problems involving sonar sensing can be posed as a nonlinear regression problem,
and, nonparametric Bayesian methods, such as the Gaussian process regression (GPR) model, are potential solution
candidates. In this work, to overcome the problem of irrelevant input features from the sonar range data, an advanced
automatic relevance determination kernel structure is proposed to be used in the GPR model instead of the commonly
used standard isotropic kernel. It is able to extract high-relevance input features even from partially trained data, thus
offering a better generalization ability while improving the prediction rates and robustness significantly.

Index Terms—Sensor applications, automatic relevance determination (ARD), Gaussian process regression (GPR), human leg localiza-
tion (HLL), mobile robot, sonar sensing.

I. INTRODUCTION However, unlike the classical GP, the GP model in [5] considered
that the output data, i.e., (x, y) coordinates are certain and stable
The problems of detecting, identifying, and following moving peo- while the input data, i.e., angular orientations and ranges could suffer
ple are considered a key problem in human–robot coexisting envi- from uncertainty and imprecision. As the sonar positioning mechanism
ronments [1], [2], [3], [4]. In our recent work [5], we addressed a involved in [5] is servo motor driven, typical irregularities such as
crucial subproblem within this genre, especially keeping in mind the friction, dead zone, saturation, etc., occur frequently, which may affect
constraints in developing countries, where the range readings obtained the rotational movement of the sonar sensor induced via a specific
with the onboard sonar sensor of a mobile robot are used for human set of software commands. This can very well infiltrate a certain
leg localization (HLL) in Cartesian coordinate space at different time extent of imprecision and uncertainty in the measurements of angular
instants, during a pursuit. The work successfully demonstrated how position and range data of the human legs. Conversely, the standard
a sonar-based HLL problem can be solved as a nonlinear regression distance measurement methodology is adopted to measure the (x, y)
problem in which two distinct sets of observed data are used to learn coordinates on a real two-dimensional (2-D) plane, which leads to
two different prediction models [5]. The observed datasets include providing certain and reliable Cartesian positional data. Thus, the work
the same input data, i.e., 1) different angular positions of the human in [5] essentially adapts the basic I/O properties of the standard GPR
leg-pair with respect to (w.r.t) a specific robot pose and 2) sonar to carry out a statistical model calibration, used for estimating the leg
readings obtained at those angles but different output data, i.e., the positions in Cartesian space.
corresponding x and y coordinates of the leg poses on the Cartesian However, with the standard isotropic squared exponential (SE)
plane. The learned models are then utilized to estimate the (x, y) kernel in GPR, the system suffers from irrelevant input features from
coordinates for an unknown leg pose associated with an arbitrary input the sonar range data. To deal with the same, we propose to replace the
combination of angle and range information. conventional isotropic SE kernel in GPR with the Matern32 (M32)
Considering sonar behavior [6], the HLL framework in [5] proposes kernel having the structure of automatic relevance determination
to apply the Gaussian process regression (GPR) [7], [8] to effectively (ARD) [9]. This specific ARD structure becomes very useful in the case
solve the regression problem formulated. The attractive analytical of independent priors over the length scales in the Gaussian covariance
properties and tractable posterior computation ability of the Gaussian models. This ARD-based kernel works as a powerful tool for the
process (GP) make it a popular tool in various machine learning, extraction of discriminative input features from the entire data, as it ju-
signal processing, and statistical applications involving multiclass diciously eliminates irrelevant input features among orientation angles
classification, identification, or nonlinear regression problems with and range measurements by setting large length-scales for them [9].
real-world noise-corrupted data. In such cases, GPR turns out to be a Additionally, an advanced predictor with varied length-scales is devel-
robust and accurate nonparametric Bayesian method offering excellent oped to enhance the robustness and accuracy of predictions. To the best
resilience against overfitting. of our knowledge and belief, this work is the first to utilize ARD-based
GPR models in the application of HLL using ultrasonic sensors.
Corresponding author: Pritam Paral (e-mail: [email protected]). Section II discusses GPR in brief. Section III presents the proposed
Associate Editor: F. Costa. HLL framework in light of the ARD kernel-based GPR. Section IV
Digital Object Identifier 10.1109/LSENS.2022.3232920 demonstrates the experimental results. Finally, Section V concludes
the letter.

2475-1472 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: JADAVPUR UNIVERSITY. Downloaded on January 20,2023 at 08:06:48 UTC from IEEE Xplore. Restrictions apply.
6000504 VOL. 7, NO. 1, JANUARY 2023

II. GLIMPSE OF GPR

The GP model is mainly used to learn an unknown function from a
set of training data. Consider the following measurement model [10]:
zk = f(sk ) + k , k ∼ N(0, g2 (sk )) (1)
where zk ∈ R denotes a noisy observation of the function f(·) at the
training point sk ∈ Rn of dimension n, and k is the zero-mean mea-
surement noise with variance g2 (sk ). The objective is to learn a machine
learning model for the predictive distribution of the target value z at a
test location s , given an observed (training) dataset D = {sk , zk }Lk=1 .
The measurements z = [z1 , z2 , . . . , zL ]T and the function value f are
jointly Gaussian distributed as [10]
     
z 0 C + V C
∼N , T (2)
f 0 C c
where C denotes the covariance matrix whose entries are [C]i j com- Fig. 1. Gathering sonar readings for the leg-pair located around rD in
puted from the covariance function c(si , s j ) at input locations si and the angular direction φk , with φk = 70◦ and rD = 20 cm. Magenta points
s j , C indicates the L × 1 covariance matrix computed by c(s , si ) represent the sonar readings. φLk and φRk , respectively, are the angles
of the left-most and right-most range readings detected [5].
between test point s and training point si , V represents the diagonal
matrix containing noise variances with elements: [V]ii = g2 (si ) = g2i ,
total M = N ∗ P number of different {r̃D , φ} polar coordinates. With
and c denotes the prior variance computed from c(s , s ) at s .
an equal number of Cartesian measurements, the combinations of polar
By putting a zero-mean GP prior over f(·) and using the multivariate
and Cartesian parameters are populated in two separate datasets as [5]
Gaussian distribution’s analytical property, the conditional distribution
that defines the posterior distribution of z [10] is obtained as follows: D1 = {(rk , θk ), xk }M
k=1 ; D2 = {(rk , θk ), yk }k=1
M
(6)
p (z |s , f , g, g , D) ∼ N(z , var (z )) (3) with r(N−1)P+1 = r̃D(N1 ) ,
θ(N−1)P+1 = φ1 , . . . , rNP = r̃D(NP) ,
θNP = φP .
Similar to [5], while acquiring the range data for the leg-pair
where g = [g1 , g2 , . . . , gP ] and g , respectively, represent the noise
T
located around rD(m) in a specific angular direction φk , a set of sonar
standard deviations at training points S = [s1 , s2 , . . . , sP ]T and test
readings, rather than a single one, are gathered by the onboard sonar
point s , f denotes the hyperparameters of the covariance function,
configuration with a 180◦ field of view. This is done primarily to curb
and z and var(z ), respectively, denote the posterior mean and
the sonar uncertainty in the process of range data acquisition. However,
variance of the test output z , which are expressed as [10]
while generating datasets in our formulated regression problem, only
z = CT (C + V)−1 z (4) single-range inputs are accepted correspond to individual angular
  instances. Therefore, taking into account the sonar imprecision, an
var (z ) = c − CT (C + V)−1 C + g2 . (5) error-tolerant equivalent range reading r̃D(m) along angle φk is generated.
k

The conditional mean z is designated as the best estimate while The averaging mechanism for obtaining r̃D(m) k
is detailed in [5]. The
conditional variance var(z ) captures the uncertainty in the estimate. scheme for acquiring sonar readings for a specific leg pose along a
particular angle is exemplified in Fig. 1.

III. HUMAN LEG-PAIR LOCALIZATION FRAMEWORK B. Proposed Leg Localization Model

A. Mode of Data Acquisition Similar to [5], in the present regression-based HLL model, two
distinct training datasets D1 = {vk , xk }M
k=1 and D2 = {vk , yk }k=1 are
M
During the pursuit, a set of sonar readings is acquired for the target used to learn two different prediction models, best described via
leg-poses at a certain time instant, and with these data, it is aimed to the unknown functions X : R2 → R and Y : R2 → R, respectively.
estimate the leg positions on the Cartesian plane at that time instant. With vk = [rk , θk ] ∈ R2 being the two-dimensional input vector, the
Following the same framework of data acquisition as in [5], two measured scalar outputs xk , yk ∈ R can be approximated as [5]
different sets of data are generated in this work as well. Each of
them contains the same input information, i.e., the angular orientations xk = X(vk ) + xk xk ∼ N(0, g2x (vk ))
of several human leg postures w.r.t a specific robotic pose and the
yk = Y(vk ) + yk yk ∼ N(0, g2y (vk )). (7)
range measurements captured at those orientation angles. However,
both the datasets comprise different output information consisting The error terms xk and yk , referred to as the equivalent model-
of the corresponding x and y coordinates of the target leg postures, ing errors (EMEs), have been introduced to take the measurement
respectively. The reference position of the robot is considered (0, 0) on uncertainties into account. The EMEs, primarily resulting from the
a horizontal 2-D plane and the robot heading is taken along the +y-axis uncertainties in input variables, can be considered independent and
direction. A set of range measurements r̃D j ≈ rD |Pj=1 are acquired for normally distributed random variables, as discussed in [5]. The noise
different leg poses of a target person while he/she is located around a variances, defined as g2x (v) ≡ βnx
2
and g2y (v) ≡ βny
2
, are assumed to
radial proximity of rD along a predefined set of equidistant angles φ j ∈ remain constant across the input space. For M number of observed
[0◦ , 180◦ ]|Pj=1 [5]. Simultaneously, the true values of corresponding leg instances, we define gx = [gx1 , . . . , gxM ]T with gxk = gx (vk ). Now,
positions are measured on the 2-D coordinate plane by the conventional for a given unknown leg posture associated with input v = [r , θ ],
line gauge-based direct distance measurement technique [5]. A set of these two learned models can be used to localize the (x , y ) coor-
N different radial distances {rD(k) }Nk=1 are considered, thus generating a dinates of the corresponding leg posture w.r.t the robotic reference

Authorized licensed use limited to: JADAVPUR UNIVERSITY. Downloaded on January 20,2023 at 08:06:48 UTC from IEEE Xplore. Restrictions apply.
 VOL. 7, NO. 1, JANUARY 2023 6000504

frame [5]. Typically, the unknown polar parameter v = [r , θ ] does

not coincide with the predefined set of parameters vk = [rk , θk ]|M k=1 ,
and thus x and y are, in general, unknown, which are required to be
estimated from known variables {xk }M k=1 , {yk }k=1 .
M

Similar to [5], the GPR is proposed to be applied in our present

leg localization problem with the aim of predicting the unknown
Cartesian coordinates x and y corresponding to a query input v =
[r , θ ], given two distinct sets of observed data [x1 , x2 , . . . , xM ]T
and [y1 , y2 , . . . , yM ]T at hand for the corresponding polar inputs
v1 , v2 , . . . , vM [5]. Following the same approach of estimating the
target output z for a test input s , as illustrated in Section II, both x
and y for v can be estimated with g2x (v ) = βnx 2
and g2y (v ) = βny
2
.

C. GPR Kernel Functions With ARD Structure Fig. 2. Functional diagram of the ARD kernel-based HLL approach.
In our previous work [5], the popular SE kernel was adopted as the
covariance function, which is given by [8]
IV. EXPERIMENTAL RESULTS

cSE (v, v ) = αSE
2
exp − v − v 2 / 2γSE 2
(8)
A. Experimental Setup
with · , αSE , and γSE , respectively, denoting the Euclidean distance For the purpose of performance evaluation, we use the real sonar data
between input points v and v , the signal amplitude, and the charac- acquired by the onboard HC-SR04 ultrasonic sensor of an economical,
teristic length-scale. The hyperparameters αSE and γSE are typically indigenously developed, two-wheel differentially driven mobile robot
denoted as SE = {αSE , γSE }. SE kernel is a stationary kernel in which based on Raspberry Pi 3 model B+ [4]. Two real-world datasets D1 =
the correlations between various points are solely influenced by the {(rk , θk ), xk }M
k=1 and D2 = {(rk , θk ), yk }k=1 are formed for different
M
term v − v , resulting in a smooth distribution. For sonar range data positions of human leg-pair in polar coordinate space. Assuming the
with numerous local fluctuations, this would be too strict. Therefore, sonar range is valid in the interval (5, 55) cm, the radial distances
in this work, Matern32 (M32) kernel is selected over the SE kernel, {rD(k) }Nk=1 are chosen as {5 cm, 7 cm, . . . , 53 cm, 55 cm} while the an-
which is expressed as [9] gles {φ j }Pj=1 are adopted as {10◦ , 20◦ , . . . , 160◦ , 170◦ } (considering
√ √ dead zone nonlinearity). The parameters N, P, and M are, respectively,
3 v−v 3 v−v
cM32 (v, v ) = αM32
2
1+ exp − (9) 26, 17, and 442. The training/testing is performed with 85%/15%
γM32 γM32 training-testing split.

where the hyperparameters αM32 and γM32 , respectively, involved

in controlling the function amplitude and smoothness, are typically B. Performance Evaluations
represented as M32 = {αM32 , γM32 }.
The estimation performance of the proposed ARD-kernel-based
In real applications, the restricted capture ability of the isotropic
HLL approach (called HLL-M32+ARD) is assessed by means of
kernels, such as the SE and M32 functions, would cause them to
three quantitative metrics: root mean square error (RMSE), mean
produce unreliable or incorrect predictions for nonlinear mapping
absolute error (MAE), and fit-goodness (R2 ) between the ground-truth
involving multidimensional input variables. For our HLL model, the
observations and predicted outputs [9], and is compared with our
inputs contain the angular orientation terms, as well as the correspond-
earlier SE-kernel-based HLL approach [5] (called HLL-SE), and the
ing range measurements, and to extract these features and enhance
conventional (r, θ ) to (x, y) conversion method. As evidenced by
accuracy, the ARD structure is embedded into the isotropic M32
Table 1, HLL-M32+ARD outperforms HLL-SE and the conventional
kernels, as given by (10) and (11) [9]
method in estimating x and y positions, uniformly for all three metrics.
    To demonstrate the effectiveness of HLL-M32+ARD, two repre-
cM32+ARD (v, v ) = αM32
2
(1 + 3ρ ) exp − 3ρ (10)
sentative cases are illustrated in Fig. 3, where estimation results for y
coordinates achieved by two competing GPR-based approaches are
with
presented. In the first case, the range measurements are varied by
r−r 2
θ −θ 2 keeping the orientation angle fixed at a certain value [see Fig. 3(a)
v = [r, θ ]; v = [r, θ ]; ρ = + (11) and (b)], whereas in the second case, the angle inputs are varied
γr2 γθ2
around a specific radial distance [see Fig. 3(c) and (d)]. To describe
where r and r are, respectively, the range readings in the directions and summarize the uncertainty related to the estimation of unknown
θ and θ . The hyperparameters γr and γθ , respectively, determine the parameters, credible intervals are used, which are especially pertinent
relevancies of range and angle inputs w.r.t the regression results. A in the context of the empirical Bayesian framework [13]. As can be
large value, in general, results in low relevancy. In the context of observed from Fig. 3(a) and (c), the mean estimation results obtained
GPR, “learning” refers to the optimization of GP hyperparameters, with HLL-SE get closer to the real data. The corresponding 95%
including kernel hyperparameters and constant noise variance, by credible intervals, however, show wide distributions, which implies
using the training dataset. In this work, Limited-memory Broyden– that high uncertainty is attained in these cases. This degradation in
Fletcher–Goldfarb–Shanno (L-BFGS) algorithm [12] is employed to the estimation performance is primarily induced by the overfitting
fit the GP hyperparameters through the maximization of log-marginal problem, implying that the solo kernel structure has a limited capability
likelihood (LML) [11]. Fig. 2 shows the functional diagram of our of generalization. With reference to Fig. 3(b) and (d), HLL-M32+ARD
ARD kernel-based approach. shows improved mean estimation over HLL-SE and, more importantly,

Authorized licensed use limited to: JADAVPUR UNIVERSITY. Downloaded on January 20,2023 at 08:06:48 UTC from IEEE Xplore. Restrictions apply.
6000504 VOL. 7, NO. 1, JANUARY 2023

Fig. 4. Effects of variation of length-scales γr and γθ on LML with unity

signal amplitude (αM32 = 1): In the processes of learning prediction
models for (a) x and (b) y positions.

of a regression-based prediction model. With the ARD-based kernel

structure, the proposed GPR-based approach can robustly learn the
prediction models by effectively eliminating irrelevant input features
and localize the (x, y) coordinates of query leg postures with enhanced
Fig. 3. Estimation results for y coordinates obtained from (a) HLL- accuracy. Real-life performance evaluations aptly demonstrate the
SE and (b) HLL-M32+ARD with range measurements being varied at superiority of our ARD kernel-based approach.
orientation angle 60◦ ; (c) HLL-SE and (d) HLL-M32+ARD with angle
inputs being varied around radial distance 25 cm.
ACKNOWLEDGMENT
TABLE 1. Comparison of Quantitative Metrics Among the Conven- Prof. S. K. Pal acknowledges the National Science Chair, SERB-DST, GoI.
tional Method, HLL-SE, and HLL-M32+ARD in Estimating x and y

REFERENCES
[1] J. Yuan, H. Chen, F. Sun, and Y. Huang, “Multisensor information fusion for people
tracking with a mobile robot: A particle filtering approach,” IEEE Trans. Instrum.
Meas., vol. 64, no. 9, pp. 2427–2442, Sep. 2015.
[2] P. Paral, A. Chatterjee, and A. Rakshit, “Vision sensor-based shoe detection for
human tracking in a human–robot coexisting environment: A photometric invariant
approach using DBSCAN algorithm,” IEEE Sens. J., vol. 19, no. 12, pp. 4549–4559,
Jun. 2019.
[3] D. Li, L. Li, Y. Li, F. Yang, and X. Zuo, “A multi-type features method for leg
detection in 2-D laser range data,” IEEE Sens. J., vol. 18, no. 4, pp. 1675–1684,
Feb. 2018.
95% credible intervals for HLL-M32+ARD distribute in considerably [4] P. Paral, A. Chatterjee, and A. Rakshit, “Human position estimation based on filtered
narrower regions, indicating that it achieves much higher estimation sonar scan matching: A novel localization approach using DENCLUE,” IEEE Sens.
credibility and lower uncertainty compared to HLL-SE. This impres- J., vol. 21, no. 6, pp. 8055–8064, Mar. 2021.
sive performance of HLL-M32+ARD can be attributed to the strong [5] P. Paral, A. Chatterjee, and A. Rakshit, “Sonar sensing-based human leg localiza-
tion using Gaussian process regression,” in Advances in Intelligent Systems and
feature extraction capabilities of the ARD structure, in addition to Computing, K.-K. Giri et al., Eds. Berlin, Germany: Springer, 2022, vol. 1412,
the high robustness offered by the M32 kernel. The ARD kernel with pp. 1029–1041.
variable length-scales can effectively learn the relevance of the input [6] A. Burguera, Y. González, and G. Oliver, “On the use of likelihood fields to perform
features and subsequently remove irrelevant features by setting large sonar scan matching localization,” Auton. Robot., vol. 26, no. 4, pp. 203–222,
May 2009.
length-scales for them. This allows for the retention of only the input
[7] C. M. Bishop, Pattern Recognition and Machine Learning. 1st ed. Berlin, Germany:
features modeling the underlying structure of the observed data rather Springer, 2006.
than the noise so that overfitting is alleviated. [8] S. Lee and J. McBride, “Extended object tracking via positive and negative informa-
To study the effects of the variation of γr and γθ on LML (loga- tion fusion,” IEEE Trans. Signal Process., vol. 67, no. 7, pp. 1812–1823, Apr. 2019.
rithmic evidence) [11], surface plots are generated corresponding to [9] K. Liu, Y. Li, X. Hu, M. Lucu, and W. D. Widanage, “Gaussian process regression
with automatic relevance determination kernel for calendar aging prediction of
the processes of learning prediction models for x and y positions. The lithium-ion batteries,” IEEE Trans. Ind. Informat., vol. 16, no. 6, pp. 3767–3777,
corresponding plots, shown in Fig. 4(a) and (c), respectively, help to Jun. 2020.
provide independent verification of the optimization results obtained [10] Q.-H. Zhang and Y.-Q. Ni, “Improved most likely heteroscedastic Gaussian process
from L-BFGS algorithm. regression via Bayesian residual moment estimator,” IEEE Trans. Signal Process.,
vol. 68, pp. 3450–3460, May 2020.
[11] C. K. Williams and C. E. Rasmussen, Gaussian Processes for Machine Learning,
vol. 2, Cambridge, MA, USA: MIT Press, 2006.
V. CONCLUSION [12] J. Nocedal and S. J. Wright, Numerical Optimization, 2nd ed. New York, NY, USA:
Springer, 2006.
An advanced HLL model based on GPR is proposed that deals with [13] D. Makowski, M. S. Ben-Shachar, and D. Lüdecke, “bayestestR: Describing effects
a major issue of irrelevant input features associated with real-world and their uncertainty, existence and significance within the Bayesian framework,” J.
sonar data. Inappropriate input features can cause erroneous learning Open Source Softw., vol. 4, no. 44, Aug. 2019, Art. no. 1541.

Authorized licensed use limited to: JADAVPUR UNIVERSITY. Downloaded on January 20,2023 at 08:06:48 UTC from IEEE Xplore. Restrictions apply.