IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL.

71, 2022 2520412

Sonar-Based Human Leg Localization Using Chaos

Enhanced Dynamic Neighborhood Learning-Based
GSA Aided sNDT Algorithm
Pritam Paral , Amitava Chatterjee , Senior Member, IEEE, Anjan Rakshit ,
and Sankar K. Pal , Life Fellow, IEEE
Abstract— The present article proposes a new human leg local- CoBiDE Differential evolution based on covariance
ization (HLL) algorithm using ultrasonic sensors in human–robot matrix learning and bimodal distribution
coexisting environments. The algorithm estimates the motion
of a human leg pair between two successive sonar scans by
parameter setting.
using a new static cluster elimination (SCE) method, an edge DBSCAN Density-based spatial clustering of applica-
feature-based leg recognition algorithm, and an advanced scan tions with noise.
matching technique. We also propose a novel, robust approach DD-WMR Differentially driven wheeled mobile robot.
to overcome bad initialization problem in sonar scan matching, DE Differential evolution.
by introducing a metaheuristic search (MHS)-based optimization
algorithm for the sonar normal distributions transform (sNDT)
DFSABC_elite ABC algorithm with depth-first search
method. The recently proposed dynamic neighborhood learning- framework and elite-guided search
based GSA (DNLGSA) has been successfully utilized in real-life equation.
scenario to solve this problem. The work also proposes a new DNLGSA Dynamic neighborhood learning-based
chaos enhanced DNLGSA (CEDNLGSA) to further improve real- gravitational search algorithm.
life performance and the proposed novel variant of the sNDT
method based on CEDNLGSA, called chaotic MHS-based sNDT
EFBLR Edge feature based leg recognition.
(CMHS-sNDT), has been demonstrated to achieve superior leg EPSO Ensemble particle swarm optimizer.
detection performance in various real-life case studies, compared FE Function evaluation.
to different contemporary state-of-the-art methods. FOChS Fractional order chaotic system.
Index Terms— Chaos enhanced DNLGSA (CEDNLGSA), FOChS-PRNG FOChS-based pseudorandom number
dynamic neighborhood learning-based GSA (DNLGSA), leg generator.
detection, likelihood field (LF), scan matching, sonar sensing. FOVlS Fractional order Volta’s system.
FWS Forward straddle.
N OMENCLATURE GA Genetic algorithm.
ABC Artificial bee colony. G-L Grünwald–Letnikov.
ARM Advanced reduced instruction set computer GSA Gravitational search algorithm.
(RISC) machine. GWO Gray wolf optimizer.
BFO Bacterial foraging optimization. HLL Human leg localization.
CEDNLGSA Chaos enhanced dynamic neighborhood HLL-sNDT HLL with sonar normal distributions trans-
learning-based gravitational search form.
algorithm. HLL-X-sNDT HLL with X-based sonar normal distrib-
CMHS Chaotic metaheuristic search. utions transform (X: CMHS, DNLGSA,
CMHS-sNDT CMHS-based sonar normal distributions GWO, CoBiDE, DFSABC_elite, IRDPSO,
transform. and EPSO).
Manuscript received 6 August 2022; revised 18 September 2022; HRI Human–robot interaction.
accepted 12 October 2022. Date of publication 25 October 2022; date of IRDPSO Improved random drift particle swarm
current version 8 November 2022. The work of Pritam Paral was supported optimization.
by the Ministry of Electronics and IT, Government of India through the
“Visvesvaraya PhD Scheme for Electronics and IT”. The Associate Editor K-S Kolmogorov–Smirnov.
coordinating the review process was Dr. Huang-Chen Lee. (Corresponding LF Likelihood Field.
author: Pritam Paral.) LPDDR Low-power double data rate.
Pritam Paral and Sankar K. Pal are with the Center for Soft Computing
Research, Indian Statistical Institute (ISI), Kolkata 700108, India (e-mail: LRF Leg reading frame.
[email protected]; [email protected]). MAE Mean absolute error.
Amitava Chatterjee is with the Department of Electrical Engineering, MHS Metaheuristic search.
Jadavpur University, Kolkata 700032, India (e-mail: [email protected]).
Anjan Rakshit, retired, was with the Department of Electrical Engineer- MT-PRNG Mersenne Twister pseudorandom number
ing, Jadavpur University, Kolkata 700032, India (e-mail: anjanrakshit1951@ generator.
gmail.com). NDT Normal distributions transform.
This article has supplementary downloadable material available at
https://doi.org/10.1109/TIM.2022.3216846, provided by the authors. NWM Newton’s method.
Digital Object Identifier 10.1109/TIM.2022.3216846 PRN Pseudorandom number.
1557-9662 © 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: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
2520412 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

PRNG PRN generator. between exploration and exploitation. RDPSO [14], a novel
PSO Particle swarm optimization. efficient PSO variant, was proposed based on the free electron
RANSAC/GD Random sample consensus/Gaussian model in metal conductors exposed to an external electric field.
determination. Later, an IRDPSO [14] was presented to further improve the
RANSAC/GF Random sample consensus/Gaussian performance of RDPSO. EPSO [15] introduced an ensemble
filtering. of PSOs with self-adaptive mechanism. New algorithms like
RCF Reading coordinate frame. GWO [16] and WOA [17] are attracting significant attention
RDPSO Random drift PSO. too.
RMSE Root mean square error. The GSA [18], inspired by the law of gravity and mass
RPF Robot pose frame. interactions, became very popular in a very short time, show-
SCE Static cluster elimination. ing superior performance in solving various real-life complex
SDRAM Synchronous dynamic random-access optimization problems, such as [19], and is often adopted due
memory. to its several significant features [18].
SJM Short jump. 1) A particle in GSA is guided by the sum of gravitational
sNDT Sonar NDT. forces acting on it from all its neighbors stored in an
SRCC Spearman’s rank correlation coefficient. elite archive K best , rather than only one fittest particle in
TLA Two legs apart. algorithms like canonical PSO model. This endows GSA
TLT Two legs together or single leg. with a unique attribute, i.e., diverse search directions.
WOA Whale optimization algorithm. 2) The gravitational constant is reduced exponentially with
time to judiciously balance the exploration and exploita-
I. I NTRODUCTION tion and speed up the convergence process. Also, the
size of K best , with the initial value K 0 = P (P is the
O VER the past two decades, HRI has become a rapidly
emerging research subject [1], accomplishing objectives
like detecting and tracking human supervisors [2], [3], [4],
number of particles), decreases linearly over time and
thus, exploration gradually disappears and exploitation
[5], [6], [7], especially using range sensing [5], [6], [7]. The begins to appear. By adjusting the size of K best , the trade-
present work proposes a new range sensing-based approach off between both phases can be approved for GSA.
which estimates the motion of the legs of a walking person 3) Better particles have larger gravitational masses which
between two successive time instances, using scan match- cause higher intensities of attraction and, as a conse-
ing [8]. Although laser range sensors are known to provide quence, the particles tend to move toward the fittest
more precise solutions, sonar sensors still continue to offer particle. This permits a faster convergence. Conversely,
acceptable alternatives, especially from size, cost and power particles with heavy inertial mass move at a slower pace
consumption perspectives, which can be very useful in devel- and therefore search the problem space more locally and
oping countries. precisely, giving the flavor of an adaptive learning rate.
Among the scan matching algorithms, NDT [9], probabilis- 4) GSA is memory-less. However, it functions efficiently
tic laser scan registration [10] approach, and sNDT, a sonar- like the methods with memory. The seminal paper [18]
based variant of NDT [8], are quite popular. In sNDT, NWM demonstrates superior convergence rate of GSA over
is utilized to minimize a score function [8]. In NWM, the PSO and Real GA, two of the most popular MHS
score function should belong to class C1 (i.e., continuously algorithms.
differentiable) and computation of first- and second-order Various GSA variants, such as an improved GSA proposed
derivatives in each iteration involve high computational costs. in [20], adaptive gbest-guided GSA [21], etc., have shown
More importantly, NWM’s performance degrades if the initial improved performance vis-à-vis basic GSA. However, very
solution is far from the true minimum point. In this work, few studies have addressed the problems caused by the K best
we propose to overcome this bad initialization problem by model [22]. Recently, proposed DNLGSA [22] incorporated
using metaheuristic-based non-gradient type, global optimiza- the local and global neighborhood topologies to effectively
tion algorithms [11], suited for complex optimization prob- balance the exploration and exploitation and enhance the per-
lems. However, determining the global optimum remains a formance of GSA. The K best model is replaced in DNLGSA.
challenging job for MHS-based algorithms [11] too, which Moreover, our experimental studies reveal that DNLGSA is
requires a judicious balance of exploration and exploitation. significantly faster and more accurate than the GSA, when
Some of the popular MHS algorithms include GA, PSO, implemented for our real-life problem.
DE, ABC, BFO, and so on [11]. However, till date, the search However, in the practical engineering applications of meta-
for new, improved MHS algorithms remains a hot research heuristics, obtaining acceptable solutions in a very short time
topic. For example, Wang et al. [12] proposed CoBiDE, always remains a point of focus. To find an acceptable or
a novel variant of DE based on covariance matrix learning near-optimal solution fast and efficiently in a real-life problem,
and bimodal distribution parameter setting, which improved a MHS algorithm must possess a superior exploration capa-
the balance between local exploitation and global exploration bility. Population diversity is very crucial in MHS algorithms
abilities of DE. Cui et al. [13] introduced a novel ABC variant to enable efficient global exploration and to avoid abysmal
with depth-first search framework and elite-guided search performance because of premature convergence [22]. Now,
equations, called DFSABC_elite, to achieve a better tradeoff population diversity in DNLGSA gets noticeably influenced

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
PARAL et al.: SONAR-BASED HLL USING CEDNLGSA AIDED sNDT ALGORITHM 2520412

by the degrees of uniformity and correlation of the random convergence performance and search ability of GSA,
numbers involved in the calculation of the particle velocities. as well as maintaining a healthy population diversity.
If a random number generator which generates sequences of 2) We believe that this is the first study to demonstrate how
numbers showing superior statistical properties can be used, a global search strategy can be effectively utilized to
then the optimal performance of DNLGSA can be further perform sonar scan matching and also the first research
improved. The present work introduces an advanced variant work to show how a global optimization-based sonar
of DNLGSA called CEDNLGSA where the authors propose scan matching framework can be utilized to achieve
to utilize a new PRNG based on the state-time histories of accurate and robust human localization. Moreover, this
a FOChS [24], to generate random numbers for calculating is, to the best of our knowledge and belief, the first work
the particle velocities. Various statistical tests for distribution to demonstrate a successful real-life implementation of
and independence also confirm the superiority of the pro- a FOChS-based metaheuristic for solving sonar scan
posed FOChS-PRNG over one of the most extensively used matching-based human localization problems.
general purpose PRNG, namely MT-PRNG [23]. Fractional- Extensive experimentations and performance evaluations
order extensions of the integer-order chaotic systems demon- demonstrate superiority of the proposed method compared to
strate more abundant dynamic characteristics compared to the the basic NWM-based sNDT and other popular metaheuristics-
original integer order versions, due to their nonlocal features based variants of sNDT formulated in this work itself.
and significant nonlinear behavior [25], and the sequences of This article is organized as follows. Section II presents
PRNs produced with them show further improved degrees of the basic framework of the HLL-CMHS-sNDT approach.
uniformity and independence. Section III discusses the mechanism for the extraction of target
In this work, we propose an inexpensive, yet appealing new leg points, using the SCE method in conjunction with the
sonar-based solution for human localization in human–robot EFBLR algorithm. Sections IV and V describe CMHS-sNDT
coexisting environments that can be particularly effective for and CEDNLGSA, respectively. Real-life performances are
solving the bad initialization problem encountered by the evaluated in Section VI. Finally, the article is concluded in
traditional local optimizer-based sonar scan matching meth- Section VII.
ods. Our main aim is to accurately estimate the motion of
the leg-pair of a walking human between two successive II. BASIC F RAMEWORK OF HLL-CMHS- S NDT
sonar scans gathered by the onboard sonar arrangement of
In the HLL-CMHS-sNDT approach, for a specific robot
a mobile robot with respect to a specific RPF [26], during
pose, a couple of sonar scans are gathered corresponding to
people tracking. For this purpose, we first utilize density-
two distinct human leg poses at two successive time instants,
based clustering, followed by our proposed SCE method and
and hence both scans are referred to two different RCFs1 but
recently developed EFBLR algorithm, to filter the scans by
a single and unique RPF2 [7]. The first scan is called the raw
eliminating noises, discarding the readings indicating the static
reference scan and the second scan is called the raw current
and nonhuman moving objects, and preserving the readings
scan. Let a raw reference scan (denoted as Rϒrf ) be described
representing the target legs. Then we employ a novel vari-
by a set of Bs 2-D points Rϒrf = {nr1 , . . . , nr Bs } acquired
ant of sNDT based on CEDNLGSA, called CMHS-sNDT,
at a RCF Cr at a sampling instant s with respect to a specific
to determine the roto-translation of the target legs between the
RPF α and a raw current scan (denoted as Rϒcr ) by a set of
filtered scans, a robust solution that can specifically overcome
As 2-D points Rϒcr = {mr1 , . . . , mr As } at RCF Cc at the
bad initialization problems. This overall proposed approach
next sampling instant s + 1 with respect to the same RPF α.
for HLL is named here as HLL-CMHS-sNDT. This cycle
The goal of our HLL-CMHS-sNDT approach is to estimate
of gathering two successive sonar scans at a certain RPF,
the motion of the legs of a walking person between the RCFs
removing outliers, static, and nonhuman moving objects and
Cr = rcs and Cc = rcs+1 , with respect to a particular
extracting target leg-pair from each sonar scan, and then
RPF α = rp[s, s+1] , by using an advanced scan matching
accurately matching the two filtered scans using CMHS-
technique [7]. Firstly, we employ DBSCAN, a density-based
sNDT to estimate the human motion between them, can be
clustering method [27], separately on Rϒrf and Rϒcr , to form
repeatedly implemented for subsequent pairs of scans, thus
arbitrarily shaped significant clusters of the data points charac-
making the proposed approach suitable for following people in
terizing static and moving objects, including the target person’s
real-life.
legs and remove outliers. Then, the data points representing the
The significant contributions of this research work are as
moving objects are extracted using the SCE method, resulting
follows.
in refined Rϒrf and refined Rϒcr . Next, the EFBLR algorithm
1) In this work, a novel, robust approach is proposed is used to identify the target leg patterns in the refined Rϒrf and
to solve the bad initialization problem encountered in refined Rϒcr , and subsequently, extract the leg points from the
sonar scan matching, by embedding a MHS-based opti- corresponding scans. Accordingly, the refined Rϒrf and refined
mization mechanism in the sNDT method, a popular
1
LF-based sonar scan matching technique. In this context, The RCF rct represents the coordinate frame that is located at the set of
we develop a new MHS algorithm in the genre of range measurements constituting a scan acquired at an instant of time t by
the onboard ultrasonic sensing configuration.
GSA based on the state-time histories of a FOChS, 2 The RPF rp represents the coordinate frame that is located at the robot
t
named CEDNLGSA, which is aimed at improving the pose at an instant of time t [26].

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
2520412 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

Rϒcr are modified to reference and current scans, represented

by ϒrf = {n1 , . . . , n B } and ϒcr = {m1 , . . . , m A }, respectively.
Next, a novel scan matching algorithm based on the theory
of LF is proposed, named here CMHS-sNDT algorithm,
to determine the best estimate of the relative position of the
target person’s leg pair between the two LRFs [7] Fr = lr s
and Fc = lrs+1 , with respect to the RPF α. In the context
of scan matching here, we redefine the original RCF as
the LRF [7]. In the CMHS-sNDT algorithm, we first build
several overlapping sNDT grids having different positions of
the origin from the set of target leg points belonging to ϒrf .
For each cell of a specific grid containing at least a certain
number of points, a model which optimally fits the inliers
(and discards the outliers) of the set of data points contained
in the corresponding cell is discovered by making use of the
RANSAC/GD method. Then, with the help of RANSAC/GF
technique, we extract the points in ϒcr which render the
generative process assumption valid [8]. After constructing
the overlapping sNDT grids and filtering ϒcr , we perform the
scan matching process by using our proposed CEDNLGSA.
The corresponding scan matching estimate is indicated by
(α)
Fr zFc , represented as a multivariate Normal distribution
(α)
Fr zFc = N (Fr z(α) (α) (α)
Fc , Fr PFc ), where Fr zFc indicates the mean
vector of the form [dx , d y , φ]T (dx and d y symbolize the trans-
lations in the x- and y-directions, respectively, and φ stands
for the rotation), and Fr P(α)
Fc denotes the covariance matrix [7].
The scan matching problem here estimates the motion of the
leg pair of a target person Fr z(α)
Fc which maximizes the overlap
between ϒrf and ϒcr .
The proposed HLL-CMHS-sNDT approach can be utilized
in human–robot coexisting environments for estimating the
trajectory of a walking person using scans within an angular
span of 180◦ obtained from a sonar sensor mounted on a
wheeled mobile robot. Our proposed HLL algorithm finds
scan matching estimate between a pair of scans gathered at
successive sampling instants, and this algorithm is repeatedly
run for subsequent pairs of scans so that the mobile robot can
follow a walking person in real-life. At first, a pair of scans is
gathered for a specific robot pose at two consecutive sampling
instants, which are used by the proposed CMHS-sNDT to
estimate the motion of the target person’s leg-pair between
the two time instants. Next, the robot moves and another pair
of scans is acquired at successive sampling instants and again
the human motion between the corresponding time instants
is estimated. This sequence of events is then performed in
an iterative fashion. Inspired by our recent work in [7], the
Fig. 1. Flowchart of the proposed HLL-CMHS-sNDT approach for HLL
flowchart of the proposed HLL-CMHS-sNDT approach for and following.
HLL and following is presented in Fig. 1.

remaining points (basically characterizing static obstacles),

III. TARGET L EG P OINTS E XTRACTION
we propose a method called SCE method. Here, we compute
A. Moving Object Recognition With SCE Method the centroids of all the clusters associated with both the
For each raw reference scan Rϒrf and raw current scan scans Rϒrf and Rϒcr , and then evaluate the displacements
Rϒcr , the DBSCAN algorithm first clusters sonar data of the cluster centroids belonging to Rϒrf from their orig-
points using the notions of density-reachability and density- inal initial positions by using the locations of the cluster
connectivity, characterized by a distance threshold  > 0 and centroids belonging to Rϒcr . This facilitates detection of the
a density threshold MinPts [27]. Next, to extract the moving movement of the moving objects between Rϒrf and Rϒcr and
objects from both the noise-free scans and eliminate the identification of the points representing those objects in both

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
PARAL et al.: SONAR-BASED HLL USING CEDNLGSA AIDED sNDT ALGORITHM 2520412

target person, a walking cat, and several static obstacles are

present in the scanned environment.

B. Target Leg-Pair Identification With EFBLR Algorithm

Then to recognize the moving target leg-pair, we employ
our recently developed EFBLR algorithm [7] that detects the
target person’s legs in a sonar scan acquired during human
following within a leader-follower framework. The algorithm,
for specific leg poses, identifies typical leg patterns, which
are readily distinguishable from other objects in a real-life
human–robot coexisting environment. The refined Rϒrf and
refined Rϒcr comprising the readings of the moving clusters
(objects) are next input to the EFBLR algorithm. In this
algorithm, a set of range measurements Arg is maintained
as [7]
 
  r̂ψ j , if r̂ψ j = NaN
Arg = rg1 , . . . , rgFS , with rg j = 
Smax , if r̂ = NaN ψj

where the constant term FS denotes the maximum possible

number of readings in a scan in our work and r̂|ψ j represents
the range measurement in the direction ψ j . The maximum
sonar range Smax , in this case, is set at 70 cm.
From the set Arg of range readings, the identification of
different leg patterns can be performed efficiently by utilizing
a method based on the geometric features of vertical edges.
Details of the methodology for detecting and listing vertical
edges are given in [7]. Given vertical edges are classified
Fig. 2. (a) and (b) Data points in a representative Rϒrf and a representative into two categories: falling and rising edges, we elicit a
Rϒcr , respectively. (c) and (d) Significant clusters of the data points in subset from the complete list of connected edges, which
Rϒrf and Rϒcr , respectively, formed with DBSCAN and the corresponding
cluster centroids. A red contour: a cluster and a magenta dot: its centroid. might correspond to one of the four possible leg patterns:
(e) and (f) Data points representing the moving objects, extracted from Rϒrf TLA, SJM, FWS, and TLT. To best represent these human
and Rϒcr , respectively, using SCE method. Accordingly, Rϒrf and Rϒcr are leg patterns, four ordered sequences comprising falling/rising
transformed into the refined Rϒrf and refined Rϒcr , respectively.
edges are defined as: 1) TLA: a quadruplet {F , R, F , R};
2) SJM: a quadruplet {F , R, F , R}; 3) FWS: a triplet
the scans. However, due to uncertainties involved in sonar {F , F , R} or {F , R, R}; and 4) TLT: a doublet {F , R},
sensing, a cluster centroid belonging to Rϒrf which has not with F and R denoting a falling edge and a rising edge,
actually moved with respect to its original initial position, may respectively. It is worth mentioning that, considering the short
report spurious displacement in Rϒcr . A radius threshold r dth valid measurement distance of the sonar sensor, we have
signifying a tolerance level has been introduced to overcome introduced a new pattern into the original EFBLR framework,
this problem. namely SJM, to capture the forward jump motion of the target
Let Cnr = {Cnri }i=1
J
and Cnc = {Cnc j } Kj=1 be the sets of person because, on occasion, he/she may be required to jump
cluster centroids associated with Rϒrf and Rϒcr , respectively. once or twice while walking. The SJM pattern is based on
For a Cnri ∈ Cnr if there exists a Cnc j ∈ Cnc such that the consideration that when one leg is swinging within the
  vertical field of view of the onboard sonar arrangement, the
Cnri − Cnc j  ≤ r dth (1)
other one is in a stance state. Fig. 3 shows the four typical leg
then the pair {Cnri , Cnc j } is referred to as the static centroid patterns, where several dimensional constraints are specified
doublet and both the centroids Cnri and Cnc j are queued into for the measures wid, sin, stp(J/F), and tog, respectively,
a list Cstat . Once the tasks of detecting the static centroid dou- denoting the single leg-width, the spatial interval between
blets and inserting them into Cstat are completed, we populate two legs, the step length for SJM/FWS pattern, and the
the remaining centroids from both the scans into another list collective width of two legs. For the settings wid ∈ [10, 20]
Cmov and designate the group of the centroids belonging to cm, sin ∈ [7, 17] cm, stp(J ) ∈ [10, 20] cm, stp(F) ∈
Cmov as the moving centroid combination. Next, the clusters (20, 40] cm, and tog ∈ [10, 40] cm, the optimal results are
corresponding to the centroids in Cstat are removed and the obtained.
clusters corresponding to the centroids in Cmov (i.e., the mov- The EFBLR algorithm for identification of the target leg
ing objects) are preserved. The proposed SCE method, which pattern corresponding to a single sonar scan is detailed in [7].
outputs the refined Rϒrf and refined Rϒcr , is illustrated with The subroutines for the detection of SJM and TLA patterns
a representative real-life example in Fig. 2, where the walking are similar, except that instead of considering the geometric

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
2520412 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

Fig. 3. Typical human leg patterns identified by the EFBLR algorithm.

constraints of the TLA pattern, one needs to adopt those of the

SJM pattern. Now, once the leg pattern for a scan is recognized
using the EFBLR algorithm, the corresponding sequences of
F /R edges and edge-indices [7] can be utilized to retrieve
the set of data points representing the target leg-pair.
The EFBLR-based module for extracting the target leg
points, which outputs the reference scan (denoted as ϒrf ) and
current scan (denoted as ϒcr ), comprising B and A points,
respectively, is illustrated in Fig. 4, where the module takes
the refined Rϒrf and refined Rϒcr in Fig. 2 as its inputs.

IV. D EVELOPMENT OF THE CMHS- S NDT

The proposed CMHS-sNDT method presents an accurate
and robust LF-based solution to the sonar scan matching
problems, where the original NDT [9] is suitably adapted
to deal with the uncertainty and sparseness associated with
sonar readings [8]. CMHS-sNDT introduces a grid-building
Fig. 4. (a) and (b) Cartesian graphs representing the remodeled range-arrays
mechanism [8], where each cell is modeled with a Normal Arg s corresponding to the refined Rϒrf [given in Fig. 2(e)] and refined Rϒcr
distribution. Also, CMHS-sNDT introduces a scan-filtering [given in Fig. 2(f)], respectively (with range on the ordinate, and angle on
procedure [8] and proposes a new chaotic metaheuristic opti- the abscissa). (c) and (d) Target leg-pair patterns corresponding to the refined
Rϒrf and refined Rϒcr , respectively, identified by the EFBLR algorithm. The
mization framework to perform scan matching. blue patterns indicate the pairs of the target person’s legs. (e) and (f) Target leg
points extracted from Rϒrf [given in Fig. 2(a)] and Rϒcr [given in Fig. 2(b)],
respectively, based on the outputs of the EFBLR algorithm. Eventually, the
A. Constructing the CMHS-sNDT Grid by Using refined Rϒrf and refined Rϒcr are modified to ϒrf and ϒcr , respectively,
comprising the data points representing the moving target leg-pair.
RANSAC/GD
In CMHS-sNDT, the distribution of a sonar point cloud is
modeled by a grid of Normal distributions constructed from LF corresponding to the cell k is defined as [8]
  
n j =1,..., B ∈ ϒrf . The 2-D data space containing ϒrf is divided x − σ, k C−1
T
σ, k x − σ, k
into M square-shaped cells and the set of points n j ∈ ϒrf lying L σ, k (x) = exp − . (2)
2
within each cell is then searched. Based on the position of
the origin of the grid, the points belonging to ϒrf are divided Then the LF L σ (x) associated with the ϒrf and the sNDT
into different groups. Then, for a given cell k of a specific grid’s origin σ can be given as [8]
grid with origin σ consisting of no less than three points, ⎧
⎪
⎨ L σ, 1 (x), x ∈ σ, 1
we calculate the mean σ, k and the covariance matrix Cσ, k
of the set of points σ, k contained in that cell, required to L σ (x) = ··· (3)
⎪
⎩
obtain a bivariate Normal distribution N ( σ, k , Cσ, k ). For our L σ, M (x), x ∈ σ, M .
sonar-based approach, where a significant number of outliers Following the seminal NDT approach, we also use four
are usually present and due to these sensors’ low angular overlapping grids [9], to compute the LF at position x [8]
resolution, dense regions of outliers (called artifacts) tend to 
be formed, we use RANSAC/GD [8] to estimate σ, k and Cσ, k L(x) = L σ (x). (4)
of N ( σ, k , Cσ, k ) that optimally fits the inliers of σ, k . 1≤σ ≤4

Fig. 5 depicts the results of constructing the CMHS-sNDT

grid using the RANSAC/GD algorithm with various values
B. Defining the LF Associated With the CMHS-sNDT Grid of narrowness factor [8]. In fact, the LF gets narrower as
Next, we model the probability of having a sonar reading at the narrowness factor decreases and a narrow LF may be
2-D point x located inside the cell k by N ( σ, k , Cσ, k ). The problematic if bad initialization data is provided.

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
PARAL et al.: SONAR-BASED HLL USING CEDNLGSA AIDED sNDT ALGORITHM 2520412

Fig. 5. (a) Sample sonar scan containing noisy and spurious leg point data.
The results of building the CMHS-sNDT grid from the scan in Fig. 5(a) by
using the RANSAC/GD with various narrowness factors: four overlapping
grids with narrowness factor set to (b) 0.001 (usually used for laser counter- Fig. 6. (a) Set of readings in ϒrf [given in Fig. 4(e)] selected as inliers by
part) and (c) 0.5. RANSAC/GD (output: cyan leg points). (b) Filtering ϒcr [given in Fig. 4(f)]
with RANSAC/GF (output: magenta leg points). (c) Filtered sets of leg points
in ϒrf and ϒcr with respect to a specific RPF α, on which CMHS-sNDT is
implemented. The blue triangle represents the robot’s position. The resulting
C. Filtering Sonar Scans scan-matching estimate: Fr z(α)
Fc = [−2.87 cm, 15.21 cm, 0.59 rad] .
T

In the CMHS-sNDT approach, ϒrf is used to construct the

LF, whereas ϒcr is used as a point cloud. Hence, it is crucial
for ϒcr to posses the similar characteristics as ϒrf . Now, the for human leg points extraction and human leg motion estima-
inliers chosen from ϒrf by the RANSAC/GD algorithm are tion are presented in Section S-I of the “Supplementary File.”
used to construct the CMHS-sNDT grid. A similar approach,
referred to as RANSAC/GF [8] is employed for ϒcr which V. P ROPOSED CEDNLGSA
filters the sonar readings in ϒcr such that the generative
A. Chaos-Based PRN Generation for the Proposed CMHS
process assumption [8] holds valid. Hence ϒcr gets split into
Algorithm
two subsets: 1) the adopted readings comprise data fulfilling
the generative process assumption and 2) the rejected readings In order to generate sequences of PRNs for CEDNLGSA,
comprising the rest data [8]. we propose to use the state-time histories of a popular FOChS,
namely FOVlS, which are calculated by making use of the
well-established G-L definition [24] of fractional calculus.
D. Performing Sonar Scan Matching This is the first work, according to the best of our knowledge
Next, CMHS-sNDT computes the scan matching estimate and belief, to demonstrate a successful real-life implementa-
by minimizing a score function SF(z) with the CEDNLGSA. tion of these concepts for sonar scan matching-based human
The score function SF(z) is defined according to [8] as localization problems.
  1) Fractional Order Volta’s System: The state-time histories
SF(z) = − L σ (z ⊕ mi ) (5) of FOVlS have been utilized to develop our proposed chaos-
mi ∈ϒcr 1≤σ ≤4 based PRNG. The fractional order version of Volta’s system
is given as [24]
where z = [d x , d y , φ]T indicates a roto-translation between
β1
the LF and the ϒcr LRF, and the operator ⊕ indicates 0 Dt y1 (t) = 1 (y1 (t), y2 (t), y3 (t), t)
composition of transformations. The score function SF(z) = −y1 (t) − ay2 (t) − y3 (t)y2 (t)
evaluates the goodness of the roto-translation z = [dx , d y , φ]T β2
between ϒrf and ϒcr and when it achieves its mini- 0 Dt y2 (t) = 2 (y1 (t), y2 (t), y3 (t), t)
mum value, the CMHS-sNDT designates the correspond- = −y2 (t) − by1 (t) − y1 (t)y3 (t)
ing roto-translation as the scan matching estimate Fr z(α) Fc . D
0 t
β3
y 3 (t) = 3 (y1 (t), y2 (t), y3 (t), t)
In our proposed CEDNLGSA, we represent a roto-translation = cy3 (t) + y1 (t)y2 (t) + 1 (6)
with the position of a particle. Fig. 6 demonstrates the
CMHS-sNDT approach with a representative real-life case where 0 < β1 , β2 , β3 ≤ 2 are the fractional derivative orders
study. and (a, b, c) ∈ R3 are system parameters.
Because of space limitations, real-life case studies showing The numerical solution of the FOVlS (6) based on the
the successful implementation of our proposed methodologies G-L definition can be given by the following expression,

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
2520412 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

as described in [24]: TABLE I

     C OMPARISON OF THE M EAN R ESULTS OF VARIOUS S TATISTICAL T ESTS
yi t j = i y1 t j −1 , y2 t j −1 , y3 t j −1 , t j −1 h βi BETWEEN FOC H S-PRNG AND MT-PRNG. T HE C RITICAL VALUES
χκ(C)
2 , (C) , AND Zs(C) ARE C ONSIDERED AT A L EVEL OF S IG -

j
(βi )  NIFICANCE ξ = 0.05 [4]. W E FAIL TO R EJECT THE N ULL
− Xk yi t j −k (7) H YPOTHESIS H0 [4] IF χκ2 < χκ(C)2
,  < (C) , AND
k=1 −Zs(C) < Zs < Zs(C) , WHERE χκ(C) 2 = 233.99,
where i = 1, 2, 3; j = 1, 2, . . . , W , for W = Ts / h, Ts is (C) = 1.36
√ = 0.03512, AND Zs(C) = 1.96 [4]. H ERE ,
L
κ = 200 AND L = 1500 [4]
the simulation time and h is a significantly small step size;
[y1 (t0 ), y2 (t0 ), y3 (t0 )] are the initial conditions. The binomial
(β )
coefficients Xk i , ∀i , are computed as follows [24]:
 
(βi ) (βi ) 1 + βi (β )
X0 = 1, Xk = 1 − Xk−1i . (8)
k
A significant number of different combinations of (a, b, c)
and (β1 , β2 , β3 ) can be chosen experimentally, for which
the necessary stability condition for the general incommen- The particle k in the local neighborhood LN i at iteration
surate fractional order nonlinear system to exhibit chaotic τ [denoted as LN i (τ )] experiences gravitational forces from
behavior [24] is satisfied by FOVlS (6). For example, when the remaining particles of LN i (τ ), and its next local velocity
we consider the system parameters a = 5, b = 85, c = is computed as [22]
0.5 and the fractional derivative (incommensurate) orders  Ml (τ )  δ
β1 = 0.89, β2 = 1.10, β3 = 0.91 [24], FOVlS satisfies the L ϑkδ (τ + 1) = Y1 · G (τ ) z (τ ) − z kδ (τ )
Dkl (τ ) +  l
necessary stability condition to remain chaotic. zl ∈LN i (τ )
zl  =zk
2) Proposed Chaotic PRNG: We now propose a new
(10)
FOChS-PRNG, which employs the time histories of the state
variables y1 , y2 and y3 of FOVlS to generate three distinct where Y1 ∈ [0, 1] is a uniform PRN. G (τ ), the gravitational
sequences of PRNs, denoted as Y1 , Y2 and Y3 , by using the constant, is calculated as [18]
following equation:  
τ
        G (τ ) = G0 · exp −γ · (11)
Yi t j = {  yi t j  × 10η −  yi t j  × 10η  } TMAX
for i = 1, 2, 3; j = 0, 1, . . . , Ts / h; η ∈ {1, 2, 3} where G0 represents the initial value of the gravitational
(9) constant, γ is the attenuation coefficient, and TMAX is the
maximum number of iterations cycles.
where, Yi (t j ) indicates the PRNs generated with the 64-bit Ml (τ ), the mass of particle l at iteration τ is [18]
floating point values of yi (t j ). The PRNs belonging to an
array Yi , i ∈ {1, 2, 3} lie in the interval [0, 1], and are μl (τ )
Ml (τ ) =  P (12)
uniformly distributed throughout the interval with virtually p=1 μ p (τ )
zero correlation. We can arbitrarily select the initial conditions
for
for the purpose of generating the pseudorandom sequences.
The results in Table I demonstrate without vagueness that fitl (τ ) − worst(τ )
μl (τ ) = (13)
our FOChS-PRNG, for three benchmark statistical tests of best(τ ) − worst(τ )
distribution and independence, namely Chi-square test (χ 2 ), where fitl (τ ) is the fitness value of particle l at iteration τ .
K-S test and SRCC test [4], is consistently superior to For a minimization problem [18]
MT-PRNG, which encouraged us to formulate CEDNLGSA,
a novel chaos-based variant of DNLGSA, for the purpose of best(τ ) = min fit p (τ ) (14)
p∈{1,2,...,P}
solving our scan matching problem. worst(τ ) = max fit p (τ ). (15)
p∈{1,2,...,P}

B. Dynamic Neighborhood Learning Scheme Also, the scalar Dkl (τ ) indicates the Euclidean distance
Let Z = [z1 , . . . , z P ] represents the positions of P particles between two particles k and l at iteration τ and is represented
constituting a GSA population, where zk = [z k1 , . . . , z kd ], k = by: Dkl (τ ) = zk (τ ), zl (τ ) 2 ;  is a small positive constant;
1, . . . , P indicates the position of the kth particle moving z kδ (τ ) and z lδ (τ ) indicate the positions of the particles k and l
through a d-dimensional search space with the velocity ϑ k = at time τ in the δth direction, respectively.
[ϑk1 , . . . , ϑkd ]. Also, let the velocities of zk at iterations τ and The particle k also gets guidance from the historically
τ + 1 in the δth dimension be denoted by ϑkδ (τ ) (referred to as best experience of the entire population, denoted by gb =
current velocity) and ϑkδ (τ + 1) (referred to as next velocity), [gb1 , gb2 , . . . , gbd ]. Apparently, the gb model is analogous
respectively. In the dynamic neighborhood learning strategy, to a global neighborhood topology. For the particle k, its next
the entire population is first randomly divided into Q non- global velocity, is defined as [22]
overlapping local neighborhoods, LN = {LN 1 , . . . , LN Q }, 
G ϑkδ (τ + 1) = Y2 · gbδ (τ ) − z kδ (τ ) (16)
each comprising S = P/Q number of particles [22] (here,
we have set P = 50 and Q = 5). where Y2 ∈ [0, 1] is a uniform PRN.

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
PARAL et al.: SONAR-BASED HLL USING CEDNLGSA AIDED sNDT ALGORITHM 2520412

The next resultant velocity of the particle k is computed

as [22]
ϑkδ (τ + 1) = Y3 · ϑkδ (τ ) + a1 · L ϑkδ (τ + 1)
+ a2 · G ϑkδ (τ + 1) (17)
where Y3 ∈ [0, 1] is a uniform PRN, and two acceleration
coefficients, a1 and a2 , are arbitrary non-negative numbers.
If a1 > a2 , exploration dominates, while if a2 > a1 , exploita-
tion dominates. Inspired by the adaptive adjustment in [22],
1/6 1/6
we set a1 = 0.5 − 0.5τ 1/6 /TMAX and a2 = 1.5τ 1/6 /TMAX . The Fig. 7. Wheeled mobile robot employed for testing and implementing the
proposed HLL-CMHS-sNDT method.
position of particle k in the δth dimension is updated as [22]
z kδ (τ + 1) = z kδ (τ ) + ϑkδ (τ + 1). (18)
E VOLUTIONARY S TATE 1 (PDV > ν): Here, the particles
Here, Y1 , Y2 and Y3 are uniform PRNs corresponding to a need to concentrate more on exploration in order to prevent
value of the states y1 , y2 and y3 of the FOVlS, respectively. the potential hazards of population stagnation. So, we carry
The local neighborhood topology drastically reduces the out reformulation of the local neighborhoods and introduce
computational complexity while maintaining the search the mutation of gb with large jump steps, given as [22]
diversity of GSA. Also, the global neighborhood topology
is beneficial to accelerate the rate of convergence. There- nextgbδ (τ ) = gbδ (τ ) + PDV · U N , if PDV > ν (22)
fore, by proficiently combining both topologies, CEDNLGSA where U N is a uniform PRN with range [−1, 1], ensuring
strengthens the balance between exploration and exploitation. mutation in diverse directions.
E VOLUTIONARY S TATE 2 (PDV ≤ ν): Here CEDNLGSA
C. Framework of Dynamic Local Neighborhood Formation should concentrate on exploitation. We introduce a mutation
and gb Mutation scale ϕ ∈ (0, 1) to adjust the step size of the population
1) Convergence Criteria: The problem of particles getting diversity-based mutation mechanism as follows [22]:
trapped in local optima is addressed by using a dynamic nextgbδ (τ ) = gbδ (τ ) + ϕ · PDV · U N , if PDV ≤ ν.
local neighborhood formation strategy and a gb mutation (23)
mechanism based on two convergence criteria: the limiting
threshold and the population diversity [22]. Furthermore, with the aim of promoting the evolution of
If the gb cannot better the self-solution after a definite our MHS algorithm to the global optimum, we adopt the next
number of successive iterations (cind ), it may cause population global best particle (nextgb) if it is better than the preceding
stagnation and premature convergence. A limiting threshold one, otherwise we preserve the previous gb.
CTH is introduced to check the possibility of falling into
stagnation. Once the risks of stagnation are identified, dynamic VI. E XPERIMENTAL R ESULTS AND P ERFORMANCE
reformulation of the local neighborhoods and mutation of the E VALUATIONS
gb are carried out to overcome it, based on the population A. Experimental Setup
diversity indicator. For computing population diversity at itera-
tion τ [denoted as PDV(τ )], we first calculate the geographical Keeping the requirements in developing countries in mind,
clustering centroid of each local neighborhood at iteration τ a low-cost, lightweight system has been implemented on an
 indigenously developed DD-WMR, shown in Fig. 7. The
zi ∈LN j (τ ) zi onboard computer is a Raspberry Pi 3 (in short: RPi) model
CLN j (τ ) = . (19)
S B+ containing a 1.4 GHz 64-bit quad-core ARM Cortex-A53
Then the distribution diversity of the centroids is given as [22] processor along with 1 GB of LPDDR2 SDRAM [28]. The
robot is equipped with an RPi Camera Module v2 comprising
Q 
1  d  2 a Sony IMX219 8-Mpixel sensor [29] and an HC-SR04
PDV(τ ) = CLN δj (τ ) − CLN δ (τ ) (20) ultrasonic sensor [30] for acquiring real-world visual and
Q j =1 δ=1
range data, respectively. Our DD-WMR is driven by only two
where motors, with a third servo motor rotating the ultrasonic sensor
Q to collect sonar readings within a 180◦ wide angular range.
δ j =1 CLN δj (τ )
CLN (τ ) = . (21) Our DD-WMR is built with ultrasonic and vision sensors
Q with an aim of developing a human detection and tracking
2) Dynamic Reorganization of Local Neighborhoods and framework in which the sensors will work in both independent
Mutation of gb: If cind > CTH , we compute population and cooperative manners so that the robot can either use
diversity and compare it with a threshold of population a single sensor or a combination of both [5]. With such
diversity ν which is employed to distinguish between explo- a provision, the system becomes potentially robust to the
ration and exploitation [22]. Based on the result, Evolutionary failure of any one sensing unit. In this research work, the
State 1 or 2 is executed. ultrasonic ranging module HC-SR04 has been independently

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
2520412 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

used to perform localization of a person walking in front of our

DD-WMR. The economical HC-SR04 ultrasonic sensor [30]
whose weight is 10 g (approximately), reads, in practice,
from 2 to 80 cm within a 30◦ cone, and is accurate to the
nearest 0.3 cm. The operating voltage and working current
of the sensor are +5 V dc and 15 mA (max), respectively
(hence can be directly powered via the RPi’s 5 V output). For
the purpose of real-life implementation, the whole software Fig. 8. Robustness of the competing HLL approaches: (a) HLL-sNDT,
(b) HLL-CoBiDE-sNDT, and (c) HLL-Z-sNDT (Z: CMHS, EPSO, GWO,
has been written in Python in combination with Cython. IRDPSO, DFSABC_elite, and DNLGSA).
Performance evaluations are carried out with the NumPy and
SciPy libraries in Python, whereas the experimental results
the random errors lie between −0.25 and 0.25 m in dx
are plotted with the Python library Matplotlib.
and d y , and between −45◦ and 45◦ in φ. These introduced
Our HLL-CMHS-sNDT approach is quantitatively com-
errors correspond to the values assigned to the initial scan-
pared with the HLL approach using the original NWM-based
matching estimate in Newton’s algorithm. On the other hand,
sNDT (referred to as HLL-sNDT) and the human localiza-
in the case of a single run of an MHS algorithm, randomly
tion approaches using six popular contemporary MHS algo-
initializing the positions of P number of particles in a specific
rithms, namely DNLGSA [22], GWO [16], CoBiDE [12],
search space of dx , d y and φ can be viewed as analogous to
DFSABC_elite [13], IRDPSO [14], and EPSO [15] based
introducing initial location errors lying within the same search
variants of sNDT formulated by us in the context of this article
space. To alleviate stochastic errors, the procedure is repeated
(referred to as HLL-Y-sNDT where Y is DNLGSA, GWO,
50 times for each experiment and scan, thus constituting
CoBiDE, DFSABC_elite, IRDPSO, and EPSO, respectively).
5 × 4 × 50 = 1000 trials per algorithm and a total of
The values of the parameters used for the competing MHS
8 × 1000 = 8000 runs.
algorithms are the same as those reported in the corresponding
seminal papers. The experiments discussed in this section are
performed using real-world sonar data sets gathered in four B. Scan Matching Performance Comparisons
diverse real-life environments: 1) a corridor with multiple We first compare, in statistical terms, the performances of
door frames and smooth concrete walls; 2) an unstructured the competing optimization algorithms in optimizing the score
room incorporating tables, chairs, and cardboard boxes; 3) an function SF(z) (5) (for details, see Section S-II of the “Sup-
enclosed space comprising a combination of concrete walls, plementary File”). Next, we evaluate and compare the scan-
concrete structures, wooden walls and wooden doors; and matching performances of the HLL approaches based on
4) an enclosed space containing stone walls, glass doors, and those optimizers, in terms of robustness, accuracy, speed
steel furniture. The acquired data sets thus involve structured of convergence, and computation time [8]. Here, the term
and unstructured regions, posing various challenging difficul- convergence refers to the ability of a method to satisfy
ties to our sonar sensing system. In each environment, two numerical convergence criteria in a certain limited number of
raw sonar scans Rϒ1 and Rϒ2 are acquired for a specific leg FEs. Considering real-life implementation, the computational
posture at two successive sampling instants t and t +1, respec- budget for all the HLL methods has been limited to 1000 FEs.
tively, with respect to a certain RPF α0 , and a third scan Rϒ3 is 1) Robustness: The robustness results for the competing
acquired for a different leg posture at the subsequent sampling HLL approaches are reported in Fig. 8. A trial is considered
instant t + 2 with respect to the same RPF α0 . Then, the successful and a correct estimate is obtained if the absolute
scans Rϒ1 and Rϒ2 , based on the SCE method and EFBLR error is less than 0.075 m in dx , 0.075 m in d y and 0.075 rad
algorithm, were transformed to ϒ1 and ϒ2 , respectively, with in φ, and these threshold values are used to differentiate
the help of Rϒ3 . Moreover, the set of readings in ϒ1 selected among true positives, true negatives, false positives, and false
as inliers by RANSAC/GD is adopted and ϒ2 is filtered with negatives [26]. Fig. 8 shows that the MHS-based HLL methods
RANSAC/GF. Clearly, the roto-translation between ϒ1 and outperform HLL-sNDT and achieve superior robustness in
ϒ2 is perfectly known to be [0 cm, 0 cm, 0 rad]T [8]. That is the face of variations in initial error in different experiments
to say, the ground truth is available. Note that, although the conducted. As we can see, all the MHS-based HLL methods
scans Rϒ1 and Rϒ2 are gathered in similar real situations, except HLL-CoBiDE-sNDT achieve 100% of true positives
ϒ1 and ϒ2 are not necessarily identical, thus setting up a even in the most challenging situation of Experiment 5.
realistic test bench for the sonar scan matching methods [8]. 2) Accuracy: For measuring accuracy, we have considered
Furthermore, by considering two scans acquired for the same the MAE of the positive estimates with respect to the ground
human pose with respect to a specific RPF, the experiments truth roto-translation. By concentrating on the positive esti-
essentially focus on the matching capabilities of the methods. mates, the experiment can assess the accuracy in real-life
A group of five experiments has been carried out in each scenarios where ground truth may not be available and the
environment, incorporating various initial location errors. The method only can differentiate between positives and negatives.
error values are randomly chosen according to uniform dis- For this test, we have selected the angular error over the trans-
tributions ranging from −0.05 to 0.05 m in d x and d y , and lational errors because the accurate computation of the former
from −9◦ to 9◦ in φ in Experiment 1. The initial location is considered a more critical component in the context of
error rises as we gradually move toward Experiment 5, where localization. As can be seen from Table II, HLL-CMHS-sNDT

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
PARAL et al.: SONAR-BASED HLL USING CEDNLGSA AIDED sNDT ALGORITHM 2520412

TABLE II
C OMPARISONS OF THE AVERAGE S CAN M ATCHING P ERFORMANCES OF THE C OMPETING HLL A PPROACHES

consistently achieves the best MAE performance between the TABLE III
rotations in the positive estimates and the ground truth rotation A SSESSMENT OF THE H UMAN F OLLOWING P ERFORMANCE OF
for all the experiments conducted. HLL-CMHS- S NDT
3) Speed of Convergence and Computation Time: Here,
the speed of convergence of a scan matching method is
defined as the number of FEs until convergence in the case
of true positive estimates. Table II summarizes the con-
vergence speed and computation time results for the eight
competing HLL approaches. It can be seen that, overall, our
HLL-CMHS-sNDT only came second best in terms of both
the convergence rate (in terms of FEs) and the computation
times, where HLL-sNDT, a non-population-based algorithm,
came first in both cases. However, in terms of robustness and
accuracy, HLL-sNDT produced the worst result, as evident
from Table II. Note that, in terms of convergence speed,
some MHS-based HLL methods show better or compara-
ble performances than CEDNLGSA-based approach when
the initial error is low, primarily because of the variation
in their population sizes. However, HLL-CMHS-sNDT out-
performs those methods as the error increases. Similarly,
although HLL-CMHS-sNDT requires fewer number of FEs
than HLL-DNLGSA-sNDT as the initial error goes up,
its computational time does not improve too much than
HLL-DNLGSA-sNDT due to the involvement of chaos-
based computations. However, HLL-CMHS-sNDT outper-
forms HLL-DNLGSA-sNDT in terms of accuracy in all the
experiments, owing to the fact that the former one with its
better exploration ability finds the promising region more early
and then efficiently exploit the knowledge about the region to
find the solution until exhaustion of the computational budget.
Hence, overall, the supremacy of HLL-CMHS-sNDT is firmly
established compared to other competing approaches, taking Fig. 9. Visual maps obtained during people following in four various real-
all performance indices into consideration. life environments with our proposed HLL-CMHS-sNDT method: (a) Map1,
(b) Map2, (c) Map3, and (d) Map 4. The gray areas indicate the spaces taken
C. Human Following Performance Evaluation up by the stationary objects.

We also perform four experiments involving real motion in

real scenarios to qualitatively evaluate the human following Results in Fig. 9 show that HLL-CMHS-sNDT is able to
performance of HLL-CMHS-sNDT, by visual observation. yield the human following results very close to the ground

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.
2520412 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 71, 2022

truth maps. The corresponding quantitative results, obtained [8] A. Burguera, Y. González, and G. Oliver, “On the use of likelihood fields
considering a corner in each experimental environment as the to perform sonar scan matching localization,” Auto. Robots, vol. 26,
no. 4, pp. 203–222, May 2009.
reference point with coordinates (0, 0), is given in Table III, [9] P. Biber and W. Straßer, “The normal distributions transform: A new
which clearly demonstrates that the RMSEs and MAEs (in approach to laser scan matching,” in Proc. IEEE/RSJ Int. Conf. Intell.
x-, y-directions), obtained between the points of the ground Robots Syst. (IROS), Las Vegas, NV, USA, Oct. 2003, pp. 2743–2748.
[10] D. Hähnell, W. Burgard, and S. Thrun, “Learning compact 3D models
truth path and the path of the robot, are considerably small in of indoor and outdoor environments with a mobile robot,” Robot. Auton.
comparison to the path lengths traversed. Syst., vol. 44, no. 1, pp. 15–27, Jul. 2003.
[11] I. Boussaïd, J. Lepagnot, and P. Siarry, “A survey on optimization
VII. C ONCLUSION metaheuristics,” Inf. Sci., vol. 237, pp. 82–117, Jul. 2013.
[12] Y. Wang, H.-X. Li, T. Huang, and L. Li, “Differential evolution based on
This work showed how metaheuristic-based search covariance matrix learning and bimodal distribution parameter setting,”
algorithms can be successfully utilized for solving sonar Appl. Soft. Comput., vol. 18, pp. 232–247, May 2014.
[13] L. Cui et al., “A novel artificial bee colony algorithm with depth-
sensor-based HLL problems in human–robot coexisting first search framework and elite-guided search equation,” Inf. Sci.,
environments. This approach is particularly useful in vols. 367–368, pp. 1012–1044, Nov. 2016.
solving sonar scan matching algorithms suffering from [14] W. T. Elsayed, Y. G. Hegazy, M. S. El-bages, and F. M. Bendary,
“Improved random drift particle swarm optimization with self-adaptive
bad initialization problems. The work also showed a mechanism for solving the power economic dispatch problem,” IEEE
successful real-life implementation of recently proposed Trans. Ind. Informat., vol. 13, no. 3, pp. 1017–1026, Jun. 2017.
DNLGSA for this purpose and proposed a new CEDNLGSA [15] N. Lynn and P. N. Suganthan, “Ensemble particle swarm optimizer,”
Appl. Soft Comput., vol. 55, pp. 533–548, Jun. 2017.
to achieve further improved performances. This novel [16] S. Mirjalili, S. M. Mirjalili, and A. Lewis, “Grey wolf optimizer,” Adv.
HLL-CMHS-sNDT approach, when implemented in several Eng. Softw., vol. 69, pp. 46–61, Mar. 2014.
real-life case studies, could, overall, sufficiently outperform [17] S. Mirjalili and A. Lewis, “The whale optimization algorithm,” Adv.
Eng. Softw., vol. 95, pp. 51–67, May 2016.
various other state-of-the-art algorithms. [18] E. Rashedi, H. Nezamabadi-Pour, and S. Saryazdi, “GSA: A gravita-
tional search algorithm,” J. Inf. Sci., vol. 179, no. 13, pp. 2232–2248,
Jun. 2009.
ACKNOWLEDGMENT [19] S. Duman, U. Güvenç, Y. Sönmez, and N. Yörükeren, “Optimal power
flow using gravitational search algorithm,” Energ. Convers. Manage.,
Prof. Sankar K. Pal acknowledges the National Science vol. 59, no. 59, pp. 86–95, Jul. 2012.
Chair, Science and Engineering Research Board, Department [20] S. Jiang, Y. Wang, and Z. Ji, “Convergence analysis and performance of
of Science and Technology (SERB-DST), Government of an improved gravitational search algorithm,” Appl. Soft Comput., vol. 24,
pp. 363–384, Nov. 2014.
India. [21] S. Mirjalili and A. Lewis, “Adaptive gbest-guided gravitational search
algorithm,” Neural Comput. Appl. vol. 25, nos. 7–8, pp. 1569–1584,
R EFERENCES Jun. 2014.
[22] A. Zhang, G. Sun, J. Ren, X. Li, Z. Wang, and X. Jia, “A dynamic neigh-
[1] K. Kosuge and Y. Hirata, “Human–robot interaction,” in Proc. IEEE Int. borhood learning-based gravitational search algorithm,” IEEE Trans.
Conf. Robot. Biomim., Shenyang, China, Aug. 2004, pp. 8–11. Cybern., vol. 48, no. 1, pp. 436–447, Jan. 2018.
[2] N. Najmaei, M. R. Kermani, and M. A. Al-Lawati, “A new sensory [23] Python 2.7.16 Documentation Webpage. Accessed: Sep. 30, 2019.
system for modeling and tracking humans within industrial work [Online]. Available: https://docs.python.org/2.7/
cells,” IEEE Trans. Instrum. Meas., vol. 60, no. 4, pp. 1227–1236, [24] I. Petráš, Fractional-Order Nonlinear Systems: Modeling, Analysis and
Apr. 2011. Simulation. Berlin, Germany: Springer-Verlag, 2011, pp. 7–184.
[3] P. Paral, A. Chatterjee, and A. Rakshit, “OPTICS-based template [25] H. Y. Jia, Z. Q. Chen, and G. Y. Qi, “Chaotic characteristics analysis
matching for vision sensor-based shoe detection in human–robot coex- and circuit implementation for a fractional-order system,” IEEE Trans.
isting environments,” IEEE Trans. Instrum. Meas., vol. 68, no. 11, Circuits Syst. I, Reg. Papers, vol. 61, no. 3, pp. 845–853, Mar. 2014.
pp. 4276–4284, Nov. 2019. [26] A. Burguera, Y. González, and G. Oliver, “A probabilistic framework
[4] P. Paral, A. Chatterjee, and A. Rakshit, “Chaos-based random sam- for sonar scan matching localization,” Adv. Robot., vol. 22, no. 11,
pling for photometric invariant shoe detection with vision sensor in pp. 1223–1241, Jan. 2008.
human–robot coexisting environments,” IEEE Sensors J., vol. 20, no. 10, [27] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, “A density-based algorithm
pp. 5424–5434, May 2020. for discovering clusters in large spatial databases with noise,” in Proc.
[5] J. Yuan, H. Chen, F. Sun, and Y. Huang, “Multisensor information 2nd Int. Conf. Knowl. Discovery Data Mining, Portland, OR, USA, 1996,
fusion for people tracking with a mobile robot: A particle filtering pp. 226–231.
approach,” IEEE Trans. Instrum. Meas., vol. 64, no. 9, pp. 2427–2442, [28] Raspberry Pi 3 Model B+ Webpage. Accessed: Mar. 10, 2021.
Sep. 2015. [Online]. Available: https://www.raspberrypi.org/products/raspberry-pi-
[6] P. Vadakkepat and L. Jing, “Improved particle filter in sensor fusion for 3-model-b-plus/
tracking randomly moving object,” IEEE Trans. Instrum. Meas., vol. 55, [29] Camera Module—Raspberry Pi Documentation Webpage. Accessed:
no. 5, pp. 1823–1832, Oct. 2006. Mar. 21, 2021. [Online]. Available: https://www.raspberrypi.org/d
[7] P. Paral, A. Chatterjee, and A. Rakshit, “Human position estimation ocumentation/hardware/camera/
based on filtered sonar scan matching: A novel localization approach [30] E. J. Morgan. (Nov. 16, 2014). HC-SR04 Ultrasonic Sensor.
using DENCLUE,” IEEE Sensors J., vol. 21, no. 6, pp. 8055–8064, Accessed: Mar. 15, 2021. [Online]. Available: https://datasheetspdf.co
Mar. 2021. m/pdf/1380136/ETC/HC-SR04/1

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 09,2022 at 09:35:29 UTC from IEEE Xplore. Restrictions apply.