Feature Article: DOI. No. 10.1109/MAES.2021.

3138948

SimHumalator: An Open-Source End-to-End Radar

Simulator for Human Activity Recognition
Shelly Vishwakarma , Wenda Li , Chong Tang , and Karl Woodbridge ,
University College London, London WC1E 6BT, U.K.
Raviraj Adve , University of Toronto, Toronto, ON M5S 3G4, Canada
Kevin Chetty , University College London, London WC1E 6BT, U.K.

rescue buttons to handle emergencies at home or out-

BACKGROUND
side [2]. However, the usage of wearable sensors and
Recent years have witnessed significant growth in the propor- expensive camera systems is limited because they can oper-
tion of the aging population and the number of people with ate only under suitable environmental conditions and are
long-term disability and chronic health conditions. Accord- affected by perceived privacy risks, comfort levels, and
ing to World Population Prospects 2019, it is expected that battery life. Moreover, the cameras can work only in the
by 2050, there will be one person in every six over the age of line-of-sight conditions and thus have limited coverage [3].
65, which is 16% higher compared to 2015 [1]. The growth X-ray signals possess good wall penetration capability.
in these numbers has already distressed the existing health- However, the technology is bulky and expensive, and the
care systems worldwide due to increased demand for dedi- radiation can potentially harm the patient. Fortunately,
cated residential care services that require continuous close radio frequency (RF) sensing, being noninvasive, can ele-
monitoring of the patients and older people. Therefore, there vate the need for in-person monitoring services. It can
is an urgent need for technologies that enable healthy living work even in nonline-of-sight scenarios (mainly through-
to ensure the health, safety, and well-being of people of all the-wall scenarios), making it a safe alternative. RF sensing
age groups, with a particular focus on supporting people liv- is low-cost, noninvasive, privacy-preserving, and can con-
ing independently in their own homes. tinuously monitor human activities remotely using com-
There is considerable interest in developing cost-effec- mercial off-the-shelf components [4], [5].
tive and reliable technologies enabling healthy living. The Humans are rarely still. When illuminated by an RF
technologies are specifically tailored for the elderly popula- source, the human motions introduce a Doppler shift to
tion, catering to their special needs of noninvasive continu- the radar signals. In addition, micromotions due to the
ous monitoring while offering them complete freedom to motion of hands and legs give rise to additional frequency
live a fulfilling life. The increased research interest in these variations known as micro-Doppler. The micro-Doppler
innovative technologies is evoked by their likely contribu- possesses unique and discriminative features when
tion to the efficiency and effectiveness of future healthcare. observed in a joint time–frequency space [6]. Extensive
The traditional sensing technologies use intrusive wearable research has exploited these signatures for indoor health-
sensors such as wrist-worn units acting as emergency care applications, such as assisted living (for fall detection
in older adults) [5],[7], biomedical applications (for nonin-
trusively monitoring patients), and detecting the debilitat-
Author current address: Shelly Vishwakarma, Wenda Li, ing motor symptoms of Parkinson’s patients [8].
Chong Tang, Karl Woodbridge, and Kevin Chetty are Micro-Doppler data have been generally gathered
with the University College London, London WC1E 6BT, through actual measurements using active and passive RF
U.K. (e-mail: [email protected]; [email protected].
sensors [6], [9], [10]. Active sensing uses the system’s trans-
uk; [email protected]; [email protected];
missions as the source of illumination, whereas passive sens-
[email protected]); Raviraj Adve is with the University
of Toronto, Toronto, ON M5S 3G4, Canada (e-mail: ing leverages existing opportunistic signals such as WiFi (or
[email protected]) other wireless) communications to capture RF returns pas-
Manuscript received May 25, 2021, revised November sively. Due to the number of advantages, such as its receive-
29, 2021; accepted December 23, 2021, and ready for only nature, low energy consumption, and presence of multi-
publication December 25, 2021. ple illuminators, the passive sensing systems have become
Review handled by Christof Schupbach. increasingly popular and are extensively used for short-range
0885-8985/21/$26.00 ß 2021 IEEE

6 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 Image licensed by Ingram Publishing

indoor monitoring applications [11]–[13]. The essential func- available: 1) marker-based and 2) marker-less. The real
tionality in all these works has been classifying human activi- advantage of using marker-based technology is capturing
ties based on their micro-Doppler signatures. More recently, more accurate, more realistic, and complex human motions.
there has been spurred growth in the usage of deep learning- Several markers are placed on the live actor’s body parts,
assisted solutions in radar signal processing due to the greater such as the head, torso, arms, and legs, to capture their 3D
availability of memory capacity and ever-increasing process- time-varying positions in space in a marker-based motion
ing speeds of computers [14]–[16]. The performances of capture system. Although the marker-based system is accu-
these algorithms are generally tied to a large amount of high- rate, the hardware restriction limits it to a single laboratory
quality training data. However, the volumes of data captured environment. In addition, wearing a bodysuit fitted with LED
are often limited and unbalanced due to the following rea- markers further increases the overhead in the data collection
sons. First, collecting real-world micro-Doppler data can be process, limiting the massive collection of data and various
laborious and costly. Second, various environmental condi- investigation scenarios. Ram and Ling [20] first presented a
tions, sensor parameters, and target characteristics affect the complete end-to-end active radar simulator of humans using
performance of the data, affecting deep learning algorithms’ a marker-based motion capture technique. The radar scatter-
performance. Therefore, it becomes necessary to simulate ings were simulated by integrating the animation data of
radar returns in indoor sensing scenarios that would generate humans with primitive shapes-based electromagnetic model-
large volumes of training data. The simulation data can be ing. Alternatively, Erol and Gurbuz [21] and Singh et al. [22]
used to preliminary evaluate different algorithms and study gathered animation data using a marker-less motion capture
the effects of radar phenomenology. It can provide a form of technology based on Microsoft’s Kinect. The Kinect sensors
virtual prototyping and move in stages toward a real system. are relatively compact, easy to carry, and set up to capture
The benefits of this incremental “divide and conquer” data compared to the marker-based system. Freely available
approach include the fact that the performance can be investi- databases of motion capture data from CMU, University of
gated safely for a range of operating conditions. Pennsylvania, and Ohio State are available at [23]–[25].
There exist multiple methods to simulate human micro- While previous works presented an active radar simulator
Doppler data. The earliest method models the human leg as a for generating target returns, no simulation tool exists for
double pendulum structure [17]. However, this model does generating radar returns in passive WiFi scenarios. Moreover,
not simulate radar returns from other human body parts, such none of the simulators presented in the prior works are pub-
as the torso and arms, contributing significantly to the micro- licly available. Therefore, there are no means of generating
Doppler returns. The second method uses a human walking synthetic databases that can augment otherwise limited mea-
model derived from biomechanical experiments [18]. Here, surement data and address the cold-start problem in radars. In
12 analytical expressions govern the motion trajectories of response to open science practices accelerating, improving,
17 reference points on the human body as a function of the and accumulating scientific knowledge for others to reuse
human’s height and relative velocity. Finally, Manfredi et al. and build upon, we have publicly released the animation
[19] presented a similar analytical model-driven simulation data-driven simulation tool capable of modeling the human
tool to characterize the radar cross-section of a pedestrian in radar signatures in passive WiFi radar (PWR) scenarios. It
a near field. However, this approach is based on a constant will assist the sensing communities in generating large vol-
velocity model. Therefore, it cannot capture variations in umes of high-quality and diverse radar datasets and bench-
more complex motions, such as falling, sitting, and jumping. mark future algorithms. The simulator’s development will
The third technique uses animation data from motion reduce the labor and expense of field testing by imitating a
capture systems to model more realistic and complex human real-world system under different operating conditions, such
motions. There are two types of motion capture technology as environmental conditions, sensor parameters, and human

MARCH 2022 IEEE A&E SYSTEMS MAGAZINE 7

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 SimHumalator: An Open-Source End-to-End Radar Simulator for Human Activity Recognition

characteristics. More importantly, the human micro-Doppler comprising high-quality micro-Doppler spectro-
data generated using the simulator can be used to augment grams for activity recognition applications.
limited experimental data. The simulator’s reliability is
7) In addition, the study demonstrates that the synthe-
improved by validating it with experimental data gathered
sized signatures can be used for data augmentation
using an in-house-built hardware prototype. The standalone
purposes to solve the practical problem associated
executable file for the simulator is available at https://uwsl.
with insufficient or unbalanced micro-Doppler training
co.uk/.
data.
To summarize, our contributions in this article are the
following.

1) We release SimHumalator as a standalone app pub-

licly available at https://uwsl.co.uk/. SimHumalator SIMHUMALATOR
uses a highly modular radar simulation framework. This section presents a user-friendly, publicly available
It allows users to select different target parameters, end-to-end simulation tool, SimHumalator, for generating
radar parameters, and signal processing parameters. human micro-Doppler radar signatures in WiFi-based pas-
The distribution package of the SimHumalator con- sive radar scenarios. The radar scatterings are simulated
sists of executable files and documentation files by integrating the animation data of humans with IEEE
comprising a quick-start guide and reference guide. 802.11-complaint WiFi transmissions to capture features
2) While previous works present an active radar simu- that incorporate the diversity of human motion character-
lator to generate target returns, SimHumalator gen- istics and the sensor parameters.
erates target returns in PWR scenarios. The SimHumalator has a highly modular radar simulation
simulator uses animation data from a human subject framework. It allows users the freedom to select different tar-
for various activities and combines it with IEEE get parameters, radar parameters, and signal processing
802.11 Standard WiFi transmissions to simulate parameters, as shown in Figure 1. We used MATLAB’s app
reflected signals. Passive WiFi sensing uses cross- designer to create SimHumalator’s standalone application.1
correlation-based processing to determine range and MATLAB allows users to create a standalone desktop appli-
Doppler information from target reflections. cation and share it royalty-free with other users that do
not have MATLAB installed on their systems. The
3) SimHumalator can generate micro-Doppler spectro- SimHumalator’s package includes an executable file that
grams of a human corresponding to three standard installs the packaged application, files including the anima-
waveforms—IEEE 802.11 g Standard at 2.4 GHz, tion data files that enable the use of the application, and the
IEEE 802.11ax Standard at 5.8 GHz, and IEEE documentation files comprising a user’s quick-start guide
802.11ad Standard at 60 GHz. The sub-6 GHz WiFi and reference guide. The app can be installed on machines
signals (IEEE 802.11a, b, and n) can be used to with or without MATLAB installed. It downloads the MAT-
track and monitor indoor occupants for either sur- LAB run-time compiler from the Internet if it is not included
veillance or healthcare purposes. On the other hand, in the installer at the packaging time (for users who do not
IEEE 802.11ad wireless transmissions at 60 GHz have MATLAB installed on their machines). We restrained
can be used to sense fine-grained motions. ourselves from including the run-time installer in the same
4) The signatures generated by SimHumalator are package as the total file size became too large. Instead, we
experimentally validated using an in-house-built choose to generate an installer that downloads the MATLAB
hardware prototype. run-time from the web and installs it along with the deployed
MATLAB application. We urge users always to install the
5) Our latest development on the simulator allows users to MATLAB run-time compiler when prompted in the app
generate a massive signature library of human micro- installation process unless they have the same version of the
Doppler signatures across different target motion char- MATLAB installed on their machines. The distribution pack-
acteristics (aspects of the target to the radar, initial posi- age of the SimHumalator release can be downloaded from
tion in space, and different type of motions), sensor https://uwsl.co.uk/. The current release of the simulator can
parameters (different PWR radar configurations— synthesize human radar returns for two types of radar sensors:
monostatic, bistatic-in line, and bistatic circular, wave- 1) frequency-modulated continuous wave radar and 2) PWR.
forms), and radar signal processing parameters [such as However, the discussion here is limited to PWR
coherent processing interval (CPI), and packet repeti- SimHumalator.
tion interval (PRI)] with a single click. The simulator’s capability to generate a diverse set of
6) The case studies presented in this work demonstrate human micro-Doppler signatures for different radar,
the feasibility of using SimHumalator to generate
1
vast amounts of a human micro-Doppler database shorturl.at/acxMZ

8 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 Vishwakarma et al.

Figure 1.
SimHumalator’s distribution package.

target, and radar signal processing parameters is briefly methodology is depicted in Figure 3. The target reflections
presented in the following sections. are simply the time-delayed and Doppler-shifted replica of
the direct WiFi transmissions. The time delay is directly pro-
portional to the target range, Doppler shift to the target’s
RADAR SIMULATION FRAMEWORK velocity, and the complex reflectivity to the target’s size,
shape, and material. The direct and the reflected signals are
A typical passive WiFi sensing setup is shown in Figure 2. then cross-correlated in the delay-Doppler plane to generate
It comprises a reference and surveillance antenna and a sig- the cross-ambiguity function (CAF).
nal processing unit. The reference antenna is a directional
antenna that captures the direct signal from the WiFi AP.
WIFI SIGNAL MODEL
On the other hand, the surveillance antenna is omnidirec-
tional to capture the reflected signals of the human targets SimHumalator uses MATLAB’s WLAN toolbox to simu-
present anywhere in the sensing area. The signals reflected late three IEEE standard-compliant waveforms: 1) IEEE
off the targets are time-delayed and Doppler-shifted direct 802.11 g at 2.4 GHz, 2) IEEE 802.11ax at 5.8 GHz, and 3)
signals. The radar signal processing unit aims to estimate IEEE 802.11ad at 60 GHz. The physical layer of IEEE
these parameters using direct and reflected signals. 802.11 Standards uses a packet-based protocol where each
SimHumalator integrates the IEEE 802.11 Standard- transmission packet [a physical layer conformance proce-
compliant WiFi signal and the human animation data to gen- dure protocol data unit (PPDU)] comprises a preamble
erate the radar scatterings off the humans. The hybridization and data, as shown in Figure 4.

Figure 2.
Typical PWR scenario comprising transmissions from WiFi access points (APs) and targets undergoing motion in the same propagation
environment.

MARCH 2022 IEEE A&E SYSTEMS MAGAZINE 9

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 SimHumalator: An Open-Source End-to-End Radar Simulator for Human Activity Recognition

Figure 3.
Radar signal model after integration of the target model with WiFi transmissions.

The preamble field is embedded with three subfields, packets with staggered packet repetition interval, that is, the
each comprising several time-domain samples in the leg- transmission of packets with different data payloads.
acy short training field (L-STF), legacy long training field
(L-LTF), and legacy signal field (L-SIG). L-STF possesses DYNAMIC TARGET MODEL
excellent correlation properties and is therefore used to
SimHumalator uses the animation files gathered using two
detect the start of the packet, L-LTF field for communica-
motion capture technologies: 1) a marker-based motion cap-
tion channel estimation, and the third preamble field L-
ture system called PhaseSpace and 2) a marker-less Micro-
SIG to indicate the amount of data transmitted (in octets).
soft’s Kinect v2 sensor system. The real advantage of using
On the other hand, the data field contains user payload,
motion capture technology is capturing more accurate, more
medium access control headers, and cyclic redundancy
realistic, and complex human motions. The data include 3D
check bits. We use only the preamble bits to form our dis-
time-varying skeletal information of the human body,
crete-time sequence yRef ½n. The simulated IEEE 802.11 g
including 25 joints, such as head center location, knee joints,
packet structures mimic real WiFi transmission formats at
elbow joints, and shoulder joints. It assumes that the radar
the 2.4 GHz band with a channel bandwidth of BW ¼
scattering centers are lying on the center of the bones, result-
20 MHz. In comparison, IEEE 802.11ax and IEEE
ing in 19 scatterers on the human body. The adopted simula-
802.11ad packet structures mimic real WiFi transmission
tion methodology is presented in Figure 6.
formats at 5.8 and 60 GHz. The channel bandwidth for
The human skeleton is embodied with elementary
IEEE 802.11ad WiFi transmissions is fixed to 1.76 GHz,
shapes to model different body parts, such as the torso,
while the bandwidth for IEEE 802.11ax transmissions can
arms, and legs using ellipsoids and the head using a sphere.
be chosen among BW ¼ 20 ; 40; 80 ; 160 MHz.
The radar scattering centers are assumed to be lying approxi-
A general formulation of WiFi transmission signal
mately at the center of these primitive shapes. SimHumala-
yRef ðtÞ is shown in (1). It comprises a continuous stream
tor uses these data to gather time-varying range rb ðtÞ and
of P transmission packets each of duration TP s at a car-
Doppler information of fDb ðtÞ of each scatterer. It then com-
rier frequency of fC , as shown in Figure 5
putes the reflectivities ab ðtÞ of each of these B primitive
1 X P X N
shapes and takes into account various factors, such as the
YRef ðtÞ ¼ pffiffi yRef ½nej2pdfðtpTP Þ ej2pfc t (1)
ðP Þ p¼1 n¼1 aspect angle ub ðtÞ, and the relative position rb ðtÞ, of the scat-
tering center on the primitive shape with respect to the radar.
where TP corresponds to the packet repetition interval,
The reflectivity of a primitive b, at any time instant t is
which is equal to the pulse repetition interval in our case, N is
given by
the total number of time-domain samples in one transmission pffiffiffiffiffiffiffiffiffiffi
packet, ts ¼ ð1=BWÞs is the sampling period, and df is zðtÞ s b ðtÞ
ab ðtÞ ¼ : (2)
OFDM subcarrier spacing. We have used uniform packet rep- r2b ðtÞ
etition interval for our simulations. However, the future devel- Here, zðtÞ subsumes propagation effects, such as attenuation,
opment of the simulator would include transmissions of antenna directivity, and processing gains, s b ðtÞ is the radar
cross-section of the primitives. The radar cross section (RCS)
of primitive shapes is well characterized at microwave fre-
quencies [26]. The RCS of an ellipsoid of length Lb and
radius Rb is given by

Figure 4. pffiffiffi p 4 2
4 Rb Lb
s b ðtÞ ¼ G 1 2
: (3)
IEEE 802.11 g Standard-complaint OFDM transmission packet R2b sin 2 u b ðtÞ þ 4 Lb cos 2 ub ðtÞ
structure.

10 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 Vishwakarma et al.

Figure 5.
IEEE 802.11 g Standard-complaint continuous packet transmission.

Figure 6.
Simulation framework for generating an electromagnetic human primitive scattering model from motion capture animation data.

The dielectric properties of human skin are incorporated into CAF PROCESSING
the RCS estimation through the Fresnel reflection coefficient
The CAF processing is implemented over the received
G. The human is assumed to be a single-layer dielectric with
data and the direct reference signal data to compute the
a dielectric constant of 80 and conductivity of 2 S/m.
delay t b and Doppler information fDb of the target. The
adopted CAF processing is shown in Figure 7. Match-fil-
HYBRID ELECTROMAGNETIC RADAR SCATTERING FROM tering is performed along the fast time samples and fast
DYNAMIC HUMANS Fourier transform along the slow time samples to generate
CAFs. The CAF processing is implemented as
The received signal is simply the attenuated, time-delayed
t b , and Doppler-shifted fDb , version of the direct transmit- Z Ti

ted signal YRef ðtÞ. Ignoring multipath, the baseband CAFðt; fd Þ ¼ Ysur ðtÞYref ðt  tÞej2pfd t dt (5)
0
received signal on the surveillance channel can be repre-
sented as
where  denotes complex conjugation, t is the time delay,
fd is the Doppler shift, and Ti is the integration time.
P X
X B
Note that the direct signal component is always strong
YSur ðtÞ ¼ ab ðtÞYRef ðt  t b  pTp Þej2pfDb pTp
p¼1 b¼1
and can mask the target returns present in the CAFs.
Therefore, the CLEAN algorithm is used to remove direct
þ zðtÞ: (4) signal interference [9]. The CLEAN algorithm subtracts
the self-ambiguity function (generated using only the ref-
Here, c ¼ 3  108 m/s is the speed of light, and zðtÞ is the erence signal) from the CAF calculated in (5) thereby sup-
additive circular-symmetric white noise. pressing stronger direct signals.

MARCH 2022 IEEE A&E SYSTEMS MAGAZINE 11

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 SimHumalator: An Open-Source End-to-End Radar Simulator for Human Activity Recognition

Figure 7.
CAF generation through the cross-correlation between the direct WiFi transmission and the reflected signals off the targets.

Multiple CAFs spanning a duration of TTotal are proc- complementary codes that hold perfect autocorrelation
essed to generate the Doppler-time spectrogram, as shown properties. Therefore, cross-correlation operation results
in Figure 8. in high sidelobe levels, significantly affecting the quality
Here, for each CPI, the peaks along the range axis are of range-time signatures. Even if the waveforms have per-
coherently added for each Doppler bin. fect autocorrelation properties, the sidelobe levels are zero
only for static target scenarios. As soon as the target starts
moving, high sidelobes appear in the range domain.
GRAPHICAL USER INTERFACE (GUI) SimHumalator offers users the freedom to select anima-
tion data files from 11 healthcare-related human motion clas-
SimHumalator allows the user to select many input param- ses. These include human rotating his body (HBR) while
eters to specify the target motion characteristics, sensor standing in a fixed position, human kicking (HK), human
operating conditions, and signal processing parameters. punching (HP), human grabbing an object (HG), human
The simulator receives the simulation inputs with the help walking back and forth in front of the radar (HW), human
of a GUI window, as shown in Figure 9. In addition, the standing up from a chair, human sitting down on a chair,
user’s manual can help visualize the parameters within the human stand up from the chair to walk, human walk to sit on
simulation inputs window represent and how changing a chair, human walk to fall on the ground, and a human stand-
these parameters affects the radar signatures. ing up from the ground to walk. The number of animation
SimHumalator can generate a massive signature files for each of these activities is presented in Table 1.
library of radar signature profiles of various human activi- Some files of the first five activities, 1) HBR, 2) HK,
ties, including CAFs and Doppler-time plots. Since we are 3) HP, 4) HG, and 5) HW, have been captured before the
simulating IEEE 802.11 g WiFi transmissions, the signal COVID-19 pandemic using a marker-based active track-
bandwidth is limited to 20 MHz, insufficient to locate tar- ing phase-space system [27]. The remaining six activities
gets in most indoor scenarios. Therefore, we focus on and some data for the first five activities were captured
extracting only the time-varying micro-Doppler informa- using Kinect v2 sensor during the COVID-19 pan-
tion in the joint time–frequency space. Moreover, the stan- demic [28]. We will continue to build our database and
dard waveforms IEEE 802.11ax and IEEE 802.11ad, provide the user community with a platform to generate
which offer a good resolution range, do not comprise substantial radar databases.

Figure 8.
Simulation methodology to generate micro-Doppler spectrograms using multiple CAFs spanning the entire duration of motion.

12 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 Vishwakarma et al.

Figure 9.
SimHumalator’s GUI for choosing PWR system parameters.

SYNTHESIZED SPECTROGRAM EVALUATION MONOSTATIC CONFIGURATION OF PWR SYSTEM WITH

The PWR signal parameters used are provided in Table 2. The VARYING TARGET ASPECT ANGLES
CPI is fixed to 0.05 s, which is sufficient to capture time-vary-
In the monostatic configuration of the PWR system, the
ing micro-Doppler features in the joint time–frequency space.
WiFi AP and radar receiver are co-located. Figure 10(a)
Note that the PRF dictates the maximum Doppler fre-
presents the micro-Doppler signature of a human moving
quency range that can be observed without aliasing
at 0 aspect angle with respect to the radar receiver (direct
(fDmax ¼ PRF=2), Doppler resolution by the CPI
line-of-sight). Since the human is walking in the direction
(dfD ¼ 1=CPI), and the choice of the nonoverlap time win-
dow is critical for successfully representing the micro-Dopp- of the radar, the micro-Dopplers are mostly positive. The
ler signature of a dynamic target. The nonoverlap time strongest returns arise from the torso due to its larger
window here is similar to the short-time window in the case cross-section, while the weaker returns are generated from
of the short-time Fourier transform. The nonoverlap time win- the motions of the arms and legs. Some of the negative
dow should ideally be low for fast-moving targets and high for Dopplers arise from the back-swing of the arms and legs.
capturing motion characteristics of slow-moving targets. In most realistic scenarios, the human motions might not
Therefore, users should choose these parameters carefully to be restricted to a single aspect angle with respect to the
obtain the desired extent and resolution spectrograms. radar. In such scenarios, the spectrograms might differ

MARCH 2022 IEEE A&E SYSTEMS MAGAZINE 13

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 SimHumalator: An Open-Source End-to-End Radar Simulator for Human Activity Recognition

Table 1. Table 2.

Motion Capture Data Information Simulated PWR Parameters

Activity Number of files Radar parameters Values

Human rotating his body 104 Carrier frequency (fc ) 60 GHz

Human kicking 104 Bandwidth (BW) 1.762 GHz
Human punching 103 Packet repetition frequency 1000 Hz
(PRF ¼ 1=Tp )
Human grabbing an object 103
Coherent processing Interval 0:05 s
Human walking back and forth 106
Maximum Doppler 500 Hz
Human standing up from chair 71
(fDmax ¼ PRF/2)
Human sitting down on chair 71
Doppler resolution (Df D ¼ 1=CPI) 20 Hz
Human stand up from chair to walk 69

Human walk to sit on chair 68

Human walk to fall on ground 60

BISTATIC CONFIGURATION WITH VARYING BISTATIC
Human standing up from ground to walk 60
ANGLE AND FIXED TARGET ASPECT ANGLE
In most passive WiFi sensing scenarios, the WiFi AP and
significantly. It could be due to the shadowing of some the radar receiver are separated by a certain distance
part of the human body if captured at different angles. resulting in a bistatic configuration of the setup. Here,
Figure 10(b)–(d) shows the spectrograms of a human the Doppler frequency fD ¼ 2vfc c cosðuÞcosðb=2Þ is gov-
walking at three different aspect angles 1) 60 , 2) 120 , erned by aspect angle u and the bistatic angle b. For sim-
and 3) 180 with respect to the radar, respectively. The plicity, the aspect angle of the target is fixed at 0 ;
signatures at 60 , shown in Figure 10(b), mostly have therefore, the Dopplers only depend upon the bistatic
reduced positive Dopplers compared to the 0 aspect angles fD ¼ 2vfc c cosðb=2Þ. Figure 10(f)–(h) shows the
angle. This is because the target is still approaching but micro-Doppler signatures of a walking human at four
with a reduced radial component toward the radar. The bistatic angles 0 , 60 , 120 , and 180 , respectively. In
Dopplers become negative when the target aspect angle is the forward scatterer position, when b ¼ 180 , the Dopp-
120 . At this angle, the target begins to move away from ler should be zero (fD ¼ 0). However, since humans are
the radar. Finally, the signatures at 180 represent a human extended targets, some micro-Doppler returns are due to
walking away from the radar, thus having all the negative the swinging motion of arms and legs. This is evident in
Dopplers. Figure 10(h).

Figure 10.
(a)–(d) Radar micro-Doppler signatures of human walking at four aspect angles 0 , 60 , 120 , and 180 with respect to the monostatic radar,
respectively. (e)–(h) Radar micro-Doppler signatures of human walking at aspect angle 0 to the radar at four bistatic radar configurations
with the following bistatic angles 0 , 60 , 120 , and 180 , respectively.

14 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 Vishwakarma et al.

Figure 11.
Micro-Doppler spectrograms of a human walking motion for (a) IEEE 802.11 g Standard at 2.4 GHz, (b) IEEE 802.11ax Standard at
5.8 GHz, and (c) IEEE 802.11ad Standard at 60 GHz, respectively.

low carrier frequency data. However, the classification algo-

MICRO-DOPPLER SPECTROGRAMS ACROSS MULTIPLE
rithms might pose some unique challenges when the mea-
STANDARDS surement data “merge” sensing inputs from sub-7 GHz and
mmWave. Therefore, to demonstrate the usefulness of mul-
Growing demand of throughput for next-generation multime-
tiband WiFi sensing systems in real-world scenarios, the
dia wireless services, such as video streaming, data transfer
algorithms would need to use nonheuristic methods to derive
among smart terminals (Internet of Things), high-dense
features for handling the diversity in the measurement data.
deployment areas, such as airports, shopping malls, stadiums,
and office spaces, has necessitated the need for creating com-
plex wireless networks. Therefore, to accommodate the CASE STUDIES
increasing demands and challenging application scenarios,
This section presents case studies to demonstrate the feasi-
most WiFi devices will soon be equipped with antennas oper-
bility of using SimHumalator to generate vast amounts of a
ating at both sub-7 GHz frequencies (to be compatible with
human micro-Doppler database comprising high-quality
legacy standards—IEEE 802.11ax and earlier) and mmWave
micro-Doppler spectrograms for activity recognition appli-
frequencies (60 GHz IEEE 802.11ad/ay and subsequent) [29].
cations. In addition, the study demonstrates that the synthe-
The sub-7 GHz WiFi signals have better wall penetration
sized signatures can be used for data augmentation to solve
capabilities and richer multipath information. Therefore, they
the practical problem of insufficient or unbalanced micro-
can be used to sense relatively large motions even through
Doppler training data. Some other studies used SimHumala-
obstacles (through walls). On the other hand, the measure-
tor for studying human micro-Doppler signature classifica-
ments at mmWave can sense fine-grained motions and pro-
tion in the presence of a selection of jamming signals [30].
vide highly directional information through beamforming.
The case studies use 802.11 g packet structures to
Figure 11 presents the micro-Doppler spectrograms of a
mimic real WiFi transmission formats at the 2.4 GHz
human walking motion corresponding to three standard wave-
band with a channel bandwidth of BW ¼ 20 MHz.
forms: 1) Figure 11(a)—IEEE 802.11 g Standard at 2.4 GHz,
2) Figure 11(b)—IEEE 802.11ax Standard at 5.8 GHz, and 3)
Figure 11(c)—IEEE 802.11ad Standard at 60 GHz. HUMAN ACTIVITY RECOGNITION USING
We keep fixed maximum detectable Doppler frequency
to 500 Hz and nonoverlap time window to 0.5 s for all the
SIMHUMALATOR-GENERATED SPECTROGRAMS
plots to demonstrate the substantial differences arising from The human micro-Doppler spectrograms simulated from
different carrier frequencies. The figures demonstrate that SimHumalator are classified for the following activities:
the Doppler spectrograms corresponding to high carrier fre- HBR, HK, HP, HG, and HW. Sample spectrograms for
quency data show finer frequency resolution than those from the five motion classes are shown in Figure 12.

Figure 12.
Radar micro-Doppler signatures for a human undergoing (a) body rotation motion, (b) kicking motion, (c) punching motion, (d) grabbing an
object motion, and (e) walking in the direction of the monostatic configuration of PWR radar at the 2.4 GHz band.

MARCH 2022 IEEE A&E SYSTEMS MAGAZINE 15

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 SimHumalator: An Open-Source End-to-End Radar Simulator for Human Activity Recognition

Table 3.

Simulation Dataset Description

Target Number of Duration of Number of spectrograms (1.5 s each) In radar configuration

class MoCap data (in each
Files file) Monostatic Bistatic

Aspect angle 0 Varying aspect angle Varying bistatic angle

(Overlapping 0:5:360 (No 0:5:360 (No overlapping)
Window 0.5 s) overlapping)

HBR 10 4.5 s 90 2190 2190

HK 20 4.5 s 180 4380 4380

HP 20 4.5 s 180 4380 4380

HG 20 4.5 s 180 4380 4380

HW 19 4.5 s 171 4161 4161

Total data in each configuration 801 19,491 19,491

These motions are periodic and thus have alternating 1.5 s with an overlapping time of 0.5 s was used over the
positive and negative micro-Doppler features. Therefore, full signature of duration 4.5 s. It resulted in nine spectro-
it becomes challenging for any classifier to discern the grams, each of duration 1.5 s from every motion capture
correct motion class. file. Table 3 summarizes the entire simulation database.
This section studies the different classification algo- We use handpicked features [31], Cadence velocity
rithms’ robustness to classify micro-Dopplers in more features (CVD) [32], and automatically extracted sparse
complex scenarios, such as varying aspect angles and features [33],[34] from the micro-Doppler signatures to
bistatic angles. Note that the human motions in these test the performance of a classical machine learning-based
repeated measurements were unrestricted, and therefore, support vector machine classifier [35]. We then compared
the micro-Doppler signatures vary due to differences in their performances with a deep convolutional neural net-
gait patterns in every simulation. The duration of each work (DCNN) that has a joint feature extraction and clas-
measurement was 4.5 s. A sliding window of duration sification framework within the same network. We

Table 4.

Classification Accuracies of Multiple Algorithms for a Simulation Database (Captured for a Fixed Aspect Angle of the
Target)

Target class/ Handpicked CVD Sparse Neural networks

algorithm features features features
DCNN AlexNet GoogLeNet ResNet18

HBR 100 98.9 99.1 100 100 100 100

HK 83.9 83 93.1 100 100 100 96.3
HP 73.8 92.3 92.4 100 96.3 96.3 100
HG 82.5 98.3 97.6 96.3 100 100 100
HW 100 94.5 100 100 100 100 100
Overall accuracy 88 93.4 96.4 99.3 99.3 99.3 99.3
(%)

16 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 Vishwakarma et al.

Table 5.

Classification Accuracies for a Simulation Database (Captured for Varying Aspect Angle of the Target With Respect to the
Radar Receiver)

Target class/ Handpicked CVD Sparse Neural networks

algorithm features features features
DCNN AlexNet GoogLeNet ResNet18

HBR 93.2 91.6 98.6 100 100 99.4 100

HK 41.8 34 81.7 94.2 91.8 88.6 93.6
HP 6.1 90.9 88.1 95.7 98.3 93.6 97.6
HG 29.7 61.5 90.2 92.5 89.3 98.5 99.4
HW 90.3 69 96.5 87.8 98.6 99.4 99.8
Overall accuracy 52.02 69.4 91.02 94.04 95.6 95.9 97.84
(%)

designed a 24-layered deep neural network comprising considered in the study share common features in
three components (convolution layer, pooling layer, and the micro-Doppler feature space because of the
activation functions). We also tested some pretrained deep proximity between different motion categories.
neural networks, such as AlexNet, GoogLeNet, and The resulting classification accuracies are pre-
ResNet18. We used 70% of each target’s spectrograms as sented in Table 4. The results demonstrate that all
the training dataset, 15% as the validation set, and the deep neural networks outperform the classical
remaining 15% as the test dataset. machine learning-based methods and achieve an
The following four classification scenarios were con- average classification accuracy of 99%. This is
sidered to give readers a better understanding of the sensi- because the deep neural networks, a cascaded
tivity of the algorithm’s performance to the simulation structure of neurons, can learn any complex func-
database. tion to create a decision boundary even for nonlin-
ear data considered in the study. The classical
 Case 1a—Train using data from a fixed zero aspect machine learning algorithms, on the other hand,
angle: The algorithms’ performances using a simu- are not capable of learning these complex discern-
lation database generated for a fixed 0 aspect ing boundaries, resulting in poor classification
angle of the target. Note that the five target classes performances.

Table 6.

Classification Accuracies for a Simulation Database (Captured Under Varying Bistatic Circular Configurations.

Target class/ Handpicked CVD Sparse Neural networks

algorithm features features features
DCNN AlexNet GoogLeNet ResNet18

HBR 93.6 91.8 99.1 80.2 82.6 87.2 91.2

HK 17.2 35.5 81.8 86.9 83.9 85.8 91.6
HP 21.2 88.7 89.7 97 95.9 91.3 99.5
HG 64.4 60.3 89.3 94.8 97.9 93.6 95.1
HW 90.9 73.8 95.8 98.6 99.2 100 99.4
Overall accuracy 57.46 70.2 91.14 91.5 91.9 91.58 95.79
(%)

MARCH 2022 IEEE A&E SYSTEMS MAGAZINE 17

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 SimHumalator: An Open-Source End-to-End Radar Simulator for Human Activity Recognition

encounter. Here, the performance is observed to be

lower compared to the previous two cases. This is
because the micro-Doppler signatures are captured
in a forward scatter geometry, resulting in micro-
Doppler being centered mostly around 0 Doppler
for almost all the motion classes. This leads to a sig-
nificant reduction in classification accuracies. The
classification results for this case are presented in
Table 6.
 Case 1d—Performance under noisy conditions:
Most realistic scenarios have environmental factors
Figure 13. that significantly affect the resulting SNRs. There-
Average classification accuracies as a function of SNR. fore, to assess SNR’s impact on the classification
performances, Gaussian noise was added to achieve
 Case 1b—Train using data from multiple aspect varying levels of SNR (-2 to 10 dB) in the simulated
angles:Next, the performance was analyzed when micro-Doppler signatures. The classification was
the algorithms were trained using a simulation data- challenging under noisy conditions due to the close
base comprising micro-Doppler signatures captured similarity of motion classes considered in the data-
at multiple aspect angles. It is a significantly more set. Figure 13 shows the variation of average classi-
challenging and realistic scenario since no aspect fication performances of different algorithms as a
angle information is available during the test phase. function of SNR. As the noise level is increased,
Note that the spectrograms used in testing have not there is an expected drop in different algorithms’
been used during training. overall performance. However, the drop in perfor-
Table 5 presents the resulting classification accura- mance is higher for classical learning-based meth-
cies across different algorithms. The average classi- ods (represented in black) than that for deep
fication accuracies are lower in this case. The learning models (represented in red). The results
reason for poor performance is that the Doppler suggest that the deep networks can extract features
spectrogram for a particular motion class at certain that are more robust to noise and, therefore, be the
aspect angles might look similar to that for other “classifier of choice” in high-noise scenarios (com-
motion classes at the same angle. Algorithms based prising various environmental factors such as noise
on deep networks perform exceptionally well even and multipath).
under diverse training and test datasets. It indicates
that these algorithms are specifically suited for
problems dealing with a great deal of diversity in
radar data. The best performing network is EXPERIMENTAL VALIDATION OF SIMHUMALATOR
ResNet18, with an average classification accuracy
The simulator’s reliability is improved by validating it
of 97.8%.
with experimental data gathered using an in-house-built
 Case 1c—Train using data from multiple bistatic hardware prototype [36]. Figure 14 presents the experi-
angles:It is a more practical scenario that one can mental setup comprising two synchronized systems: 1) an

Figure 14.
Experimental setup comprising two synchronized systems: 1) a motion capture Kinect v2 sensor and 2) a PWR system for monitoring human
activities in indoor scenarios.

18 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 Vishwakarma et al.

Figure 15.
Velocity–time profile using motion capture data, simulated and measured Doppler spectrograms of a human undergoing (a)–(c) sitting fol-
lowed by standing up motion, (d)–(f) stand up from chair and starts walking, (g)–(i) walk to fall motion, respectively.

infrared motion capture Kinect v2 sensor and 2) a noncon- slightly in the radar direction while sitting down. How-
tact physical activity monitoring PWR system. ever, the strength of the micro-Doppler is low compared
The PWR system was set up using two Yagi antennas, to bulk body motion. After a 5 s delay, the human sub-
each with a gain of 14 dBm, two National Instruments ject stands up from the chair, resulting in primarily posi-
(NI) USRP-2921 [37] and a Netgear R6300 transmitter tive Dopplers. Similarly, Figure 15(d)–(f) presents
acting as the WiFi AP. The WiFi AP was configured to spectrograms corresponding to a human standing up
transmit an 802.11 g Standard-compliant waveform at a from a chair and before walking. Figure 15(g)–(i),
center frequency of 2.472 GHz. The PWR system used presents spectrograms corresponding to a human transi-
one antenna, a reference antenna, to capture direct WiFi tioning from walking to falling. The strength of the sig-
transmissions from the AP. Simultaneously, the second is nals of the spectrograms shown in Figure 15(a)–(i) may
used as a surveillance antenna to capture signals reflected not be the same; however, the envelope of the velocity–
off targets in the propagation environment. The reference time profile is visually very similar across all. The simu-
WiFi signal and the surveillance signal were then cross- lations do not synthesize environmental factors like
correlated to generate the radar micro-Doppler signatures noise, propagation loss, occlusions, and multipath. There-
in real time. fore, the simulated spectrograms are clean compared to
Figure 15 shows the qualitative comparison between the measured signatures.
the micro-Doppler spectrograms generated using the PWR
measurement system and the SimHumalator.
Figure 15(a)–(c) shows the velocity–time profile
(generated using motion capture data) and simulated and
EXPERIMENTAL DATA AUGMENTATION
the measured spectrograms, respectively, for a human SimHumalator can effectively generate high-quality
undergoing a motion of first sitting down on a chair fol- micro-Doppler spectrograms from the motion capture
lowed by standing up from the chair. The qualitative data. Therefore, it can be used to synthesize signatures for
similarity between all the signatures is evident from the data augmentation purposes to solve the practical problem
figure. As the human sits down, there is a negative Dopp- associated with insufficient or unbalanced micro-Doppler
ler due to the bulk body motion. The positive micro- training data [36],[38]. This section presents classification
Doppler arises due to an arm motion and legs moving results from various data augmentation schemes. The

MARCH 2022 IEEE A&E SYSTEMS MAGAZINE 19

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 SimHumalator: An Open-Source End-to-End Radar Simulator for Human Activity Recognition

Figure 16.
Augmentation study: The training dataset changes with the percentage of simulated data (s) augmented with the measurement training data.
The percentage of simulation data is varied to study the impact of data augmentation on classification performance. s=0 is a special case
where training and test data only comprise measured spectrograms.

augmentation scheme adds additional simulation data, CONCLUSION

increasing the training dataset’s overall size, which only
used measurement data. The study can be summarized in This work presents an open-source motion capture data-
Figure 16. The total measurement data (M) are divided driven simulation tool, SimHumalator, that can generate
into the training and test sets according to the split ratio of large volumes of human micro-Doppler radar data at
m. Then, the s percentage of the total simulation data (S) multiple IEEE WiFi Standards (IEEE 802.11 g, ax, and
is added to the training set, resulting in a training set size ad). SimHumalator can simulate human micro-Doppler
of ðmM þ sSÞ. radar returns as a function of a diverse set of target
First, the classification performance is tested when parameters, radar parameters, and radar signal process-
only measurement data are used for training and test sce- ing parameters. The simulator’s reliability is improved
narios. Next, the training dataset was created by combin- by validating it with experimental data gathered using
ing the simulation and experimentally measured data. an in-house-built hardware prototype. A qualitative com-
Only unseen measurement data were used to test the pre- parison between the simulated and real measured spec-
trained VGG16 model’s performance during the training trogram shows that the SimHumalator can effectively
process. Figure 17 presents the augmentation results. generate high-quality micro-Doppler spectrograms from
The smaller the volume of measured data in the train- the motion capture data. This article contributes to
ing set, the more pronounced the classification accuracy. developing a radar simulator, which can help the
The outcomes highlight that for cases where only limited research community augment the limited measurement
experimental data are available for training, classification training data to improve the overall classification perfor-
performance can be improved up to 8%. mance and address the cold-start problem typically
In addition, the data generated by the simulator can be encountered in radar research. Furthermore, it can help
used for replacing a small amount of low-quality mea- simulate an electronic system without risk and efficiently
sured data and filling a small number of imbalanced clas- perform the “What if” analysis. Interested researchers
ses in practical applications. can download the simulator from https://uwsl.co.uk/.
While we rigorously tested the app internally and with
our Beta users, there can still be a few undiscovered
bugs or issues. Therefore, we believe that user feedback
is crucial for developing the app.
Future simulator development will be focused on mak-
ing the simulated data as close to real-world user scenarios
as possible. To achieve this, we hope to incorporate both
user feedback and environmental factors such as noise,
multipath effects, and target and surveillance area geomet-
rical factors. We believe that this simulator can become a
flexible and comprehensive tool to benchmark future algo-
rithms and generate large volumes of realistic simulation
data for the user community through a user-friendly GUI.
In addition, it will reduce the expense and labor involved
in experimental data acquisition by other researchers and
Figure 17.
help radar and sensing communities benchmark future
Augmentation study: Classification accuracies of the percentage
of measurement data augmented with simulation data (s=100%). machine learning algorithms.

20 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 Vishwakarma et al.

[13] S. A. Shah and F. Fioranelli, “RF sensing technologies for

ACKNOWLEDGMENTS
assisted daily living in healthcare: A comprehensive
This work was supported by the U.K. Engineering and review,” IEEE Aerosp. Electron. Syst. Mag., vol. 34,
Physical Sciences Research Council (EPSRC), through no. 11, pp. 26–44, Nov. 2019.
the OPERA Project, under Grant EP/R018677/1. [14] Y. Lin and J. Le Kernec, “Performance analysis of classifi-
cation algorithms for activity recognition using micro-
Doppler feature,” in Proc. IEEE 13th Int. Conf. Comput.
Intell. Secur., 2017, pp. 480–483.
REFERENCES
[15] Y. Lin, J. Le Kernec, S. Yang, F. Fioranelli, O. Romain,
[1] “World aging population.” Accessed: Jan. 18, 2022. [Online]. and Z. Zhao, “Human activity classification with radar:
Available: https://www.un.org/en/development/desa/ Optimization and noise robustness with iterative convolu-
population/publications/pdf/ageing/WPA2017_Highlights.pdf tional neural networks followed with random forests,”
[2] S. Chernbumroong, S. Cang, A. Atkins, and H. Yu, IEEE Sensors J., vol. 18, no. 23, pp. 9669–9681, Dec.
“Elderly activities recognition and classification for appli- 2018.
cations in assisted living,” Expert Syst. Appl., vol. 40, [16] B. Zhou et al., “Simulation framework for activity recogni-
no. 5, pp. 1662–1674, 2013. tion and benchmarking in different radar geometries” IET
[3] G. Diraco, A. Leone, and P. Siciliano, “An active vision Radar, Sonar and Navigation, vol. 15, no. 4, pp. 390–401,
system for fall detection and posture recognition in elderly 2021.
healthcare,” in Proc. IEEE Des., Automat. Test Euro. [17] G. E. Smith, K. Woodbridge, and C. J. Baker, “Multistatic
Conf. Exhib, 2010, pp. 1536–1541. micro-Doppler signature of personnel,” in Proc. IEEE
[4] S. Vishwakarma and S. S. Ram, “Detection of multiple Radar Conf., 2008, pp. 1–6.
movers based on single channel source separation of their [18] R. Boulic, N. M. Thalmann, and D. Thalmann, “A
micro-Dopplers,” IEEE Trans. Aerosp. Electron. Syst., global human walking model with real-time kinematic
vol. 54, no. 1, pp. 159–169, Feb. 2018 . personification,” Vis. Comput., vol. 6, no. 6, pp. 344–
[5] B. Erol and M. G. Amin, “Radar data cube analysis for fall 358, 1990.
detection,” in Proc. IEEE Int. Conf. Acoust., Speech Signal [19] G. Manfredi, P. Russo, A. De Leo, and G. Cerri, “Efficient
Process., 2018, pp. 2446–2450. simulation tool to characterize the radar cross section of a
[6] V. C. Chen, F. Li, S.-S. Ho, and H. Wechsler, “Micro- pedestrian in near field,” Prog. Electromagnetics Res. C,
Doppler effect in radar: Phenomenon, model, and simula- vol. 100, pp. 145–159, 2020.
tion study,” IEEE Trans. Aerosp. Electron. Syst., vol. 42, [20] S. S. Ram and H. Ling, “Simulation of human microDop-
no. 1, pp. 2–21, Jan. 2006. plers using computer animation data,” in Proc. IEEE
[7] J. Le Kernec et al., “Radar signal processing for sens- Radar Conf., 2008, pp. 1–6.
ing in assisted living: The challenges associated with [21] B. Erol and S. Z. Gurbuz, “A Kinect-based human micro-
real-time implementation of emerging algorithms,” Doppler simulator,” IEEE Aerosp. Electron. Syst. Mag.,
IEEE Signal Process. Mag., vol. 36, no. 4, pp. 29–41, vol. 30, no. 5, pp. 6–17, May 2015.
Jul. 2019. [22] A. D. Singh, S. S. Ram, and S. Vishwakarma, “Simulation
[8] X. Yang et al., “Freezing of gait detection considering of the radar cross-section of dynamic human motions using
leaky wave cable,” IEEE Trans. Antennas Propag., virtual reality data and ray tracing,” in Proc. IEEE Radar
vol. 67, no. 1, pp. 554–561, Jan. 2019. Conf., 2018, pp. 1555–1560.
[9] K. Chetty, G. E. Smith, and K. Woodbridge, [23] “CMU Graphics Lab Motion Capture Database,” Accessed:
“Through-the-wall sensing of personnel using passive Dec. 24, 2019. [Online]. Available: http://mocap.cs.cmu.edu/
bistatic WiFi radar at standoff distances,” IEEE Trans. [24] “Accad motion capture lab database,” Accessed: Dec. 24,
Geosci. Remote Sens., vol. 50, no. 4, pp. 1218–1226, 2019. [Online]. Available: https://accad.osu.edu/research/
Apr. 2012. motion-lab/mocap-system-and-data/
[10] V. C. Chen, The Micro-Doppler Effect in Radar. Norwood, [25] M. M€uller, T. R€oder, M. Clausen, B. Eberhardt, B. Kr€uger,
MA, USA: Artech House, 2019. and A. Weber, “Documentation MoCap database HDM05,”
[11] B. Tan, Q. Chen, K. Chetty, K. Woodbridge, W. Li, Universit€at Bonn, Bonn, Germany Tech. Rep. CG-2007-2,
and R. Piechocki, “Exploiting WiFi channel state Jun. 2007.
information for residential healthcare informatics,” [26] E. F. Knott, J. F. Schaeffer, and M. T. Tulley, Radar Cross
IEEE Commun. Mag., vol. 56, no. 5, pp. 130–137, Section. Rijeka, Croatia: SciTech Publ., 2004.
May 2018. [27] “Phase-space motion capture system,” Accessed: Aug. 10,
[12] W. Li, B. Tan, and R. Piechocki, “Passive radar for oppor- 2020. [Online]. Available: https://www.phasespace.com/
tunistic monitoring in e-health applications,” IEEE J. [28] Microsoft, “Kinect sensor,” 2021. [Online]. Available:
Transl. Eng. Health Med., vol. 6, 2018, Art. no. 2800210, www.xbox.com/en-US/kinect
doi: 10.1109/JTEHM.2018.2791609.

MARCH 2022 IEEE A&E SYSTEMS MAGAZINE 21

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.
 SimHumalator: An Open-Source End-to-End Radar Simulator for Human Activity Recognition

[29] F. Restuccia, “IEEE 802.11 BF: Toward ubiquitous Wi-Fi [34] S. Vishwakarma and S. S. Ram, “Dictionary learning with
sensing,” 2021, arXiv:2103.14918. low computational complexity for classification of human
[30] D. Dhulashia, M. Ritchie, S. Vishwakarma, and K. Chetty, micro-Dopplers across multiple carrier frequencies,” IEEE
“Human micro-Doppler signature classification in the pres- Access, vol. 6, pp. 29793–29805, 2018.
ence of a selection of jamming signals.” in Proc. IEEE [35] S. S. Keerthi, S. K. Shevade, C. Bhattacharyya, and K. R.
Radar Conf., 2021, pp. 1–6. Murthy, “A fast iterative nearest point algorithm for sup-
[31] Y. Kim and H. Ling, “Human activity classification based port vector machine classifier design,” IEEE Trans. Neural
on micro-Doppler signatures using a support vector Netw., vol. 11, no. 1, pp. 124–136, Jan. 2000.
machine,” IEEE Trans. Geosci. Remote Sens., vol. 47, [36] C. Tang, S. Vishwakarma, W. Li, R. Adve, S. Julier, and
no. 5, pp. 1328–1337, May 2009. K. Chetty, “Augmenting experimental data with simula-
[32] R. Ricci and A. Balleri, “Recognition of humans based on tions to improve activity classification in healthcare mon-
radar micro-Doppler shape spectrum features,” IET Radar, itoring,” in Proc. IEEE Radar Conf., 2021, pp. 1–6.
Sonar Navigation, vol. 9, no. 9, pp. 1216–1223, 2015. [37] NI USRP 2921. [Online]. Available: http://sine.ni.com/
[33] G. Li, R. Zhang, M. Ritchie, and H. Griffiths, “Sparsity- nips/cds/view/p/lang/en/nid/212995
driven micro-Doppler feature extraction for dynamic hand [38] S. Vishwakarma, C. Tang, W. Li, K. Woodbridge, R.
gesture recognition,” IEEE Trans. Aerosp. Electron. Syst., Adve, and K. Chetty, “GAN based noise generation to aid
vol. 54, no. 2, pp. 655–665, Apr. 2018. activity recognition when augmenting measured WiFi
radar data with simulations.” in Proc. IEEE Int. Conf.
Commun., 2021, pp. 1–6.

22 IEEE A&E SYSTEMS MAGAZINE MARCH 2022

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on October 20,2022 at 16:38:46 UTC from IEEE Xplore. Restrictions apply.