Human Target Detection, Tracking, and Classification Using 24-GHz FMCW Radar
This paper presents a millimeter wave radar in the 24-GHz ISM band for detection, tracking, and classification of human targets. Linear frequency modulation of the transmit signal and two receive antennas enable distance and angle measurements, respectively. Multiple consecutive frequency chirps are used for target velocity calculation. Hardware as well as firmware concepts of the proposed system are described in detail. Various algorithms for human detection and tracking are investigated and combined to a new signal processing routine optimized for compactness and low-power to run on a microcontroller. Additionally, a novel Doppler-compensated angle-ofarrival estimation method as well as a one-class support vector machine for human classification are proposed to further enhance the human detection and tracking performance. The achieved performances of the designed hardware and the implemented algorithm are verified in extensive measurements. The distance and angle errors of the realized radar sensor are at most 25 cm along a measurement range of 18m and 10° for a two-sided angle sweep of 65°, respectively. The achieved range resolution is 0.9 m. Dedicated verifications of the most important signal processing routines are presented to verify their functionality and experiments with several human targets illustrate the performance and limits of the overall tracking algorithm. It is shown that range, velocity, and angle of up to five humans are correctly detected and tracked. The presented one-class classifier successfully distinguishes human targets from other quasi-static targets like trees and shadowing effects of human subjects on walls.
IEEE SENSORS JOURNAL 1
Human Target Detection, Tracking, And
Classification Using 24 GHz FMCW Radar
Christoph Will, Student Member, IEEE, Prachi Vaishnav, Abhiram Chakraborty, Member, IEEE,
and Avik Santra, Senior Member, IEEE
Abstract—This paper presents a millimeter wave radar in the
24-GHz ISM band for detection, tracking, and classification of
human targets. Linear frequency modulation of the transmit
signal and two receive antennas enable distance and angle mea-
surements, respectively. Multiple consecutive frequency chirps
are used for target velocity calculation. Hardware as well as
firmware concepts of the proposed system are described in
detail. Various algorithms for human detection and tracking are
investigated and combined to a new signal processing routine
optimized for compactness and low-power to run on a micro-
controller. Additionally, a novel Doppler-compensated angle-of-
arrival estimation method as well as a one-class support vector
machine for human classification are proposed to further enhance
the human detection and tracking performance. The achieved
performances of the designed hardware and the implemented
algorithm are verified in extensive measurements. The distance
and angle errors of the realized radar sensor are at most 25 cm
along a measurement range of 18 m and 10◦for a two-sided
angle sweep of 65◦, respectively. The achieved range resolution
is 0.9 m. Dedicated verifications of the most important signal
processing routines are presented to verify their functionality
and experiments with several human targets illustrate the per-
formance and limits of the overall tracking algorithm. It is
shown that range, velocity, and angle of up to five humans are
correctly detected and tracked. The presented one-class classifier
successfully distinguishes human targets from other quasi-static
targets like trees and shadowing effects of human subjects on
walls.
Index Terms—Millimeter wave radar, radar detection, radar
tracking.
I. INTRODUCTION
AUTOMATIC detection of human targets offers a vari-
ety of applications in modern society. In smart homes,
human target detection can be utilized in light sensors for
instance to automatically switch the light on or off if a person
enters or leaves the room, respectively. This idea can be
enlarged to smart cities with smart street lighting. In the
industrial area, fast detection of humans helps robots to avoid
collision and security systems to find intruders. Also in the
automotive area, there is always a need for enhanced safety
and security, especially regarding autonomous cars. Human
target detection in general is enabled by advanced sensor
systems consisting of a hardware platform and a dedicated
firmware, which adjusts the component settings and gathers
raw sensor data. Depending on the system concept, either the
firmware itself pre-processes these data and evaluates them by
algorithms or it communicates with a host processor, that takes
charge of the signal processing.
All authors are with Infineon Technologies AG, 85579 Neubiberg, Germany.
The authors declare no conflict of interest.
Various sensor concepts can be used for human target detec-
tion. The most obvious approach is to use a camera together
with an image processing and feature detection algorithm [1],
[2], [3]. Advanced algorithms like a combination of temporal
differencing and template matching enable tracking of the
classified targets [1]. Special RGB-D cameras improve the
performance by utilizing the additional depth information [4].
Next to these single camera systems, also distributed camera
systems for three dimensional tracking have been published
[5]. All camera-based systems suffer from high costs and the
lack of privacy. Alternative sensor concepts use acoustic waves
[6], [7], millimeter waves [8] or non-visible light [9], [10]
and therefore ensure more privacy. Apart from camera-based
sensor fusion [9] and lidar systems [11], [12], which have
a high accuracy and precision, these sensors are additionally
considerably cheaper. This paper focuses on radar-based sys-
tems as their electromagnetic waves have some advantages
compared to the other concepts. While ultrasonic systems
as well as cameras require line-of-sight, pulsed, broadband
electromagnetic waves can penetrate walls [13]. In contrast
to infrared or lidar sensors, radar systems also work in harsh
environments and are less affected by rain, dust or fog [14].
Besides, they can be aesthetically hidden without affecting
the operating performance. Furthermore, radar systems can
detect micro motions like vital signs by interferometric signal
evaluation even in the micrometer range [15], [16]. Multiple
targets can be detected by signal modulation [8] or multiple-
input multiple-output (MIMO) systems [17].
Radar-based human tracking has been researched for several
years meanwhile, using diverse radar concepts. Frequency-
modulated continuous wave (FMCW) [18], [19], [20], [21]
and ultra-wideband (UWB) [8], [22] radar sensors use multiple
frequencies to obtain distance information of one or more
targets. For short-range human tracking up to 5 m, a 24-
GHz FMCW radar based on Six-Port interferometry [14]
utilizing vital-Doppler standard deviation was published [18],
[19]. Other FMCW radars use inverse synthetic aperture radar
(ISAR) images for indoor tracking of a single person [20]
or a Kalman filter with track management table and adaptive
thresholding for multi-target tracking [21]. A Kalman filter is
also used in UWB radar-based multi-target tracking combined
with an expectation maximization algorithm [8]. The UWB
radar in [22] uses a motion filter and tracking point cohesion
to track one target even behind a wall. Another common
approach is to combine several FMCW [23], [24] or UWB
[25], [26], [27] radar sensors to a multistatic radar system.
Target detection by motion filtering or an ordered statistics
IEEE SENSORS JOURNAL 2
constant false alarm rate (CFAR) detector each combined with
tracking point cohesion enables the tracking of one person
through a wall [23] or three humans in a cluttered environment
[24], respectively. Two UWB receive (RX) radar sensors are
used each for single target tracking in Cartesian coordinate
system [25], indoor human tracking [26], and through-wall
tracking of two persons [27]. Alternative approaches use
pulse-Doppler radar together with an alpha-beta filter for
real-time embedded tracking of two targets [28], synthetic
aperture radar (SAR) with ground moving target indication for
automatic target recognition from airplanes [29], MIMO radar
with two-dimensional MUSIC (MUltiple SIgnal Classification)
correlation for indoor multi-path environment [30], and passive
acoustic radar with moving target detection and tracking to
measure coordinates of a flying multicopter [31]. 2D tracking
of two humans by an unmodulated CW radar sensor is enabled
by three RX antennas and the evaluation of the direction of
arrival (DOA) [32].
Radar-based classification is an arising topic, while most
publications focus on specific human activities [33] or gestures
[34] and on target recognition of pre-defined classes like small
drones and birds [35]. A rather open research topic is human
target classification, which means to distinguish humans from
other targets with radar systems. Instantaneous presence detec-
tion using Doppler radar detects if a human is present or not
[36]. Other publications limit the targets to pre-defined classes
like ”one person”, ”two persons”, and ”vehicle” [37]. In [38]
radar classification of indoor targets to distinguish human tar-
gets from harsh clutter using through-the-wall radar systems is
investigated. Doppler spectrogram-based features are utilized
in [39] to distinguish humans from dogs, bicycles, and cars.
All of these publications pre-define two or more specific target
classes, some combine several targets in one non-human class
to generate a two-class problem. Most of these classifiers are
based on a support vector machine (SVM) [33], [35], [38],
[36], [39], since an SVM is a convenient classifier for binary
problems like human classification. The disadvantage of a two-
class SVM with one class for humans and the other for non-
human targets is the required pre-selection of the non-human
targets, which are used to train the classifier. The classification
for un-trained targets is not predictable and can easily result
in false human target detection. This issue can be solved by
one-class classifiers for outlier detection. A one-class SVM
is trained only with data of the desired target class, in the
presented human targets. With appropriate features, all non-
human targets should be detected as outliers. Examples for
one-class SVMs from literature are image-based handwritten
signature verification [40], sheet metal forming limit detection
with camera [41] or oil spill detection using airborne SAR
[42].
This paper presents an FMCW radar sensor for multi-
target tracking with all algorithms to detect and track up to
five humans running on a microcontroller. In comparison to
the state-of-the-art, the researched signal processing routine
is tailored and optimized to detect and track more than
three targets in an outdoor environment within an angular
range of over 100◦and a distance up to 20 m. While some
parts of the proposed signal processing routine are smartly
combined already-known procedures, two novel algorithms are
introduced to ensure high robustness against losing tracks and
assigning ghost targets. First, an advanced angle estimation
algorithm for humans targets is proposed, since common angle
estimation using the range-Doppler peak often results in wrong
estimates as the individually moving parts of a human body
induce different angles even and especially for (quasi-)static
humans. The presented algorithm uses the centroid of the
Doppler spread along the slow-time and utilizes a Doppler-
compensation for static humans. Second, a one-class SVM
classifier with handcrafted features is introduced to distinguish
quasi-static humans from other static but slightly moving
targets (like trees in outdoor scenarios) to mitigate false human
target detection. To the author’s knowledge, the proposed one-
class classifier is the first radar-based approach to distinguish
humans from any other target without using pre-selected
target classes. The complete proposed algorithm pipeline is
simultaneously optimized with respect to compactness and
energy-efficiency to enable an implementation on an ARM M4
cortex real-time embedded platform. The implementation of all
signal processing routines on the micro-controller along with
measurements in complex outdoor scenarios shifts laboratory
research to real-world scenarios and underlines the feasibility
of the proposed system.
The paper is structured as follows. In Sec. II, the challenges
of this research are described. The hardware and firmware
concept for radar-based human target detection and tracking
are designed on system level in Sec. III. The sections IV, V,
and VI explain the algorithms for human detection, tracking,
and classification, respectively. The measurement results are
shown in Sec. VII and compared to the state-of-the-art in
Sec. VIII. A conclusion is given in Sec. IX.
II. CHALLENGES OF HUMAN DE TE CTION & TRACKING
Human detection and tracking with radar systems is of
high interest regarding consumer, industrial, and automotive
applications. However, a robust detection of humans in a noisy
clutter environment is a challenging topic. Major challenges
are multi-path reflections especially in bounded environments
and the differentiation of human-related movements from non-
human movements like animals or trees. Since millimeter-
wave radar systems are very sensitive to slight motions, multi-
path reflections from static objects and walls induce ghost
targets at their positions. It is a challenging task to distinguish
human targets from these ghost targets.
Furthermore, in contrast to completely static targets like
walls or clutter, multi-path effects and speckle lead to high
variations of the received signal strength reflected from a hu-
man [43]. Destructive interferences can temporarily extinguish
the backscattered signal from the human target. In this case,
the signal strength of the human target can be way below
the detection threshold for several consecutive measurements,
whereby target information is lost. Thus, in order to be able to
track a human target, it is required to account for this issue.
The reflected signal from targets like walls always possess
high signal strength as compared to human targets and hence,
for a human target being very close to such targets, the signal
IEEE SENSORS JOURNAL 3
strength of a human target is easily overshadowed by such
strong targets. Therefore, it is also necessary to diminish or
extinguish the backscattered signals from large static targets
by intelligent signal processing routines.
III. SYS TE M DESIGN
The hardware and firmware concepts of the proposed radar
system are designed in this section.
A. Hardware Concept
Figure 1 shows the block diagram of the microwave multi-
channel FMCW radar sensor designed to work in the 24-GHz
ISM band. The sensor consists of a radio frequency (RF) unit
formed by a 24-GHz monolithic microwave integrated circuit
(MMIC) and microstrip patch antennas, a low noise fractional-
N phase-locked loop (PLL) for frequency stabilization and
FMCW ramp generation, a programmable gain low noise
analog baseband (ABB) section and a digital part consisting
of Cortex-M4-based signal processing and control unit. The
sensor is powered by 5V direct current (DC) supply and on-
board low noise low dropout regulators (LDOs) are used to
supply each of the sensor sections. The complete radar system
is fabricated on a 10-mil-thick Rogers 4350B substrate-based
single printed circuit board (PCB) comprising six metal layers.
The MCU on board is programmed via a serial-wire debug
(SWD) interface using an external breakable debugger unit not
shown in the block diagram. The sub-sections below describe
each of the functional blocks in detail.
1) 24-GHz Front-end MMIC: The RF unit consists of the
24-GHz multichannel transceiver MMIC BGT24MTR12 with
one transmit and two receive sections, impedance matching
structures, harmonic filters, and microstrip patch antennas. The
MMIC is designed in a Silicon-Germanium (SiGe) Bipolar
technology [44] with ft/fmax of 200/250GHz from Infineon
Technologies and is enclosed in a small 32-pin VQFN package.
The transceiver consists of a voltage controlled oscillator
(VCO) with prescaler outputs for frequency generation and
control, a transmitter (TX) chain including a power amplifier
(PA) and buffer amplifiers for the TX and local oscillator (LO)
sections, and two homodyne quadrature receiver sections with
low noise amplifiers (LNA) and highly-linear modified gilbert
type mixers. The multiple receive channel configuration en-
ables angle-of-arrival (AoA) estimation of the signal reflected
from the radar target.
The VCO is a free-running, fundamental frequency oscilla-
tor with a low phase noise of -85 dBc/Hz at 100-kHz offset.
It is controlled by two tuning inputs, one for coarse pre-
adjustment (COARSE) and one for fine-tuning (FINE). The
VCO is followed by two prescaler blocks. The first prescaler
has an output frequency of 1.5 GHz and the second prescaler
has a 23-kHz square-wave output. The prescalers enable to
interface the VCO with an external frequency control circuit
like a PLL or a digital-to-analog converter (DAC).
The TX section includes a power amplifier with differential
output. Its typical output power is +11 dBm. A part of the
TX signal is used as LO signal for the mixers in the receiver.
A passive poly-phase filter (PPF) generates quadrature sig-
nals for the mixers LO input. The receiver sections have a
single-sideband noise figure (NFSSB) of 12 dB and a voltage
conversion gain (CG) of 26dB. Figure 2 shows the perfor-
mance characteristics of the transceiver over temperature at
24.125 GHz.
Depending on the target distance and radar cross section
(RCS), the signal received at the sensor may completely
saturate the analog-to-digital converters (ADCs) in the signal
processing unit. Therefore, depending on the strength of the
received signal, the transceiver needs to reconfigure the TX
output power and the RX conversion gain. Integrated power
sensors based on RF to DC converters at the TX ports enable
to monitor the output power and adjust it via a 3-bit serial
peripheral interface (SPI) setting. The transmitter output power
can be reduced by 9 dB from its maximum value. On the
receiver side the gain of the LNA can be reduced by a
single gain-step of 5 dB. Further gain configuration at the
receiver side is provided by programmable gain amplifiers
(PGAs) in the intermediate frequency (IF) stage as described
in Sec. III-A3. The built-in modified Gilbert-based quadrature
downconversion mixers help to achieve a high input 1-dB
compression point (IP1dB) of -12 dBm and convert the RF
signal directly to zero-IF. An on-chip temperature sensor
constantly monitors the temperature of the MMIC. The chip
is controlled via a simple 3-wire SPI. When fully enabled, the
chip consumes approximately 700 mW from a 3.3-V power
supply. Since the MMIC does not support a power down mode,
the transceiver is connected to the power supply over a PMOS
switch, which enables to operate the sensor in duty-cycle mode
for low power applications.
The transmitter output of the MMIC is differential. These
differential outputs are first connected over matching structures
followed by a Wilkinson power combiner on the PCB. The
matching structures compensate for the bond wire inductance
and other parasitic effects due to the VQFN package. Fol-
lowing the power combiner, a microstrip harmonic filter is
used to attenuate the harmonics around 48 GHz. The harmonic
filter provides an attenuation larger than 20 dB and has a
simulated loss of approximately 0.5 dB. The filter path then
goes over a DC block and a feedthrough via to the other
side of the PCB to the antennas. The simulated loss for the
entire RF section connecting the TX output from the MMIC
to the antennas on the other side of the board including the
vias is approximately 2 dB. There are DC shorts before the
feedthrough vias for enhanced electrostatic discharge (ESD)
protection. The receiver input of the MMIC is single-ended.
The RX input is connected over a matching structure, a DC
block and a feedthrough via to the antennas on the other side
of the board. The simulated loss for the entire RF section
connecting the RX input at the MMIC to the antennas on the
other side of the board including the vias is approximately
1 dB. Also in this path, DC shorts before the feedthrough
vias enhance ESD protection. All feedthrough vias for ESD
protection go from top layer to the adjacent inner ground layer.
They are visible in the photography of Fig. 3, which shows
the top and bottom view of the sensor module. The size of the
module is 45 mm by 50 mm in dimension.
IEEE SENSORS JOURNAL 4
Fig. 1. 24-GHz multi-channel FMCW radar system.
A tapered series-fed 1x6 linear antenna array with low side
lobe levels is used for the transmission and reception of the
microwave signal. Figure 4 shows the measured transmit field
of view of a sensor module. The sensor has a typical effective
isotropic radiated power (EIRP) of +18 dBm and a 3-dB half
power beam width (HPBW) of 19◦and 76◦in the E- and
H-plane, respectively.
2) Phase-Locked Loop : A fractional-N PLL is used to
stabilize the VCO on the 24-GHz MMIC and generate the
frequency ramps for target tracking. The PLL consists of a
phase detector with programmable frequency dividers for both
the high frequency RF input signal and the low frequency
reference clock signal. The 1.5-GHz prescaler output from the
MMIC is connected to the high frequency RF input of the PLL
and an external low phase noise 40-MHz reference oscillator
clock is connected to the reference input. The phase detector
compares the phase differences between the two input signals
and generates a voltage proportional to the phase error which is
converted to a correction current by a charge pump. The output
of the charge pump is filtered with a third order passive loop
filter formed by discrete components and given to the tuning
port of the VCO. For the aimed application, the coarse and
fine tuning pins of the MMIC are tied together to minimize
the tuning sensitivity. The loop filter is designed to support
frequency ramps up to 300 us and at the same time fulfill the
ETSI and FCC specifications for unwanted emissions.
3) Analog Baseband Section: Depending on the target in
front of the antennas, the analog IF output signal from the
MMIC can be very low in amplitude (µV to mV range). To
process these low amplitude signals it is necessary to amplify
the IF signals with analog amplifiers. Another commonly
occurring issue in a homodyne architecture-based FMCW
radar system is the feedthrough of the TX signal into the
RX part due to limited isolation on the PCB, also called TX-
to-RX leakage. Consequently, there is always a dominating
low frequency component at the receiver output of the radar
IC. The value of this low frequency component depends on
the value of the frequency ramp settings. This low frequency
signal is further amplified by the baseband amplifiers and
may completely saturate the radar IF chain. This effect is
inherent to all FMCW radar systems and cannot be eliminated
completely in the analog domain. The TX-to-RX leakage also
limits the minimum distance that can be measured by the
radar. However, the effect of this effect can be minimized
using appropriate filtering. This requires the implementation
of filtering stages prior to the amplification of the IF signal in
the baseband section. Therefore, the baseband section of this
sensor is designed to have a bandpass characteristic.
The MMIC provides both in-phase and quadrature (I/Q)
phase IF signals from each of its receive channels. Since
IEEE SENSORS JOURNAL 5
-50 0 50 100 150
10
11
12
Temperature (◦C)
Power (dBm)
(a)
TX output power
-50 0 50 100 150
20
25
30
Temperature (◦C)
Gain (dB)
(b)
IRX1 QRX1 IRX2 QRX2
-50 0 50 100 150
5
10
15
Temperature (◦C)
Gain (dB)
(c)
IRX1 QRX1 IRX2 QRX2
Fig. 2. (a) Transmitter output power, (b) CG, and (c) NFSSB of the
BGT24MTR12 over temperature at 24.125 GHz.
the I/Q signals are provided differentially from the MMIC,
there are four IF output signals per receive channel: IFI,
IFIx, IFQ, and IFQx. Each IF path comprises two stages
of low noise analog baseband amplifiers with a bandpass
characteristic. The first stage consists of two dual-channel low
noise operational amplifiers with a gain of 23.5 dB, which also
perform the differential to single-ended conversion of each IF
signal pair, thereby reducing the eight IF signal lines from the
MMIC to four. The second amplification stage consists of four
programmable gain amplifiers for each of the IF signals output
by the first amplification stage. The PGAs in the IF section
(a) (b)
Fig. 3. (a) Top and (b) bottom view of the fabricated sensor module.
Fig. 4. Measured radiation pattern in E- and H-plane of the sensor.
10 100 1000 1e4 1e5 1e6 1e7
-60
-40
-20
0
20
40
60
80
Frequency (Hz)
Voltage Conversion Gain (dB)
G= 1 G= 2 G= 4 G= 8
G= 16 G= 32 G= 64 G= 128
Fig. 5. Simulated transfer characteristics of the analog baseband for various
PGA gains.
along with the configurable gain of the MMIC LNAs enable
an automatic gain control of the receive section based on the
received signal strength from a target. The second IF stage
in combination with the first IF stage provide a total IF gain
of 65.5 dB with a 3-dB bandwidth from 13 kHz to 105 kHz.
Figure 5 shows the simulated transfer characteristics of the
analog baseband section for different PGA gain settings. For
the current sensor implementation, the desired IF frequencies
appear between 10 ... 100 kHz depending on the target distance
from the sensor.
4) Signal Processing and Control Unit: The sensor plat-
form uses the 32-bit ARM Cortex-M4 144-MHz microcon-
troller unit (MCU) XMC4700 in a 194-pin BGA package from
Infineon Technologies, to perform the radar signal processing.
The MCU takes care of communication with all sub-systems
on the sensor module, enables data acquisition, performs the
complete radar signal processing including sampling and FFT,
and sends the results via its universal asynchronous receiver
transmitter (UART) or universal serial bus (USB) interface
to an external device. Four 12-bit versatile ADCs (VADC)
in the MCU implement the radar signal sampling on the IF
signals from the PGAs and also acquire various sensor data
from the MMIC. With a 2048-kB flash and 352-kB RAM,
the MCU enables highly accurate range-Doppler processing.
Several general purpose input/output (GPIO) pins from the
MCU are available on the header pins of the sensor module
to interface with external circuits.
IEEE SENSORS JOURNAL 6
Fig. 6. Concept of the FMCW frame format.
TABLE I
FRA ME CO NFIG UR ATION PA RAM ET ERS .
Number of chirps 16
Pulse repetition time 500 µs
Up-chirp time 300 µs
Samples per chirp 128
Range FFT size 256
Doppler FFT size 32
B. Firmware Concept
The radar firmware is based on a frame concept, whose
configuration is illustrated in Fig. 6 and whose parameters
are listed in Tab. I. Each frame has a duration of 150 ms and
starts with enabling the PLL for 3 ms, followed by NC= 16
identical frequency chirps with a repetition time of 500 µs.
After this 11-ms on-time, the BGT and PLL are disabled
for power saving. The single chirps consist of an up-chirp
of 300 µs, a down-chirp of 100 µs, and a waiting time of
100 µs. Although only the up-chirps are sampled for signal
processing, a down-chirp instead of a frequency jump is used
to prevent illegal spurs in the adjacent frequency bands. The
waiting time before a new up-chirp starts ensures a correct
settling of the start frequency. All frequency changes are
automatically controlled by the pre-set PLL and all frequency
ramps are in the range of 24.025 ... 24.225 GHz. The closest
25-MHz bands at the ISM band limits are not used to avoid
regulatory violations in the adjacent frequency bands. The
used bandwidth B= 200 MHz results in a theoretical range
resolution ∆R=c
2B≈60 cm, whereas the real resolution is
typically degraded due to applied windowing. During the PLL
off-time, the algorithm is processed on the microcontroller
within the first approximately 36 ms and the remaining time
serves for host communication.
A flow chart of the complete algorithm is shown in Fig. 7.
Correct raw data acquisition as the first step is ensured by
setting the start time and sampling frequency of the ADC so
that the required NS= 128 samples are equally distributed
within ramp start and end. For each frame, a three-dimensional
data cube Φ∈ CNS×NC×NRX containing the complex-valued
baseband signals is obtained. The first dimension contains
all samples per chirp (fast-time) for range estimation, the
second dimension belongs to the different chirps per frame
(slow-time) for velocity estimation, and the third dimension
corresponds to the NRX = 2 RX antennas for AoA estimation.
In the second step, a fast-Fourier transform (FFT) is applied
to the fast-time dimension with a zero padding of factor two.
The resulting FFT is called range FFT and has a size of
256 values. Subsequently, coherent pulse integration along the
slow-time dimension aims to improve the signal-to-noise ratio
and moving target indication (MTI) filtering is used for clutter
suppression. The loop expresses the mutual influences over
consecutive frames. Both signal processing routines as well
as all other steps mentioned in the following are explained
in detail in Sec. IV. In the next step, up to five human
targets are detected in the range dimension by peak detection
utilizing an adaptive threshold. The targets’ velocities are
estimated by the so-called Doppler FFT along the slow-time
dimension for the corresponding range bins. Zero padding
with a factor of two is also applied to this FFT resulting
in a size of 32 values. Computing the Doppler FFT solely
for the target indices reduces the computational complexity
with regard to the implementation on the microcontroller. This
range-Doppler processing is done for both antennas, for which
reason the corresponding peak pairs are used for subsequent
AoA estimation to determine the angle of each target to the
radar sensor. In the last step of each frame, the calculated target
parameters are fed to the tracking algorithm, which takes care
of track association, track management, and track filtering.
Due to its low complexity, the chosen tracking algorithm is
based on alpha-beta filtering, which is explained in Sec. V.
IEEE SENSORS JOURNAL 7
Raw Data Acquistion
Range FFT
Coherent Pulse Integration
Peak Detection
Doppler FFT
AoA Estimation
Target Tracking
MTI Filtering
Fig. 7. Concept of the tracking algorithm.
IV. HUMAN DET EC TI ON
This section investigates three signal processing routines for
human target detection to be implemented on the microcon-
troller. While moving target indication filtering and chirp data
integration are well-known concepts from literature, Doppler-
compensated angle-of-arrival estimation is a newly proposed
algorithm for enhanced angle estimation of quasi-static human
targets.
A. Moving Target Indication Filtering
TX-to-RX leakage is a common challenge in radar systems
and is also present in the Infineon BGT24MTR12 radar chip.
Due to the bandpass filtering in the baseband, this leakage
leads to a low-frequency component, which results in high
amplitudes in the first range FFT bins. Next to extensive
hardware-related cancelation methods [45], [46], there are
also software-based approaches to remove this leakage. One
possible software solution is to calibrate the system before
usage measuring the system itself by placing an absorber
in front of the antennas to generate a reflection-free envi-
ronment. At runtime these calibration values are subtracted
from the measurement data. On the one hand, this method
simultaneously extinguishes other impairments of the radar
hardware, on the other hand, the calibration values are highly
temperature dependent. Another software solution is MTI
filtering [47], which is a type of an in-situ calibration and
therefore temperature independent.
MTI filters in principle are low-order, simple finite impulse
response (FIR) designs. At each time stamp the absolute
maximum value over slow time of each range bin is denoted
by ri,max. The MTI filter value tiis the weighted average of
this maximum value and the previous MTI filter value ti−1
with a weight of α:
ti=α·ri,max + (1 −α)·ti−1.(1)
In the first time stamp t0is initialized with Zero. For each
range bin, the previous MTI filter value is subtracted from
ri,max to obtain the filtered range FFT value ri,filt:
ri,filt =abs(ri,max −ti−1). (2)
This filtered value is then utilized for the subsequent target
detection. MTI processing performs a linear filtering which
leads to a diminished signal strength for static targets while
maintaining the signal strength of the moving targets. There-
fore, MTI processing helps to remove completely static targets,
while non-static targets like humans are retained.
B. Chirp Data Combination
After computing the range FFT for each chirp and applying
the MTI filter it is required to combine the data over all
chirps of a frame for subsequent target detection. One possible
data combination strategy is coherent integration [47], which
combines the phase and magnitude of the range FFT data
coherently over all chirps of a frame. Coherent integration
is based on the mean Rmean of the range FFT data riover
NCchirps:
Rmean =1
NC
·
NC
X
i=1
ri(3)
ri=ai·ejφi(4)
Here, amplitude and phase of the range FFT value for chirp i
are denoted by aiand φi, respectively. This method enhances
the signal-to-noise ratio (SNR), but is has to be considered that
all phase values of all chirps have to be aligned, since phase
misalignment leads to signal distortion. Another approach is
to find and use the chirp iwith the maximum absolute range
FFT value Rmax over all NCchirps:
Rmax =NC
max
i=1 (yi)(5)
This approach assures that the FFT data with highest am-
plitudes is chosen for target detection and prevents phase
misalignments. Both approaches are evaluated in Sec. VII,
but only the variant using the maximum absolute value is
implemented in firmware due to the enhanced signal quality
for target detection and the necessity of phase alignment.
C. Doppler-Compensated Angle-Of-Arrival
In case of human targets, the individually moving body
parts like torso or shoulders induce different velocities and
(wrong) angles in the Doppler FFT. To overcome this issue,
the proposed procedure uses the centroid of the Doppler spread
along the slow-time for angle estimation. Furthermore, the
vital signals of static and quasi-static humans modulates the
data along slow-time, which results in high variations in the
angle estimation. In the presented approach these modulations
are compensated to stabilize the angle estimations.
The basic approach to determine the AoA is to evaluate
the phase difference between both RX antenna beams, also
referred to as phase-comparison monopulse. The AoA θis
calculated by geometrical considerations based on an incoming
plane wavefront with
θ= arcsin( λ
2π·∆φ
d), (6)
where dis the geometric distance between two antennas, ∆φ
is the observed phase difference between both RX signals, and
Citations (15)
References (50)
... However, any form of people surveillance with cameras raise privacy concerns and also underperform in indoor scenarios with limited illumination. Even though radar-based solutions [8] [9][10] [11][12] [13] are immune to such limitations but suffer from issues such as missed detection caused due to occlusion, low resolution data which lowers the chances of target identification and time-varying radar signal strengths caused due to superposition of reflections coming from different body parts. These challenges makes the use of short-range and low-cost radars using conventional approaches for people counting solutions in dense scenarios ineffective. ...
... The FMCW radar chipset BGT60TR13C from Infineon Technologies AG in Fig. 2 has been used for the proposed solution. The most commonly used mode for FMCW radars are to use sequence of frequency chirps with short ramp-times, delays between chirps to save power and at the end of the sequence of chirps for data acquisition and processing [8]. The frequency chirps with bandwidth of 1.0 GHz within the 60-GHz band and pulse repetition time of 400 µs was used. ...
Conference Paper
Full-text available
Apr 2020
Souvik HazraShow abstract
... In [33], Multiple-Input-Multiple-Output (MIMO) architectures have been employed for coherently combining the same signal with different phases, thus greatly improving the target sensing. In [34], multiple human targets have been detected and tracked by using a multireceiver radar board, exploiting an angle estimation algorithm to ensure high robustness against false alarm and ghost targets. In [35], Doppler processing and spectral estimation are concurrently exploited for detecting the vital signs of multiple targets located at distances shorter than the radar spatial resolution, eliminating the mutual interference by means of a range integration algorithm. ...
Article
Full-text available
Sep 2020
Show abstract
... Radar has evolved from automotive applications such as driver assistance systems, safety and driver alert systems, and autonomous driving systems to low-cost solutions, penetrating industrial and consumer market segments. Radar has been used for perimeter intrusion detection systems [3], gesture recognition [4] [5] [6], human-machine interfaces [7], outdoor positioning and localization [8], and indoor people counting [9]. ...
Chapter
Sep 2020
Show abstract
... The amount of acquired data for this class remained the same as for the previous ones. Our objective with class i) was to address the problem of distinguishing humans from quasi-static objects, which is a known challenge in the area of human target detection [31]. ...
Article
Jun 2020
Show abstract
... Although limited cyclic motion can be canceled at close range under constrained conditions using interferometric and multistatic techniques [55]- [62], such approaches greatly increase in cost and complexity for observation at longer standoff. Recent demonstrations introducing the capability to track people in motion implicitly require the subjects to come to rest in order to detect and interpret vital signs [54], or else simply detect the presence of vibration without the ability to analytically reconstruct it [4], [63]. Airborne motion compensation techniques leveraging on-board VOLUME XX, 2017 1 inertial measurement and/or global navigation sensors [47]- [52], [64] can correct for the dynamics of the interrogating radar but not for the dynamics of the target. ...
Article
Full-text available
Feb 2020
Show abstract
Thesis
Aug 2020
Show abstract
Thesis
Aug 2020
Show abstract
Chapter
Jan 2021
Show abstract
Conference Paper
Mar 2020
Article
Full-text available
Apr 2020
Show abstract
Show more
Skip Ad
Advertisement
Recommendations
Project
Tobias Steigleder
Christoph Ostgathe
Christoph Will[...]
Sven SchellenbergerRadar-based (no touch) measurement of heart-beat, breathing and autonomic nervous System function. GUARDIAN's goal is to develop a six-port radar system which allows to reliably measure heart-beat ... [more]View project
Project
Avik Santra
Souvik Hazra
Raghavendran Vagarappan Ulaganathan[...]
Developing scalable and robust industrial and consumer applications of mm-wave radar using deep learning and signal processing techniquesView project
Project
Abhiram Chakraborty[...]
Designing Smart Sensing systems with ultra low-power, low noise and small form factor 24GHz MMICs.View project
Project
Kaushal Jadia[...]
Conference Paper
December 2007
Firstly, the echo model of moving target for SAR is constructed in this paper. According to this model, the method of detection of moving targets based on Matched Fourier Transform is discussed. Compared with the bilinear time-frequency distribution algorithms, the Matched Fourier Transform will not be affected by the cross-terms when multiple moving targets exist. Moreover, because of remaining ... [Show full abstract]Read more
Conference Paper
November 1997
This paper presents a new configuration of Synthetic Aperture Sonar (SAS): the squint SAS where the sonar beam is not perpendicular to the sonar track. Though worse than in the ideal case of a side looking SAS, the azimuth resolution of a squint SAS is in most cases better than the azimuth resolution of a conventional beam. This point is interesting in mine hunting because it allows, in a forward ... [Show full abstract]Read more
Conference Paper
Full-text available
April 2019
Nonlinear frequency modulated (NLFM) pulse compression waveforms have become a mainstream methodology for radars across multiple sectors and missions, including weather observation, target tracking, and target detection. NLFM affords the ability to generate a low-sidelobe autocorrelation function and matched filter while avoiding aggressive amplitude modulation, resulting in more power incident ... [Show full abstract]View full-text
Article
August 2012
To meet the public security and criminal investigation requirement, this paper studies detecting and tracking of human motion through related theories on image processing and pattern recognition. The video information is obtained from video monitoring system (VMS). Background modeling is based on mixture Gaussian model. The background subtraction is used for detecting moving targets, separating ... [Show full abstract]Read more
