1

A Survey Object Classification Using FMCW Radar

and Learning-Based Methods
Abdul Wase Khan (539017), Member, IEEE, Muhammad Shayan (536722), Member, IEEE, Mehran Rizwi
(539521), Member, IEEE, Mohammad Bilal Abbasi (539754), Member, IEEE MS SEECS, National University
of Sciences and Technology (NUST)
1st Semester Students

Abstract—Frequency Modulated Continuous Wave (FMCW) are estimated [2]. In contrast to continuous-wave (CW) radar,
radar has emerged as a robust sensing modality for human which is only able to estimate velocity, or pulsed radar, which
and object classification due to its resilience to adverse weather is only able to estimate range.
conditions, independence from illumination, and inherent privacy
preservation. This work presents a comprehensive survey and With FMCW radar, it is possible to extract several parame-
comparative analysis of learning-based FMCW radar classifi- ters of the targets at the same time. Recent mmWave FMCW
cation approaches. Classical signal-processing and handcrafted radars working in 24 GHz and 7781 GHz have high resolution,
feature-based methods are first reviewed, followed by recent deep small size, and low-power capabilities that allow them to be
learning techniques using radar-derived image representations. used in embedded and automotive systems.
Key challenges, including clutter, low signal-tonoise ratio targets,
limited generalization, and hardware constraints, are systemati- Early FMCW radar concentrated on mainlydetection and
cally identified. To address these challenges, multiple learning- tracking but recent studies have changed the current research
based classification frameworks are evaluated through prac- to re- liable classification of humans and objects. Conventional
tical studies, including ResNet-18, YOLOv8n, EfficientNet-B0, signal- processing-based methods have been developed on the
and MobileNetV3-RD architectures. All models are assessed on
basis of thresholding, cluster- ing, and handcrafted features of
FMCW radar–derived datasets for binary human–bicycle classi-
fication. Experimental results demonstrate that while YOLOv8n range-Doppler or micro-Doppler images. Nevertheless, these
achieves the highest numerical accuracy, its performance is methods tend to lack in generalization and need features that
influenced by dataset limitations and increased computational are developed by experts. New opportunities have arisen due
complexity. Lightweight models such as EfficientNet-B0 and to the development of machine learning and deep learning,
MobileNetV3-RD offer improved efficiency and suitability for
which enables the improvement of radar-based classification
embedded deployment, with competitive accuracy. Among all
evaluated approaches, the fine-tuned ResNet-18 model provides through the learning of discriminative representations using
the best overall trade-off between accuracy, robustness, and radarderived data. The current paper is an extension of this
generalization. The findings confirm that radar image–based deep progress by surveying the currently used FMCW radar-based
learning combined with transfer learning offers a practical and classification framework and assessing a deep learning-based
scalable solution for reliable FMCW radar-based human and
system that applies to binary human-object classification [3].
object classification.
Index Terms—FMCW radar, human detection, object classifi-
cation, mmWave radar, deep learning, ResNet, range FFT. II. L ITERATURE R EVIEW
A fundamental aspect of the research on FMCW radar-
I. I NTRODUCTION based Human and object classification has attracted a lot
One of the key needs in the contemporary sensing ap- of research interest concerning the increasing need to have
plications is human and object classification in autonomous robust, all weather and privacy preserving sensing systems.
vehicles, intelligent surveillance systems, smart environments, FMCW radar is not, as opposed to optical sensors, sensitive
and safety monitoring. RGB cameras Vision based systems to the illumination conditions and can be reliably used in
are capable of giving rich spatial information but perform the conditions of low visibility like fog, dust, rain and other
poorly in case of poor light, bad weather conditions as well as adverse conditions [4].All these benefits have made radarbased
occlusions. Moreover, camera-based solutions present severe sensing an important enabling technology in such uses as
privacy concerns and hence they are not extensively adopted autonomous driving, intelligent surveil- lance, healthcare mon-
in most real-life situations. Such issues have inspired the study itoring, and smart environments. The initial studies in FMCW
of other modalities of sensing that are robust, reliable, and not radar were mainly target detection and tracking, but not
privacy invading [1]. semantic classification. The traditional radar signal processing
Frequency Modulated Continuous Wave (FMCW) radar has methods that were commonly used to estimate the range and
become one of the solutions because of its capability to work velocity of the target were Fast fourier Transform (FFT),
in the absence of lighting condition as well as being strong in Constant False Alarm Rate (CFAR) detection, Moving Target
fog, rain, dust, and darkness. FMCW radar uses a frequency Indication (MTI) and Kalman filtering. Information on range
signal that is continuously varying and processes reflected is generally acquired through onedimensional FFT processing
echoes such that target range, velocity and angular information of the intermediate-frequency (IF) signal and velocity can
 2

be determined through Doppler processing of a series of a similar way, with handcrafted feature-based methods applied.
chirps. Despite these techniques being useful in identifying Although these techniques are computationally fast and sim-
the presence of the targets and determining their kinematic pler to apply than deep learning models, their performance
attributes, they do not give much information on the identity relies on expertly crafted features and intensive fine-tuning,
of the targets and semantic class [5]. Sophisticated methods which means that they do not generalize to unknown targets
like the Synthetic Aperture Radar (SAR) and Inverse Synthetic and also to different environmental conditions [6]. The recent
Aperture Radar (ISAR) have also been considered to be developments in radar equipment and signal processing have
effective in object recognition, but these methods tend to take made it possible to create image-like representations, including
long time observations, controlled movements and complicated range-angle and range-Doppler heatmaps. These images offer
hardware setup and as such are not as effective in real time the spatial context and enable working with radar data with the
human and object classification in dynamic environments [3] help of computer vision methodology. Gupta et al. used range-
. angle images based on the mmWave FMCW radar data and
Human detection has been a central theme in FMCW used machine learning and deep learning models to classify
radar research due to its importance in safety- and security- various targets, which had high classification accuracy.
critical applications. Will et al. developed a 24 GHz FMCW Deep learning has emerged as a powerful alternative to tra-
radar system capable of detecting, tracking, and classifying ditional feature-based methods by enabling automatic learning
multiple human targets in outdoor environments. Their ap- of discriminative features directly from radar-derived images.
proach combined classical radar signal processing with a one- Convolutional Neural Networks (CNNs) have been success-
class Support Vector Machine (SVM) to distinguish humans fully applied to range–Doppler maps, spectrograms, and other
from clutter and other objects. While the use of a one-class radar image representations, consistently outperforming classi-
classifier addressed the challenge of unseen target signatures, cal machine learning approaches in complex and cluttered en-
the method relied heavily on handcrafted Doppler- and range- vironments. Transfer learning has further reduced the data re-
based features, which limited its scalability and adaptability quirements for training deep models by fine-tuning pretrained
across different environments and radar configurations. CNN architectures such as ResNet, VGG, and MobileNet on
etection because it is used in safety and security-sensitive radar datasets. Despite these advances, much of the existing
applications. Will et al. designed a multi-target 24 GHz literature focuses on multi-class classification or complex
FMCW radar system that can identify, track, and classify a radar representations, with relatively limited attention given
number of human targets in the outdoor setting. Their method- to simplified binary classification tasks and systematic evalua-
ology was an integration of classical radar signal processing tion of generalization performance under practical deployment
and a one-class Support Vector Machine (SVM) to identify [Link], to achieve high-quality range-angle
human beings and clutter and other objects. Although the images, it is frequently necessary to have complicated antenna
one-class classifier was used to solve the issue of invisible array design or mechanically rotating radars, which add cost,
target signatures, the approach heavily depended on manually- complexity to preprocessing, and computational complexity to
created Doppler- and rangebased features and was therefore systems and restrict their usefulness in real-time and resources
restricted in its scalability and ability to adapt to various limited systems. Deep learning can be considered a significant
settings and radar setups. alternative to more traditional feature-based approaches and
Micro-doppler analysis is also widely used as a human clas- allows the automatic acquisition of discriminative features
sification method. Micro-Doppler signatures only detect fine using radarpainted images. CNNs have effectively been used
grain motions, which occur due to human body parts, like arms to range- Doppler maps, spectrograms and other radar image
and legs, and they present very discriminative information representations and in such complex and cluttered scenarios
to distinguish between humans and animals,and vehicles, as they can always be more successful than classical machine
well as to identify human activities and movements. Although learning techniques. Transfer learning also enabled a smaller
effective, micro-Doppler based methods also generally need amount of data to be used to train deep models, through
long observation windows in time and are computationally fine-tuning already trained CNN architectures in radar data
expensive, which is not as compatible as real-time embedded including ResNet, VGG, and MobileNet. Although these im-
systems and low-power applications. provements have taken place, simplified binary classification
In order to address the shortcomings of the rule-based problems and systematic assessment of generalization effec-
detector, numerous researchers have employed the classical tiveness as constrained by real-life deployments have received
machine learning methods with hand-crafted features based on little attention in the existing literature although the term
radar signal representations. Bhatia et al. suggested an FMCW multi-class classification or complex radar representation is
radar-based object classification system that derives statistical widespread.
information of range FFT plots in terms of peak amplitude,
spectral width, variance, and area under the FFT curve. The
III. C HALLENGES IN FMCW R ADAR -BASED H UMAN AND
inputs to these classes included these features, and it was
O BJECT C LASSIFICATION
revealed that range FFT features have discriminative informa-
tion to classify objects, and this was done using the Logistic Although FMCW radar-based sensing has advanced much,
Regression and Naive Bayes classifier. The range- Doppler human and object classification are nevertheless difficult issues
maps and the point-cloud representations have been treated in to be resolved because of a set of signal-level, algorithmic, and
 3

Fig. 1. Time–frequency spectrogram of a bike target captured using FMCW Fig. 2. Time–frequency spectrogram of a human target obtained from FMCW
radar. radar.

system-level constraints. The current literature illuminates Practical deployment is also further restricted by compu-
that there are various underlying issues which permeate tational complexity and real-time constraints. More sophisti-
strength, generalization and practical application of FMCW cated methods of classification, like micro-Doppler analysis or
radar classification systems. multi-dimensional radar imaging, like range -angle representa-
Clutter, multipath propagation, and environmental variabil- tions, consume sub-stantial computational resources. YOLO-
ity can be considered as one of the most intractable problems. based models and three dimensional con- volutional networks,
The surfaces and objects found inside and in urban areas cause among others, are deep learning architectures that add to pro-
high reflections on the walls, furniture, ground surfaces, and cessing latencies and power consumption. Embedded and edge
fixed objects, which may distort target signatures producing devices have a big challenge in meeting these requirements
false detections. Rain, fog, dust and moving vegetation are because of limited memory, computation and energy resources.
environmental factors that add to the difficulty in interpreting One of the main design issues is to be able to achieve an
signals. Despite the fact that clutter effects are reduced by effective trade-off between classification accuracy and real-
methods like Moving Target Indication (MTI) filter- ing, back- time performance [4].
ground subtraction, and Constant False Alarm Rate (CFAR) Limitation of hardware and system configuration is also
detection, they are not able to completely avoid false alarms very significant. Radar representations of high quality might
especially when the non-human objects share motion patterns need complicated hardware systems, such as giant antenna ar-
with human beings [5] . rays, mechanical rotating radars, or numerous-input numerous-
The other urgent problem is the deterioration of perfor- output (MIMO) systems. These systems raise the costs, the
mance of classification in low signal-noise ratio (SNR) sce- complexity of calibration and use of power. Moreover, the
narios. Low radar cross sections (RCS) are common with performance of classification is also dependent on the antenna
small targets, objects that are distantly located, and humans placement and orientation as well as the geometry of the
that are partially covered, which leads to the reflection being environment especially in confined spaces like vehicle cabins
close to the noise floor. This is particularly significant in where tolerances can have great impact on the performance of
usage in detecting drones, detecting children, and detecting the signals in terms of quality and reliability [7].
hidden objects. Such micro-motions as breathing or minute Also, most deep learning designs are optimized on optical
movements of limbs produce Doppler shifts which are hard images and might not utilize radar-related properties to their
to differentiate noise without long observation windows or best advantage. The absence of massive, heterogeneous, and
sophisticated filters, which makes them hard to apply in real- standard radar data contributes to the further overfitting and
time. cross-domain generalization. The hardware behavior of real
Most of the existing FMCW radar classification techniques FMCW radar systems is also non-ideal such as phase noise,
are based on handcrafted features that are developed based on mixer nonlinearity, finite ADC resolution and timing jitter, [8].
range FFT, Doppler spectra or micro-Doppler representations. The algorithms that have been tested and worked with simple
Although these features may perform well in closed systems, or synthetic information are often not precise in practical hard-
expert knowledge is vital and requires an expert parameter ware implementation. This means that noise- aware modeling
tuning. The handcrafted features are not always able to gen- and realistic data representation is necessary, and most of the
eralise in other settings, radar configurations and classes of existing classification methods do not directly consider such
objects, and therefore when the system is brought out to view impairments in training or testing [8] .
unseen variable scenarios, the performance will be degraded. By and large, there is an inherent trade-off between robust-
 4

ness and system simplicity of FMCW radar classification [9]. constructed as radar spectrograms. These spectrograms record
Simple-feature and simple-classifier-based lightweight meth- motion-related features and micro- Doppler patterns which
ods are computationally efficient and do not offer robustness give discriminative features to the various object classes. The
and adaptability, but more complex deep-learningbased and spectrograms are perceived as image-like inputs that can be
imaging-based methods are more accurate at the expense of used with the help of learningbased classification models [13].
higher hardware and complexity of processing. The challenge The resulting spectrograms are in turn inputted into deep
of creating classification systems that are both robust, have the learn- ing-based classification networks. Various architec-
ability to generalize, work efficiently, and can be deployed tures are considered, such as ResNet-18 and YOLOv8n,
is open to research. Algorithms developed and validated on EfficientNet-B0 and MobileNetV3-RD. Such models are
simplified or synthetic data frequently fail when deployed trained on discriminative features of radar-derived images and
on real hardware. As a result, noise- aware modeling and are used to do binary classification to tell the difference
realistic data representation are essential, yet many existing between human and bicycle targets [14] [15].
classification approaches do not explicitly account for these FiLastly, the output of classification is consumed to identify
impairments during training or evaluation. the target category. In the case where the output has been
Besides these problems, there has been insufficient focus determined to be in line with the anticipated target features, the
on the systematic assessment of pretrained deep convolutional target gets successfully classified either as human or a bicycle.
neural networks on radar-based image representations on This last stage of decision guarantees predictable classification
simplified binary classification tasks. Majority of the current without being sensitive to noise, clutter and environmental
literature focuses on multi-class classification or sophisticated fluctuations. Based on the suggested methodology, several
radar representations resulting in a gap in the comprehension learningclassification strategies are used to determine how well
of the generalization accomplishment of transfer learning- FMCW radar images based on images can be classified as
based classifications to viable binary human-object separation human and objects.
[10] .
A. ResNet-18-Based FMCW Radar Classification
IV. M ETHODOLOGY This work investigates the effectiveness of a fine-tuned
ResNet-18 convolutional neural network for binary object clas-
Here, the procedures of human and object classification in
sification using FMCW radar–derived images. The classifica-
the FMCW radar system are introduced in general. A detailed
tion task focuses on distinguishing between human targets and
processing chain, starting with the radar signal capture process
bicycle using a dataset consisting of 7,004 labeled samples
through to the end target classification process is shown in
[16] [17]. All images are resized to 224 × 224 pixels and
Fig. 3. The suggested framework combines traditional radar
normalized using standard ImageNet statistics to match the
signal processing technology with classification models that
pretrained network requirements. The dataset is divided into
are based on learning so as to attain strong and dependable
training (70%), validation (15%), and testing (15%) subsets to
performance [11]. The radar sends out the chirps of fre-
ensure unbiased performance evaluation [18] [19].
quencymodulated linear formats and the echoes that bounce
The ResNet-18 trained model is used with the last fully
off the targets are received and down-converted to the base-
connected layer adjusted to give two classes as outputs. The
band. They are signals which carry messages of target range,
Adam optimizer is used with a learning rate of 0.001 and the
velocity and motion characteristics and are sampled to get
cross-entropy loss to train the model in ten epochs. It has been
raw radar data. Standard signal processing is used to process
shown through experimental results that there is a stable con-
the radar data that is acquired. The transformation of range
vergence and high generalization of the experimental results
data is performed by a Fast Fourier Transform (FFT), whereas
across all dataset splits. Accuracies of 94.78% on the training
the transformation of velocity-related features is done by the
set, 92.29% on the validation set, and 93.82% on the test set
Doppler processing of multiple chirps. Reduction and suppres-
are achieved. The analysis of the confusion matrix shows that
sion of noise and clutter methods are adopted to improve the
there are no errors in the classification of all human samples,
visibility of the target and minimize false detection leading to
and there is only a slight misclassification of bicycle samples
informative radar representations to be further analyzed [12].
[20] [21]. These results prove the relevance of transfer learning
Following signal processing, a detection algorithm deter-
with ResNet-18 to effectively and successfully classify human
mines the presence of a valid target. If no target is detected,
beings and objects with the use of radar under the FMCW
the system reports a No Object condition and terminates
radar.
further processing. When a target is detected, the pipeline
proceeds to feature representation and classification to avoid
unnecessary computation. A detection algorithm is used after B. YOLO-Based FMCW Radar Object Classification
signal processing to ascertain the existence of a valid target. In The article assesses a deep learning method, based on
case of nondetection of the target the system sends a No Object the YOLO, to classify radar spectrogram images of radar
condition and stops the rest of processing. Upon the detection objects. The research is inspired by the previous studies
of a target, the pipeline moves on to feature representation and that use mmWave-YOLO-based systems to perform real-time
classification in order to eliminate unnecessary computation. In object detection in driving and aerial sensing systems. The
the case of targets detected, time-frequency representations are range angle time (3D), spectrogram data are not available;
 5

Transmit and receive Process raw data using Apply detection

signal using radar signal processing techniques algorithm

Is target No
No object
detected?

Yes

Spectrogram
generation

Classification
algorithm

Human Target is Bike

Human detected Bike detected
human or bike?

Fig. 3. Flow graph of the proposed FMCW radar–based human and object classification methodology.

hence, a lightweight YOLOv8n model is used to work with literature, a classification methodology based on a radar image
twodimensional radar spectrogram images. The classification is offered with the usage of the prepared EfficientNet-B0
task is aimed at differentiating human and bicycle targets. The convolutional neural network. The data is based on 2 classes,
dataset is divided into training (70%), validation (15%), and human and bicycle, and has a training set training (70%), val-
testing (15%) subsets, with all images resized to 224 × 224 idation (15%), and testing (15%) to ascertain reproducibility.
pixels. The model finishes the training process with 60 epochs EfficientNet-B0 is chosen because of its good tradeoff between
and Adam optimizer with a learning rate of 0.001. The accuracy and computational efficiency and it can be used in
convergence is fast as experiment results indicated an im- the embedded case. The convolutional layers are frozen, and
provement in training accuracy up to more than 99 percent only classification head is trained to minimize the cost of
as compared to about 70 percent before the experiment. Last computation and overfitting [24] [25].
accuracies some are 99.79% (training), 98.41% (validation), The Adam optimizer, sparsely categorical cross-entropy
and 98% (testing) are [Link] the findings indicate loss, and validation loss to early stopping is used to train the
good performance on the provided data, the study addresses model to prevent overfitting. Accuracy, loss curves, confusion
that the good performance is due to a lack of data variety matrices, and class-wise performance are some of the measures
as well as repeated patterns which means that there might be used to perform performance evaluation. The findings show
generalization constraints in the actual implementation of the that the suggested EfficientNet-based algorithm is superior
research [22] [23]. to classical handcrafted feature-based ones and has fewer
complexities than more heavy deep learning architectures [26].
C. EfficientNetB0-Based FMCW Radar Classification Frame- The paper concludes that lightweight CNNs will be an efficient
work and realistic approach to the problem of FMCW radar-based
human and object classification by transfer learning.
This paper will provide the comparative analysis of object
classification procedures based on FMCW radars and suggest a
light deep learning architecture based on transfer [Link] D. MobileNetV3-Based FMCW Radar Object Classification
paper examines three exemplary methods that include hand- This paper examines the concept of the FMCW radar-based
designed feature-based classification that applies range FFT object detection and classification that is applied in security
features, machine learning based vehicular in-cabin sensing al- and human-object discrimination. The paper identifies the
lows detection of child presence, and noise-conscious FMCW shortcoming of the traditional imaging systems and encourages
radar simulation platform. These works all point to the trade- the application of millimeter-wave FMCW radar since it has
offs in feature engineering, system-level robustness and data the capacity to penetrate clothes and nonmetallic objects
[Link] the conclusions offered by the analyzed without invading privacy. The radar-derived images like range-
 6

Doppler maps, micro-Doppler spectrograms and range angle correctly and has less misclassifications only with bicycle
maps are regarded as grayscale images that can be analized targets. ResNet-18 has the advantage of residual connections
using deep learning [27]. that enhance the stability and noise resilience of training, and
The identity of humans and bicycles is classified using a it is therefore effective in radar-acquired image classification.
binary architecture of deep learning on nearly 7,000 radar ResNet-18 benefits from residual connections, which improve
images, which is referred to as MobileNetV3-RD. The model
is particularly designed to take radar-domain inputs and make
use of depthwise separable convolutions and squeeze-and-
excitation blocks to have low latency and power consumption.
The training is carried out after 22 epochs, which gives a
classification accuracy of about 96.29%.
Confusion matrix analysis as a method of performance
evaluation shows great level of diagonal dominance, which
signifies credible separation of classes. There are also low mis-
classification rates of both classes, just due to partial visibility
of the target and structural resemblance among them. The
findings indicate that the MobileNetV3 architectures offer a
good tradeoff between accuracy, robustness, and computational
efficiency, and therefore they can be used in real-time and
embedded FMCW radar classification systems [28].
Fig. 5. Comparison of training and validation accuracy across epochs

V. R ESULTS
training stability and robustness to noise, making it well suited
Despite the fact that each of the four learning-based methods for radar-derived image classification.
has done very well in terms of performance in FMCW The EfficientNet-B0 model has a higher performance than
radar-based human and object classification, the degree of classical handcrafted featurebased meth- ods and at the same
their effectiveness can be affected upon generalization ability, time low computational complexity. Its advantage is efficiency
complexity of computation, and the possibility of practical and applicability to embedded sys- tems but the absence of
implementation. The YOLOv8n approach is the one that has explicitly reported final accuracy constraints direct quantitative
the highest numerical accuracy, with its best performance of comparison with other deep learning models.
98% on the test data. Nonetheless, such a performance is
done on fairly small and monotonous data, and the assignment
specifically points out that this high accuracy might not apply
to practice. YOLO- based models are also computationally
more expensive and are not optimized to perform pure clas-
sification, which can create a constraint on suitability to low-
power embedded radar systems [29] [30].

Fig. 6. Accuracy vs Loss plot of EfficientNet-B0 model.

MobileNetV3-RD model has high accuracy of about 96.29%

and has the lowest latency and power consumption as com-
pared to the other models that have been tested. It is especially
suitable to be deployed on the edge due to its radar-domain
adaptation as well as lightweight architecture. Nonetheless,
there is slight misclassification when there is partial visibility
and low quality of radar images.
Fig. 4. Training accuracy progression over epochs, showing rapid convergence
In sum, the ResNet-18-based model has the lowest trade-
in early epochs off between accuracy, robustness, and generalization, which
is why it is the most appropriate model to use in ensuring
The model based on ResNet-18 is highly balanced and reliable FMCW radar-based human and object classification.
reliable in its performance compared to the others that have YOLOv8n is more accurate in terms of the number, but the
been studied. The test accuracy (93.82%) of it is lower than possibility of overfitting and the complexity also decreases
the one of YOLOv8n, but it exhibits a better generalization be- its usefulness. The MobileNetV3-RD and EfficientNet-B0 are
havior, as shown by stable validation and consistent confusion better used in a resource-constrained environment, where
matrixes. The model is able to classify all human samples efficiency holds greater importance than accuracy.
 7

TABLE I
C OMPARISON OF L EARNING -BASED FMCW R ADAR C LASSIFICATION A PPROACHES

Model Accuracy Dataset Strengths Limitations

ResNet-18 Train: 94.78% 7004 FMCW radar images Strong generalization, stable Minor misclassification of bikes as
Val: 92.29% (Humans vs Bikes) convergence, effective transfer humans; moderate computational
Test: 93.82% learning without handcrafted cost compared to lightweight mod-
features els
YOLOv8n Train: 99.79% FMCW radar spectrogram Very high accuracy, fast con- Performance influenced by limited
Val: 98.41% images vergence, suitable for real-time dataset diversity; potential overfit-
Test: 98% (Humans vs Bikes) detection ting and reduced real-world gener-
alization
EfficientNet-B0 Higher than classical ML FMCW radar image Lightweight architecture, low Exact accuracy not explicitly re-
(Better than ∼87%) dataset computational cost, good bal- ported; relies on transfer learning
(Humans vs Bikes) ance between accuracy and ef- assumptions
ficiency
MobileNetV3-RD Overall Accuracy: 96.29% ∼7000 Range–Doppler Low latency, low power Slight confusion under partial vis-
images consumption, edge- ibility; performance dependent on
(Humans vs Bikes) deployment friendly, radar- radar image quality
domain adaptation

level of accuracy. The best overall compromise in terms of

accuracy, robustness and generalization was the fine-tuned
ResNet-18 model out of all the approaches evaluated. Its
consistent convergence properties, good performance in valida-
tion, as well as its usefulness in transfer learning testify to the
appropriateness of pretrained residual networks as an approach
to FMCW radar-based classification without manual feature
engineering. All in all, the findings confirm that radar-based
image representations together with deep learning are a viable
and scalable approach to human and object classification in
adverse settings. This framework could be further extended to
multi-class classification in future work, the realistic training
that is noise-aware and the deployment of the framework into
a real-time embedded radar system could be done to show the
safety-critical application.

R EFERENCES
Fig. 7. Training and validation accuracy curves for the MobileNet model
[1] S. Campbell et al., “Sensor technology in autonomous vehicles: A
review,” in Proc. 29th Irish Signals Syst. Conf. (ISSC), Jun. 2018, pp.
1–4.
VI. C ONCLUSION [2] Y. Cheng, J. Su, M. Jiang, and Y. Liu, “A novel radar point cloud
generation method for robot environment perception,” IEEE Trans.
This paper has provided a survey and experimental in- Robot., vol. 38, no. 6, pp. 3754–3773, Dec. 2022.
vestigation of human and object classification on FMCW [3] M. Kotur, N. Lukić, M. Krunić, and Ž. Lukač, “Camera and lidar sensor
radar systems based on learning processes. The paper revis- fusion for 3d object tracking in a collision avoidance system,” in Proc.
Zooming Innov. Consum. Technol. Conf. (ZINC), May 2021, pp. 198–
ited classical signal-processing solutions, manually designed 202.
feature-based machine learning solutions and the recent deep [4] R. Gourova, O. Krasnov, and A. Yarovoy, “Analysis of rain clutter
learning solutions to radar image classification. Such issues detections in commercial 77 ghz automotive radar,” in Proc. Eur. Radar
Conf. (EURAD), Oct. 2017, pp. 25–28.
as environmental clutter, low signal-to-noise ratio goals, weak [5] J. W. Marck, A. Mohamoud, E. vd Houwen, and R. van Heijster, “Indoor
generalization, hardware limitations, and the trade-off between radar slam: a radar application for vision and gps denied environments,”
the robustness and complexity of systems were identified as in Proc. Eur. Radar Conf., 2013, pp. 471–474.
[6] M. Bauduin and A. Bourdoux, “Impact of phase noise on fmcw and
key. To manage these problems, several learning classification pmcw radars,” in Proc. IEEE Radar Conf. (RadarConf), May 2023, pp.
frameworks have been discussed with the help of practical 1–6.
assign- ments like ResNet-18, YOLOv8n, EfficientNet-B0, and [7] D. Steinhauser, P. Held, B. Thöresz, and T. Brandmeier, “Towards safe
autonomous driving: Challenges of pedestrian detection in rain with
MobileNetV3-RD architectures. All the models were tested on automotive radar,” in Proc. 17th Eur. Radar Conf. (EuRAD), Jan. 2021,
binary human-bicycle classification on FMCW radar-derived pp. 409–412.
datasets. The comparative analysis revealed that although the [8] T. Brühl, L. Ewecker, R. Schwager, T. Sohn, and S. Hohmann, “Making
radar detections safe for autonomous driving: A review,” in Proc. 10th
YOLOv8n had the best numerical accuracy, it had a limited Int. Conf. Vehicle Technol. Intell. Transp. Syst., A. V. Vinel, K. Berns,
capacity to perform because of its dataset limitations and J. Ploeg, and O. Gusikhin, Eds., May 2024, pp. 299–310.
increased complexity. Efficient architecture lightweight models [9] M. Gerstmair, A. Melzer, A. Onic, R. Stuhlberger, and M. Huemer,
“Highly efficient environment for fmcw radar phase noise simulations
like EfficientNet-B0 and MobileNetV3-RD provided efficient in if domain,” IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 65, no. 5,
and fitted well into an embedded system, with an acceptable pp. 582–586, May 2018.
 8

[10] K. Siddiq, M. K. Hobden, S. R. Pennock, and R. J. Watson, “Phase noise

in fmcw radar systems,” IEEE Trans. Aerosp. Electron. Syst., vol. 55,
no. 1, pp. 70–81, Feb. 2019.
[11] P. K. Rai, H. Idsøe, R. R. Yakkati, A. Kumar, M. Z. A. Khan,
P. K. Yalavarthy, and L. R. Cenkeramaddi, “Localization and activity
classification of unmanned aerial vehicle using mmwave fmcw radars,”
IEEE Sensors Journal, vol. 21, no. 14, pp. 16 043–16 053, 2021.
[12] A. Kosuge, S. Suehiro, M. Hamada, and T. Kuroda, “mmwave-yolo:
A mmwave imaging radar-based real-time multiclass object recognition
system for adas applications,” IEEE Transactions on Instrumentation
and Measurement, vol. 71, pp. 1–10, 2022.
[13] N. Kumchaiseemak, I. Chatnuntawech, S. Teerapittayanon, P. Kotcha-
pansompote, T. Kaewlee, M. Piriyajitakonkij, T. Wilaiprasitporn, and
S. Suwajanakorn, “Toward ant-sized moving object localization using
deep learning in fmcw radar: A pilot study,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 60, pp. 1–10, 2022.
[14] C. J. Li and H. Ling, “High-resolution, downward-looking radar imaging
using a small consumer drone,” in 2016 IEEE International Symposium
on Antennas and Propagation (APSURSI). IEEE, 2016, pp. 2037–2038.
[15] X. Zhuge and A. G. Yarovoy, “A sparse aperture mimo-sar-based uwb
imaging system for concealed weapon detection,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 49, no. 1, pp. 509–518, 2011.
[16] E. Turppa, J. M. Kortelainen, O. Antropov, and T. Kiuru, “Vital sign
monitoring using fmcw radar in various sleeping scenarios,” Sensors,
vol. 20, no. 22, p. 6505, 2020.
[17] G. Y. Tian, A. Al-Qubaa, and J. Wilson, “Design of an electromagnetic
imaging system for weapon detection based on gmr sensor arrays,”
Sensors and Actuators A: Physical, vol. 174, pp. 75–84, 2012.
[18] M. Rezaei, H. Zamani, M. Fakharzadeh, and M. Memarian, “Quality
improvement of millimeter-wave imaging systems using optimized dual
polarized arrays,” IEEE Transactions on Antennas and Propagation,
vol. 69, no. 10, pp. 6848–6856, 2021.
[19] Y. Jang, I. Jeon, T.-K. Kim, and W. Woo, “Metaphoric hand gestures for
orientation-aware vr object manipulation with an egocentric viewpoint,”
IEEE Transactions on Human-Machine Systems, vol. 47, no. 1, pp. 113–
127, 2016.
[20] K. A. Smith, C. Csech, D. Murdoch, and G. Shaker, “Gesture recognition
using mm-wave sensor for human-car interface,” IEEE sensors letters,
vol. 2, no. 2, pp. 1–4, 2018.
[21] S.-J. Ryu, J.-S. Suh, S.-H. Baek, S. Hong, and J.-H. Kim, “Feature-based
hand gesture recognition using an fmcw radar and its temporal feature
analysis,” IEEE sensors Journal, vol. 18, no. 18, pp. 7593–7602, 2018.
[22] C. Liu, Y. Li, D. Ao, and H. Tian, “Spectrum-based hand gesture recog-
nition using millimeter-wave radar parameter measurements,” IEEE
Access, vol. 7, pp. 79 147–79 158, 2019.
[23] S. Ahmed and S. H. Cho, “Hand gesture recognition using an ir-uwb
radar with an inception module-based classifier,” Sensors, vol. 20, no. 2,
p. 564, 2020.
[24] J.-W. Choi, S.-J. Ryu, and J.-H. Kim, “Short-range radar based real-time
hand gesture recognition using lstm encoder,” IEEE Access, vol. 7, pp.
33 610–33 618, 2019.
[25] J.-W. Choi, C.-W. Park, and J.-H. Kim, “Fmcw radar-based real-time
hand gesture recognition system capable of out-of-distribution detec-
tion,” IEEE Access, vol. 10, pp. 87 425–87 434, 2022.
[26] J. Liu, K. Zhang, Z. Sun, Q. Wu, W. He, and H. Wang, “Concealed
object detection and recognition system based on millimeter wave fmcw
radar,” Applied Sciences, vol. 11, no. 19, p. 8926, 2021.
[27] G. L. Charvat, Small and short-range radar systems. CRC Press, 2014.
[28] O. Schumann, J. Lombacher, M. Hahn, C. Wöhler, and J. Dickmann,
“Scene understanding with automotive radar,” IEEE Transactions on
Intelligent Vehicles, vol. 5, no. 2, pp. 188–203, 2019.
[29] R. Prophet, M. Hoffmann, M. Vossiek, C. Sturm, A. Ossowska, W. Ma-
lik, and U. Lübbert, “Pedestrian classification with a 79 ghz automotive
radar sensor,” in 2018 19th International Radar Symposium (IRS).
IEEE, 2018, pp. 1–6.
[30] S.-L. Jeng, W.-H. Chieng, and H.-P. Lu, “Estimating speed using a side-
looking single-radar vehicle detector,” IEEE transactions on intelligent
transportation systems, vol. 15, no. 2, pp. 607–614, 2013.