IEEE Circuits and Systems Magazine - Q1 2022 - 68

Wider bandwidth is envisioned to be leveraged for integrated RF transceivers to realize
enhanced sensing resolution and improved communication efficiency.
a 3.8% error rate and faster average acquisition time in
less than 1 second compared to other existing methods.
AI algorithms have been used in radar sensors to
directly detect vital signs [156]-[158], where conventional
machine learning and deep learning (DL) are used
to extract the vital signs information from radar waves
reflected by the targeted human.
Conventional machine learning usually operates
based on short-time Fourier transform (STFT) or wavelet
transform (WT) to extract manually designed features.
A well-trained classifier is used to realize subject person
localization and respiration pattern classification [156].
After extracting relative features from UWB radar signals
based on STFT processing, the support vector
machine (SVM) is served as a classifier to remotely
identify the subject's location via a set of labeled patterns.
The main limitation of a conventional machine
learning method is that it is difficult to achieve superior
accuracy on complicated classification tasks. Also,
the manually designed features are limited by the designers'
knowledge and experience, where some vital
information is missing. To address the issue, deep convolutional
neural networks (CNNs) are used to ensure
superior performance on image classification and vital
signs detection. Deep learning methods have been adopted
in biomedical radar systems to achieve accurate
reconstruction on vital sign signals, by which accurate
respiration pattern recognition, heartbeat rate estimation
for health status monitoring, and personal identification
can be achieved robustly [157], [158]. In [157],
a 1D CNN model is leveraged for UWB radar signal processing
to accurately recognize four typical respiration
patterns: Eupnea, Bradypnea, Tachypnea and Apnea.
The proposed deep CNN model can achieve 95.8% average
accuracy, which is better than the performance of
the conventional ML method using STFT extractor and
SVM classifier operating at 87.6% accuracy. S. Wu et.al.
proposed a DL-based signal reconstruction scheme to
enhance the performance of a person-specific longterm
UWB radar monitoring system [158]. By utilizing
the time deviation between the radar sensor and the
ECG devices, the heartbeat signals can be reconstructed.
The DL-based scheme achieved better accuracy on
heart rate estimation compared to the conventional
time-frequency processing method.
Based on CNN techniques, hand gestures and human
gaits can be recognized accurately when operating
in coordination with chip-scale radar imaging,
68
IEEE CIRCUITS AND SYSTEMS MAGAZINE
which demonstrates the capability to realize seamless,
robust, and contactless human-machine interactions
[103], [152]. In [101], ensemble learning is used to enhance
the classification capabilities of IR UWB radar
systems for the monitoring and localization of people
inside vehicles. Moreover, machine learning algorithms
can be leveraged to enhance the accuracy of localization
under none-line-of-sight (NLOS) interferences. In
[102], by using relevance vector machine (RVM) and
two-step iterative (TSI) algorithms, the UWB localization
accuracy is improved, and the coverage area is
expanded. In [103], a deep learning method leveraging
the combination of CNN and long-short term memory
(LSTM) is employed to enable accurate UWB NLOS/LOS
classification. In [104], an FPGA-based object detection
processor is operating with a cell-based histogram of
oriented gradient (HOG) feature descriptor and a support
vector machine (SVM) classifier, demonstrating
the potential to be adapted for embedded radar signal
feature classifications. A time-domain-based artificial
intelligence (AI) radar system-on-chip is presented in
[152], where 1D-CNN and LSTM are leveraged synergistically
to enable accurate recognitions on dynamic and
static gestures. Moreover, AI-assisted RF transceivers
can realize human emotion recognition, demonstrating
the potential for comprehensive health status monitoring
in the future [130].
Vi. Future Perspectives
In the future, wider bandwidth is envisioned to be leveraged
for integrated RF transceivers to realize enhanced
sensing resolution and improved communication efficiency.
Improved RF front-end designs, ASIC, and specific
intelligent signal processing algorithms for wideband TRX
sensing and communications are needed [125], [126]. Specific
TRX signal processing techniques supported by AI
would be developed in coordination with integrated wideband
RF transceivers to enable ubiquitous sensing and
communications with low latency, high accuracy, compact
size, and low power consumption [102], [103], [151].
Compared to the mm-Wave transceiver chips, subTHz
and THz transceivers have broader bandwidth,
demonstrating promising capabilities to enable more
accurate sensing and communications with significantly
higher data rate [114], [135], [136], [146]. A typical subTHz
phased-array transmitter is demonstrated in [135],
where the transmitter works at 318 GHz to 370 GHz range
with 344 GHz carrier frequency. The transmitter has
FIRST QUARTER 2022
IEEE Circuits and Systems Magazine - Q1 2022

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q1 2022
