Multibin Breathing Pattern Estimation by Radar Fusion for Enhanced Driver Monitoring

Monitoring the status of the driver is a crucial aspect of health monitoring inside vehicles as it helps to identify potential health or safety risks that could affect a driver’s ability to operate a vehicle safely. This includes monitoring for fatigue, distraction, or impairment, among other things, which can potentially cause car crashes. Although many solutions for health monitoring in private vehicles have been proposed, most of them are inconvenient to use or have the risk of leaking private information. Radars have the potential to address the above drawbacks by their inherent privacy protection and contactless operation in addition to their high accuracy, convenience, affordable price, and resilience to environmental factors. Among many possible radar configurations, millimeter Frequency Modulated Continuous Wave (FMCW) radars can accurately detect range and monitor displacements that are essential in breathing pattern monitoring. Breathing pattern monitoring is one of the key signatures of the driver’s health. An accurate estimation of the breathing pattern enables the detection of breathing abnormalities, including Tachypnea, Bradypnea, Biot, Cheyne–Stokes, and Apnea. The breathing pattern can be estimated from both the chest and abdomen. For this purpose, we employed two 60-GHz FMCW radars. The proposed algorithm is capable of detecting the mentioned breathing abnormalities through breathing rate (BR) estimation and breath-hold period detection. In addition, the proposed method in this article estimates BR based on the multiple range bins. We conducted a study on the human radar geometry problem inside a vehicle to determine the accurate number of range bins for BR estimation. The experimental results demonstrate a maximum BR error of 1.9 breaths/min using the proposed multibin technique. In addition, the dual radar fusion system can detect breath-hold periods with minimal false detections.


Multibin Breathing Pattern Estimation by Radar
Fusion for Enhanced Driver Monitoring Ali Gharamohammadi , Graduate Student Member, IEEE, Mohammad Pirani , Senior Member, IEEE, Amir Khajepour , Senior Member, IEEE, and George Shaker , Senior Member, IEEE Abstract-Monitoring the status of the driver is a crucial aspect of health monitoring inside vehicles as it helps to identify potential health or safety risks that could affect a driver's ability to operate a vehicle safely.This includes monitoring for fatigue, distraction, or impairment, among other things, which can potentially cause car crashes.Although many solutions for health monitoring in private vehicles have been proposed, most of them are inconvenient to use or have the risk of leaking private information.Radars have the potential to address the above drawbacks by their inherent privacy protection and contactless operation in addition to their high accuracy, convenience, affordable price, and resilience to environmental factors.Among many possible radar configurations, millimeter Frequency Modulated Continuous Wave (FMCW) radars can accurately detect range and monitor displacements that are essential in breathing pattern monitoring.Breathing pattern monitoring is one of the key signatures of the driver's health.An accurate estimation of the breathing pattern enables the detection of breathing abnormalities, including Tachypnea, Bradypnea, Biot, Cheyne-Stokes, and Apnea.The breathing pattern can be estimated from both the chest and abdomen.For this purpose, we employed two 60-GHz FMCW radars.The proposed algorithm is capable of detecting the mentioned breathing abnormalities through breathing rate (BR) estimation and breath-hold period detection.In addition, the proposed method in this article estimates BR based on the multiple range bins.We conducted a study on the human radar geometry problem inside a vehicle to determine the accurate number of range bins for BR estimation.The experimental results demonstrate a maximum BR error of 1.9 breaths/min using the proposed multibin technique.In addition, the dual radar fusion system can detect breath-hold periods with minimal false detections.
Index Terms-Advanced driver assistance systems (ADAS), driver status monitoring, millimeter-wave radar, radar fusion.

I. INTRODUCTION
I N RECENT years, radar technology has found applications in a diverse range of human monitoring applications, such as vital sign monitoring [1], occupancy detection [2], and gesture recognition for human-computer interfaces [3].The escalating interest in human monitoring is driven by the Ali Gharamohammadi and Amir Khajepour are with the Mechanical and Mechatronics Engineering Department, University of Waterloo, Waterloo, ON N2L 3G1, Canada (e-mail: aligharamohammadi@gmail.com).
Mohammad Pirani is with the Mechanical Engineering Department, University of Ottawa, Ottawa, Ontario ON K1N 6N5, Canada.
George Shaker is with the Electrical and Computer Engineering Department, University of Waterloo, Waterloo ON N2L 3G1, Canada.
Digital Object Identifier 10.1109/TIM.2023.3345909growing amount of time commuters spend in transit each day.Radar sensors offer distinct advantages in this context, particularly their ability to detect and monitor individuals through nonmetallic objects.Additionally, these sensors excel in preserving privacy, making them a preferred choice for in-cabin applications [2].
Monitoring a driver's breathing can be an important factor in determining their health and well-being during driving.Breathing patterns can provide important information about a driver's level of relaxation, stress, or fatigue, all of which can impact their ability to operate a vehicle safely.According to statistics from the years 2021 and 2022, 93.8% of Americans (who are 16 years of age or older) spend 61.3 min or more per day in a car [4].Since driving accounts for a considerable amount of our daily lives, it is essential to track drivers' status, particularly breathing patterns, consistently to identify potential safety problems as early as possible.Driver status monitoring is the most practical factor that can enable advanced driver assistance systems (ADAS) to control the vehicle in emergencies, e.g., when the driver encounters Apnea, fatigue, stroke, trauma, or stress [5].These abnormal conditions can be detected by analyzing breathing patterns.As a result, drivers' breathing monitoring is a crucial task that can be accomplished by contact and noncontact sensors.
Several contact and noncontact technologies are deployed for the driver's breathing monitoring.Contact technologies require a constant attachment to the human body all the time, which makes them less convenient for daily commuting.In addition, wearing the sensor needs an extra driver's attention and can cause distraction.Conversely, noncontact sensors are less intrusive and can monitor drivers breathing without causing distraction [6].
Vision-based sensors use image sequences to monitor BRs.These sensors have gained attention because of their affordability and ease of use, even without requiring prior knowledge [7].These sensors mostly detect motions to monitor breathing patterns.Another approach is based on the changes in the intensity of reflected light [5].Nonetheless, the vulnerability to the illumination level and privacy concerns are the main drawbacks of vision-based sensors.
Radar offers numerous benefits compared to noncontact and contact sensors.First, the radar functionality remains unaffected at different illumination levels.Second, radar preserves privacy since it does not show the entire or the partial human body [7].These benefits can be leveraged in millimeter-wave radars since they are less susceptible to noise due to the higher carrier frequency, and electromagnetic wave penetrates the human body to a less extent at these frequencies due to the higher reflection coefficient of skin [8].In addition, millimeter-wave radar has small packaging, low power consumption, and low price.Consequently, millimeterwave radar is a promising technology for estimating breathing patterns within a vehicle.
Estimating breathing patterns by millimeter-wave FMCW radars encounters some challenges for in-cabin applications.First, breathing patterns can be estimated from the chest and abdomen displacements.However, due to the low amplitude of chest wall displacements during abdominal breathing, both the chest and abdomen should be monitored by a dual radar to prevent false detections.Nevertheless, recent in-cabin investigations employed a single radar for estimating breathing patterns [18].
Second, due to the confined space inside a vehicle, the passengers' reflections can interfere with the driver's reflections.These interferer reflections pose a challenge in the detection of abnormal breathing patterns that involve in a breath-hold period [19].Since the driver is nonvibrant during a breath-hold period, an interfering reflection from a vibrant passenger can be misinterpreted as a driver's reflections.
Third, normal driving activities like turning the wheel around and shoulder checking can distort the breathing pattern even when they are not periodic [20].Some of the studies employed an expensive and intricate system to mitigate the effect of these movements [21], [22].These studies used a less practical package size for this purpose, which adds discomfort for the drivers.Off-the-shelf millimeter-wave radars offer a compact size radar that is capable of sweeping a high-frequency bandwidth.Consequently, this high-frequency bandwidth leads to improve range resolution [23], [24], [25], [26].As the human body captures more than one range bin in these radars inside the vehicle, multiple range bins can be used for breathing pattern estimation.The proposed multibin approach estimates breathing patterns from different human body parts, which can be used for body motion detection.Another multibin approach is also explored in [5], where the multiple bins are selected from rang-azimuth map.
This study has the following contributions to tackle these challenges.
1) We propose a novel dual radar fusion approach for accurate monitoring of drivers' breathing patterns.The proposed method utilizes two radars that are separated both in the time domain and in the covered spots to avoid interference and improve accuracy.Time-domain separation is based on the signal parameters, specifically tailored for the driver status monitoring application.Additionally, the radars are placed in distinct positions to separate covered spots.The use of dual radar fusion with these separation techniques can provide more reliable and precise monitoring of breathing patterns, which is crucial for ensuring the health and safety of drivers.
To the best of our knowledge, our proposed approach represents the first instance in which dual radar fusion, incorporating both time and space separations, has been employed for driver status monitoring inside a vehicle.2) Our proposed algorithm can detect a range of breathing abnormalities, including Tachypnea, Bradypnea, Biot, Cheyne-stokes, and Apnea.Although most in-cabin articles focused on breathing rate (BR) estimation [5], [27], [28], [29], [30], [31], our focus is on the detection of these abnormalities while driving.3) We propose a multibin breathing pattern estimation method to remove incorrect estimations caused by random body movements.These incorrect estimations due to body motion can be eliminated from the BR estimation.We selected the multiple bins from the range profile since our employed radar has better range resolution than azimuth resolution, and the azimuth bins can be corrupted more by body movement.4) We present a novel signal processing chain that addresses several key aspects.First, our proposed signal processing chain provides a comprehensive framework that encompasses all the necessary steps for estimating BR and detecting breath-hold periods using a dual radar system.Second, we introduce an innovative approach for accurate range detection, which relies on the utilization of multiple detected ranges instead of a single detection.Additionally, we highlight the impact of clutter cancellation on the breathing waveform, particularly during breath-hold periods.The remaining part of this article is organized as follows.Signal and system design for breathing pattern estimation are introduced in Section II, followed by experimental studies in Section III.Section IV evaluates the proposed signal and system performance and discusses future works.Section V concludes this article.

II. SIGNAL AND SYSTEM DESIGN TO ESTIMATE
THE BREATHING PATTERNS The analog and digital signal processing chains for breathing pattern estimation by dual radar are presented step-by-step in this chapter.

A. Analog Signal Processing Chain
Fig. 1 shows a simplified homodyne block diagram of the FMCW radar prior to the analog-to-digital converter (ADC) [32].The receiver amplifies received signals from objects in the environment.These reflected signals are influenced by the properties of the objects, such as range and radar cross section.Then, a mixer correlates the transmitted and received signals.This mixer produces both low-frequency and highfrequency components.A low-pass filter is applied to select the low-frequency components in the next step.Finally, an ADC converts analog signals to discrete signals.
This study uses the BGT60TR13C FMCW radar from Infineon Technologies.This sensor can generate an FMCW signal from 58 to 63.5 GHz modulated by the sawtooth wave.The employed ADC has a 12-bit resolution, providing a 74-dB dynamic range, and its sampling rate can reach up to 4 M samples/s [33].

B. Geometry of Multibin BR Estimation
The human body in front of a high-range resolution radar captures more than one range bin, especially when it is close to the radar [5].Therefore, several range bins around the main range bin of the target can be employed to improve breathing pattern estimation accuracy.If the radar is in front of the subject's chest, the main range bin is the range of the middle of the chest from the radar, which is the closest part of the body to the radar.Other organs of the human body capture additional range bins.If the estimated BR from one of the range bins is incorrect due to the random body movement and noise, it can be removed from estimations.
Fig. 2 represents a radar in front of the subject's chest.The farthest part of the human body which is influenced by lung movements in the chest radar is the shoulder.Assuming that the radar is positioned at r 1 from the middle of the chest and that the distance from the middle of the chest to the shoulder is b 1 , we can use the Pythagorean theorem to calculate the distance between the radar and the shoulder as follows: The human body captures the radar range profile from r 1 to R 1 .Consequently, the number of range bins that can be employed for breathing pattern estimation is as follows:

C. Dual Radar Placement
Abdominal breathing causes low chest wall displacement, which can lead to false estimations of the breath-hold period.Therefore, abdomen and chest displacements should be monitored separately by two distinct radars to eliminate these false detections.Fig. 3 displays dual radar placement in front of chest and abdomen to separate them based on the covered area.The a 1 and a 2 spots should not overlap to accomplish this separation.As a result, the inequality between a 1 , a 2 , and torso length is given by of the antenna.The HPBW-H of BGT60TR13C is 35 • [23].The chest radar distance from the middle of the chest (r 1 in Fig. 3) is 20 cm, and the abdominal radar distance from the abdomen (r 2 in Fig. 3) is 40 cm. a 1 and a 2 are 6.4 and 12.8 cm based on (4), respectively.The summation of a 1 and a 2 is less than the torso length.Therefore, this design satisfies (3), and the chest radar and abdominal radar are separated based on the covered area.Nevertheless, since both radars are at the same frequency and close together, cross-interference appears in their receivers if they operate simultaneously.

D. Dual Radar Signal Design
There are two approaches to fuse two FMCW radars.The first approach involves selecting desperate carrier frequencies without frequency coverage.Because the employed radars are identical, the operational frequency bandwidth can be divided between these two radars.However, the drawback of this division is the loss of range resolution by at least 50%.The second approach entails applying time division multiplexing, where one radar is switched off while the other is switched on and vice versa.Yet, this approach results in a time delay between the radars, which should be minimized considering that the goal is to make decisions using both radars.Therefore, by calculating and incorporating the minimum time delay in the system, both radars can benefit from high-range resolution signals without interfering with each other.
Both pulse repetition frequency (PRF) and pulse duration should be considered to calculate the minimum delay between radars.The maximum frequency of breathing is almost 0.6 Hz [35].Hence, the sampling frequency should be more than 1.2 Hz based on the Nyquist theorem.Because only one chirp signal, which is the average of all received chirps, is employed from each pulse, the sampling of the breathing vibration is equal to PRF.As a result, the PRF should be more than 1.2 Hz.Higher PRF provides more samples from breathing vibration, which enhance breathing pattern estimation and can be beneficial in phase unwrapping [32].However, higher PRF reduces interpulse repetition time, and the hardware should be able to generate that PRF.Because two radars are employed in this study, interpulse repetition should be more than the pulse duration.The PRF is set to 20 in this article.
The pulse duration can be determined according to several parameters of the target.The first parameter is the range resolution as follows: where c is the light velocity in free space and B is the frequency bandwidth [26].The maximum frequency bandwidth should be employed to have accurate results.The maximum frequency bandwidth of BGT60TR13C is almost 5 GHz [23], which results in a 3-cm range resolution.Another parameter to determine chirp time is the number of samples per chirp, which depends on the maximum range and range resolution as follows: where R max is the maximum range of the target (96 cm in this article).The linear chirp slope is defined as follows: where T is the active chirp time.In this article, the chirp slope is set to 78.128 MHz/µs, with 64 samples/chirp.Hence, the ADC sample rate is 1 MHz, and the active chirp time is 64 µs.
Because breathing is a slow activity, the reflected chirps in a pulse are very similar to each other.We consider eight chirps for each pulse.The minimum delay between radars based on all the discussed parameters is as follows: where N c is the number of chirps per pulse, f s is the ADC sample rate, and T idle is the idle time.The minimum delay based on ( 8) is 0.552 ms with an interpulse repetition time of 49.5 ms.Table I summarizes the designed signal parameters to achieve minimum delay.In conclusion, fusing multiple radars can achieve finer range resolution without interference by designing a tailored signal.This provides significant benefits in precision-critical applications, such as remote sensing and target tracking.However, optimization is crucial for optimal performance.

E. Digital Signal Processing Chain
The FMCW radar is capable of generating periodic signals to scan the surrounding environment.In doing so, the radar produces a chirp signal that sweeps from a low frequency to a high frequency.A group of chirps generated periodically with an idle time between two sequential chirps is called a frame.The radar sensor can generate a couple of frames per second.As a result, the received signals transmitted back from the environment are 3-D, comprising information on range, slow time, and frame.
Fig. 4 depicts the proposed multibin signal processing chain for breathing pattern estimation.First, the dc value of the signal needs to be removed; otherwise, the phase quality is affected [32].Second, the first fast Fourier transform (FFT), which is called fast-time FFT, is applied to a chirp signal to generate a range profile map.The range profile map is discretized by the range resolution of the FMCW radar.Then, a 3-D cube of chirps is reconstructed, which has whole captured information in different ranges.
FMCW radar divides the range into smaller units by range resolution.Each unit is called a bin.The human body in front of a high-range resolution radar captures more than one range bin, especially when a human is close to the radar.In the proposed method, four range bins around the main range bin of the driver are employed to have accurate estimations in the chest radar.The reflected signal from the middle of the chest, which is the first reflection in the chest radar, may appear one range bin before the detected peak.Other organs of the human body capture further range bins.To avoid unnecessary computations for empty range bins, we can determine which range bins are captured by the human body.
A robust novel algorithm is presented in this article to detect the main range bin of the target.First, clutter cancellation is applied to remove zero-doppler reflections [32].Since the human body is supposed to be less active, clutter cancellation is applied on chirps through the frames in a processing Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.window.If the clutter cancellation was applied through the slow time on the chirps in a frame, the target's signals would be suppressed since the time difference between the chirps in a frame is low.
Finally, an outlier removal is applied on detected main range bins from multiple chirps.In terms of outlier removal, if an estimated range bin is not within the following interval, it is considered an outlier: where R is the mean of all detected range bins and σ R is their standard deviation.After removing outliers, the mean of all range bins after outlier removal is selected as the main range bin of the target for breathing pattern estimation.The breathing pattern, which corresponds to chest wall displacement, is the angle of the beat signal in a specific range bin.This angle can be wrapped when the chest wall displacements are large.If the target's phase changes between −π and +π, there is no phase wrapping.Phase wrapping always occurs when the target's displacement exceeds half of the wavelength.It can be calculated as follows: Since a human's chest wall displacement can reach 12 mm, the phase wrapping can usually occur at 60 GHz.The obtained signal after phase unwrapping is called the vibration signal, which has all the vibrations in the selected range bin.The constant range of the target adds a constant phase in the breathing signal, which can be removed by mean subtraction.The breathing signal can be achieved by canceling undesired vibrations, such as heartbeat and random body motion.A fourth-order Butterworth bandpass filter with cutoff frequencies of 0.1 Hz to 0.7 is applied to remove the unwanted vibrations in this article.The detected peak from the FFT of the breathing signal determines the BR.
The chest wall displacement is a criterion that can determine if there is a breath hold.The bandpass filter suppresses vibrations outside of desired band in the vibration signal.Therefore, the estimated frequency by peak detection is often in the breathing band.Since peak detection determines a peak in the breathing band in the frequency domain, there is a peak even when the input signal is noise.As a result, the amplitude of this detected peak should be compared with a threshold.This threshold discriminates breathing periods from breathhold periods.
The threshold-level determination needs to be done statistically by collecting data from different participants.These participants were asked to hold their breath, while there were no other body movements.The amplitude of the detected peak from the filtered vibration signal obtained from 20 s is considered a breath-hold displacement.Approximately 300 measurements were collected from a group of eight participants to estimate statistical parameters.
Since the amplitude of the detected peak in a breath hold scenario is nonnegative, its distribution is log-normal [36].The mean and standard deviation of data on a millimeter scale after taking natural logarithm are −4.6 and 0.9, respectively.The threshold level can be determined using the empirical rule.Since missing a breath-hold period has more cost than having some false detections, the threshold level should be higher than almost all the amplitudes.According to the empirical rule, 99.7% of the amplitudes fall within three standard deviations of the mean [37].The threshold based on the three-sigma rule of thumb is −1.89 and is approximately equivalent to 0.15 after applying the exponential function.If the amplitude of the detected peak is less than 0.15, this period is considered a breath-hold.
The estimated BRs from multiple range bins are compared with each other to determine an accurate estimation in the next step.An outlier removal, which can remove the corrupted range bin by random body motions, is applied to the estimated BRs.Finally, the mean of all the remained estimations determines the BR.The breathing patterns are classified as abnormal if the estimated BR is below 12 breaths per minute (bpm) or above 20 bpm.

III. EXPERIMENTS
The experimental setup, experimental protocols, and the abnormal breathing algorithm are discussed in this section.

A. Experimental Setup
The sensor selection is based on two factors, including the carrier frequency and the size of the radar package.First, higher carrier frequencies have a better signal-to-noise ratio (SNR) due to the noise power reduction by increasing the carrier frequency.The Federal Communications Commission (FCC) also supports the 60-GHz band to be used in in-cabin applications [7].As a result, the 60-GHz band is selected for our proof-of-concept studies.Second, BGT60TR13C from Infineon is employed in this article.The dimensions of the whole package of this radar are 64 × 25.4 mm [23].Due to these compact dimensions, it can be placed inside the vehicle without any discomfort for drivers.
The sensor placement has a direct effect on measured displacement by radar.Three factors should be considered, including the angle between the radar and the human body, the distance between the radar and the human body, and signal blockage by other objects inside a vehicle.First, if the radar is placed in front of displacement, it can estimate displacement accurately since the radar measures radial displacements.A displacement with high amplitude can be less affected by noise.Second, the reflected power in the radar equation is related to distance by a power of four.Therefore, radar measurement can be less affected by noise at low distances.Third, if a portion of radar signals before reflecting from the human body are blocked by objects, especially metallic ones, the SNR of the signal from the human body drops in a longer range.As a result, a radar that is placed in front of displacement in the closest range without being blocked by any objects can estimate displacements more accurately.
Fig. 5 depicts the displacements of the chest and abdomen walls of a sitting human during breathing in and out.Due to the angle between chest displacements and the horizontal axis, the chest radar should be attached to the ceiling while tilted toward the chest.This area is almost unused inside the vehicle.However, the abdomen radar cannot be placed properly since the best placement is under the wheel, which can block most radiated signals.Fig. 6(a) shows the chest and abdomen radars placements in this article.The abdomen radar is tilted toward the middle of the abdomen as can be seen in Fig. 6(b).The distance between the chest radar and the central point of the chest equates to approximately 20 cm, while the distance between the abdominal radar and the midpoint of the abdomen measures approximately 40 cm.The chest radar is affixed to the ceiling via adhesive tape, while the abdominal radar is securely fastened to a holder, allowing for adjustable angulation toward the abdominal center.

B. Experimental Protocol
Eight subjects (four females and four males; age: 30 ± 6 years; weight: 75 ± 20 kg; height: 172 ± 13 cm; and BR: 17 ± 8 bpm) without having any respiratory disease participated in these experiments.The participants were asked to sit behind the wheel as a driver and mimic different breath patterns using an online metronome [39] in a single-subject or dual-subject scenario while wearing a heavy jacket or shirt.In addition, various driving scenarios, including shoulder check (left then right), turning wheel 90 • , and drinking a can, are examined in this study.The participants were asked to do these activities every 30 s.In the shoulder check scenario, the participants checked the left shoulder first.After 30 s, they did a right shoulder check.In the drinking a can scenario, the participants first picked a can from drink holder inside the vehicle and mimicked drinking and then returned the can to the drink holder.Whole the scenario was almost 4-5 s.
Estimated BR is compared with the actual BR based on the root mean square error (RMSE).The RSME can be Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.calculated as follows: where N is the number of data, y(n) is the nth measurement, and ŷ(n) is its corresponding prediction.

C. Recognition of Abnormal Breathing Algorithms
Five abnormal breathing patterns, including Tachypnea, Bradypnea, Biot, Cheyne-Stokes, and Apnea, are investigated in the driving scenario in this study.Tachypnea is characterized by a BR of more than 20 bpm.On the other hand, Bradypnea is low BR, which is characterized by less than 12 bpm.Biot is characterized by some shallow breath patterns followed by some periods of Apnea.Cheyne-Stokes is characterized by two stages.First, there is a progressive increment in both deepness and rate.Then, both the deepness and rate reduce gradually, which results in Apnea at the end [40].
Fig. 7 represents two straightforward approaches to detect different abnormal breathing patterns, including BR estimation and breath-hold detection.The BR estimation can be used for the detection of both Tachypnea and Bradypnea, which are BRs of more than 20 bpm and less than 12 bpm, respectively.Breath-hold detection can be employed in the diagnosis of three different abnormalities, including Biot, Cheyne-Stokes, and Apnea.Both Biot and Cheyne-Stokes have an Apnea period.Therefore, all these three abnormalities can be diagnosed by beath-hold detection.It is worth noting that this study focuses on detecting abnormalities in various driving scenarios.

IV. RESULTS AND DISCUSSION
A. Validation of the Proposed Algorithm Fig. 8 compares the effect of two different clutter cancellations, including clutter cancellation on chirps through all the frames and the clutter cancellation on a frame, on target detection in the chest radar.The target cannot be detected if the clutter is canceled in a frame because target reflections are also suppressed by clutter cancellation.However, if the clutter is suppressed through all the frames, it can detect the target in a true range after suppressing clutters reflections.Clutter cancellation applied on the range profile is only deployed for accurate range detection.However, clutter cancellation can affect the estimated breathing pattern.Fig. 9 evaluates the effect of clutter cancellation applied on range profile in breathing pattern estimation.Although the measurement has a breath-hold period, the processed signals by clutter cancellation cannot represent any breath-hold period.However, while the chest wall displacement is low, unprocessed signals accurately represent the breath-hold period.
Fig. 10 depicts the various reflections in different ranges in both the chest and abdomen radars over time.There is a breath-hold in this scenario, which is specified in both radars accurately.First, both radars represent the target in the correct range as they are at 20 and 40 cm, respectively.Second, radar detects multiple range bins for a subject as it is close to the subject.The chest radar can detect four range bins for a subject in this study.Third, both radars are not impacted by interference like other subjects' reflections since they are out of maximum range.Fourth, the chest radar is less affected by multiple reflections than the abdomen radar.As a result, the chest radar is reliable for multibin breathing pattern estimation, and the abdomen radar can associate with the chest radar in detecting breath-hold periods.
An intuitive way to improve frequency resolution is to increase the processing window length.A longer processing window is also less susceptible to interference like body motions.Although a longer processing window is more robust to interference, it is less adaptive to BR changes.Given that the minimum normal BR is 12 bpm, the duration of one breath is 5 s.The minimum window length is 20 s by considering four cycles of this breathing pattern.Fig. 11 shows the consistency of BR estimation in 35 s, while BR is 16 for different processing window lengths, including 20, 30, 40, and 50 s.The BR is estimated every second for these window lengths.A participant was asked to turn the wheel every 30 s in this measurement while mimicking different BRs.The results indicate that a window length of 40 s can estimate BR by approximately a maximum error of less than 1 bpm.For longer windows, the accuracy would be slightly higher than 40 s.However, longer windows are unable to detect breath-hold periods accurately.If we assume that the minimum duration of breath-hold is 20 s, the window length of 40 s covers 50% of this period.Therefore, longer windows may lead to missing breath-hold periods.
To examine the interference effect in dual radar system depicted in Fig. 6(a), three disparate measurements are performed.First, the radars started to radiate signals simultaneously while there was no metallic surface in front of the abdomen radar.Therefore, the radars interfered with each other, and the power level of interference in the chest radar obtained from the range-doppler map after zero-doppler cancellation exceeds −30 dB in the chest range, as shown in Fig. 12(a).Second, the metallic surface was placed in front of abdomen radar to avoid interference while the time delay was still zero.Fig. 12(b) illustrates the suppression of the interference level by almost 3.5 dB.Finally, the metallic surface was removed, and a time delay of 5 µs was set between two radars.Fig. 12(c) demonstrates a 3-dB suppression in the power level of interference after considering this time delay.periods accurately.They tend to overestimate breath-hold periods, but the use of dual radar improves accuracy.
Table II compares the single-bin and multibin approaches based on the average RMSE in different BRs inside the vehicle while the subject had a slight body movement.Four subjects participated in the data collection for each BRs.The length of measurement for each BRs was 3 min.The multibin approach exhibits lower RMSE in most cases, especially in high BRs.However, the proposed multibin approach exhibits a slight  Nevertheless, the introduced error is negligible.The maximum error in the multibin approach reaches 1.2 bpm.However, the RMSE of the single-bin approach exceeds 2 bpm when BR is high.
Table III represents the average RMSE of various BRs in driving scenarios obtained from the chest radar.Four different scenarios are examined in this study, including turning the wheel, shoulder check, shoulder check while wearing a thick jacket, and drinking a can.Eight subjects are participated in this measurement for 3 min for each BR.The participants were asked to do these activities every 30 s.Among the examined scenarios, turning the wheel is the most challenging

B. Discussion
Table IV presents a comparison between the approaches utilized in this article and those described in recently published articles.First, it is noteworthy that most articles report a maximum estimation error for breath rate (BR) of less than 2 bpm.Similarly, the multibin approach proposed in this article also demonstrates an error within this range.Second, the processing window length is another important criterion addressed in this study.While many recent articles suggest a window length of 60 s, we propose a window length of 40 s.This choice is motivated by the fact that a longer window fails to accurately detect breath-hold periods.Additionally, it is worth mentioning that the analysis of driving scenarios, such as steering wheel movements, is a prevalent focus in recent studies.
Furthermore, previous studies mainly focused on accurately estimating BR rather than detecting abnormalities.Similar to our approach, many of these studies achieved an RSME of less than 2 bpm in BR The proposed approach achieved a maximum estimation error of 1.9 bpm.It is worth noting that the BR estimation without considering the chest wall and abdomen wall displacements cannot detect breath-hold periods and breathing abnormalities.The chest radar incorrectly identified 37 periods as breath-hold, while the abdomen radar incorrectly 26 periods as breathhold.However, there is no false detected beath-hold by the proposed dual radar approach.
The proposed approach in this article handling high random body motions within a processing window.One of the challenging motions inside the vehicle is turning the wheel, which is a periodic movement [8].If this activity lasts more than 10 s, the BR estimation is challenging, especially when the actual BR is more than 20.Although the BR usually drops to less than 10 bpm and RMSE is high, the abnormality is detected even if mislabeled.This phenomenon commonly occurs when the driver is stressed and steers the wheel sharply.
This study focused on the detection of abnormal breathing by dual radar.However, dual radar application is not limited to the detection of breathing pattern estimation.The abdomen radar in the proposed setup in this article can be employed for drowsy driver detection.Since normal breathing is a nonstop activity, the detection of drowsy drivers based on the BR is less reliable.
Because people are less active during sleep, the second can be employed for activity recognition.As a result, if the driver is less active than normal and BR is also low, it can be considered a drowsy V. CONCLUSION Radar, as a contactless sensor, offers the benefit of operating in various environmental conditions while preserving privacy.This study employed dual radar fusion system to monitor the respiratory patterns of drivers.This system can detect five breathing abnormalities that are more possible to happen for drivers, including Tachypnea, Bradypnea, Biot, Apnea.Two features are extracted from breathing patterns to detect abnormalities: BR and displacement amplitude.The maximum error of BR estimation this article by multibin method reached 1.9 breaths in the driving scenario as noted Table IV.However, the error for a single bin approach 4.36 breaths.The proposed dual radar can recognize the breath-hold period without false detections.These fused radars have less interference with each other while they benefit from the finest range resolution.Furthermore, a proper signal design can mitigate the effect of passengers' reflections, which could otherwise interfere with the driver's reflections.

Manuscript received 24
July 2023; revised 17 November 2023; accepted 6 December 2023.Date of publication 22 December 2023; date of current version 4 January 2024.This work was supported in part by NSERC and in part by MITACS.The Associate Editor coordinating the review process was Dr. Zhengyu Peng.(Corresponding author: Ali Gharamohammadi.)

Fig. 3 .
Fig. 3. Dual radar placement in front of chest and abdomen.

Fig. 5 .
Fig.5.Displacements of the chest and abdomen walls of a sitting human during breathing in and out[38].

Fig. 7 .
Fig. 7. Approaches to detect different abnormal breathing patterns in this article.

Fig. 8 .
Fig. 8.Comparison between different clutter cancellation approaches on the target detection in chest radar.

Fig. 9 .
Fig. 9. Effect of clutter cancellation on breathing pattern estimation while there is a breath-hold in the scenario.(a) Estimated breathing pattern with clutter cancellation.(b) Estimated breathing pattern without clutter cancellation.

Fig. 13
Fig.13depicts the amplitude of the detected peak in the breathing pattern in the frequency domain over time while there were five breath-hold periods in this measurement from chest and abdomen radars, respectively.Both chest and abdomen radars can detect these breath-hold periods.However, two false detections with low displacements were detected in only one of the radars.As a result, reliable breath-hold detection can be achieved by dual radar fusion system.Radar signals are sensitive to movements and can detect all breath-hold

Fig. 12 .
Fig. 12. Power level of interference on the range-velocity map in different measurements.(a) No time delay and no metallic surface.(b) No time delay and a metallic surface.(c) Proper time delay and no cupper surface.

Fig. 13 .
Fig. 13.Amplitude of the detected peak in the breathing pattern in the frequency domain over time while there were five breath-hold periods in this measurement.(a) Chest radar.(b) Abdomen radar.

TABLE II COMPARISON
BETWEEN SINGLE-BIN AND MULTIBIN APPROACHES BASED ON AVERAGE RMSE (BPM) IN VARIOUS BRS INSIDE THE VEHICLE degradation in BR estimation when BR values are 14 or 16.

TABLE III AVERAGE
RMSE (BPM) OF CHEST RADAR IN DIFFERENT BRS IN DRIVING SCENARIOS