Table I and enhance the signal SNR with receiving-end beamforming PROCESSING TIME ON DIFFERENT PLATFORMS to conduct infant respiration monitoring.Unlike traditional Single User Multiple Users processing methods for FMCW signal,CFMCW [14]uses PC (RESPTRACKER) 0.1338s 0.3030s cross-correlation to increase the accuracy of the acoustic- PC (Delay-and-Sum) 14.5805s based breath sensing.Xu et al.[27]leverage Energy Spectrum Raspberry Pi (RESPTRACKER) 2.0830s 3.7712s Density(ESD)of a single-frequency acoustic signal,Ensemble Empirical Mode Decomposition,and Generative Adversarial different devices using different methods.The source data is Network to reconstruct the breathing signal in a driving six-channel recorded sound signals at a sample rate of 48 kHz scenario.The vital problem of acoustic based approach is the with 32 bit float precision.We observe that our algorithm out attenuation of sound in the air,which makes it hard to expand performs the delay and sum scheme by more than ten times. the sensing range to more than two meters. Even on resource constrained platforms like Raspberry Pi,our Beamforming of Wireless Signals:Beamforming is an system can handle the incoming data efficiently where the important technique in wireless communication,as it can average processing time for a 20 s observation slot is 2.0830 s enhance signal strength from/to different direction with receiv- and 3.7712 s for single user and multiple users.The reason ing/transmitting array.In the wireless transmission scenario, why the delay-and-sum method so slow is that it has to search Phaser [28]enables phase array signal processing on COTS exhaustively in all azimuth and elevation to find the valid delay device which increases spatial resolution,decreases phase and combine the signal. error and suppressed the multipath interference.Wang et al. [29]use a blind distributed beamforming on both uplink and VI.RELATED WORK downlink to increase the backscatter sensing distance to 64 We summarize recent works related to respiration tracking meters.For wireless sensing and tracking,mdTrack [30]gives according to the following three categories. a multi-dimensional Wi-Fi localization model and drastically Respiration Monitoring with Wireless Signals:Wireless increases passive localization divisibility.Vasisht et al.[31] signals are widely used for non-invasive vital sign monitoring analyze time of flight in different frequency band and between [9],[21]-[23].BreathTaking [21]leverages the received signal different TX/RX pairs to reach decimeter-level localization strength between different pairs of network devices to conduct accuracy.WiDar 2.0 [32]and FreeSense [33]calculate the the contactless breath monitoring for single person on the bed. AoA and ToF to match and localize the moving subject.In DeepBreath[9]uses multiple FMCW transceivers and the ICA sound signal processing,Roy et al.[34]set up a speaker array algorithm to separate different users'respiration signal.Liu et to achieve long-range ultrasound attacks on voice assistants. al.[22]extract breath and heart beats from the CSI gathered Moutinho et al.[35]address the inverse problem of localizing by commodity Wi-Fi devices.Wang et al.[11]propose the microphones with speaker arrays that are playing predefined Fresnel Zone model of Wi-Fi sensing in which the subject's sounds.Shen et al.[36]leverage microphone array and respiration can be hardly recognized by the CSI amplitude.To reflections from the wall to localize sound source.Most of tackle this challenge,Zeng et al.[23]exploit the complement- the existing beamforming techniques ignore the possibility of ary between amplitude and phase of complex CSI data to cover utilizing the multipath effect to enhance the received signal. the blind point of Fresnel Zone and further use CSI ratio of two antennas to calculate the accurate phase of reflections [24]. VII.CONCLUSION Yang et al.[25]leverage the high distance resolution of UWB radars to separate different subjects'respiration and use In this paper,we present new insights on how to tackle image processing techniques to detect sleep apnea.ViMo [26] the design challenges for long-range,multiple users domestic leverages the high spatial resolution (distance,azimuth and respiration tracking systems.We propose to exploit the mul- tipath effect to recombine the reflections in order to improve elevation)of 60GHz millimeter wave antenna array to extract both the respiration rate and the heart rate of multiple subjects. system sensitivity and robustness.In this way,we expand the sensing range of acoustic respiration patterns from the Although wireless devices can provide strong signal with good quality,they are expensive and the monitoring process might 0.7 to 1.0 meters in previous works to a room-scale of 3.0 interfere with normal data transmissions. to 4.0 meters.We believe our new insights could bring new Respiration Monitoring with Acoustic Signals:Acoustic opportunity for domestic sensing application. signal travels much slower than wireless signals.The sampling rate of 48 kHz from COTS microphones provides a fine range VIII.ACKNOWLEDGEMENT resolution of 0.7 cm.while similar resolution on RF-based We would like to thank our anonymous reviewers for systems requires Gigahertz of bandwidth.So,recent works their valuable comments.This work is partially supported by exploit acoustic signal to perform device-free breath sensing. National Natural Science Foundation of China under Numbers Apneapp [10]transmits 18~20 kHz FMCW sound signals to 61872173,61902177,61972254 and 61832005,Natural Sci- estimate breathing frequency and detect sleep apnea passively.ence Foundation of Jiangsu Province of China under number Wang et al.[13]expand the frequency band of acoustic signal BK20190298 and Collaborative Innovation Center of Novel by transforming audible white noises into FMCW signals Software TechnologyTable I PROCESSING TIME ON DIFFERENT PLATFORMS Single User Multiple Users PC (RESPTRACKER) 0.1338s 0.3030s PC (Delay-and-Sum) 14.5805s - Raspberry Pi (RESPTRACKER) 2.0830s 3.7712s different devices using different methods. The source data is six-channel recorded sound signals at a sample rate of 48 kHz with 32 bit float precision. We observe that our algorithm out performs the delay and sum scheme by more than ten times. Even on resource constrained platforms like Raspberry Pi, our system can handle the incoming data efficiently where the average processing time for a 20 s observation slot is 2.0830 s and 3.7712 s for single user and multiple users. The reason why the delay-and-sum method so slow is that it has to search exhaustively in all azimuth and elevation to find the valid delay and combine the signal. VI. RELATED WORK We summarize recent works related to respiration tracking according to the following three categories. Respiration Monitoring with Wireless Signals: Wireless signals are widely used for non-invasive vital sign monitoring [9], [21]–[23]. BreathTaking [21] leverages the received signal strength between different pairs of network devices to conduct the contactless breath monitoring for single person on the bed. DeepBreath [9] uses multiple FMCW transceivers and the ICA algorithm to separate different users’ respiration signal. Liu et al. [22] extract breath and heart beats from the CSI gathered by commodity Wi-Fi devices. Wang et al. [11] propose the Fresnel Zone model of Wi-Fi sensing in which the subject’s respiration can be hardly recognized by the CSI amplitude. To tackle this challenge, Zeng et al. [23] exploit the complement￾ary between amplitude and phase of complex CSI data to cover the blind point of Fresnel Zone and further use CSI ratio of two antennas to calculate the accurate phase of reflections [24]. Yang et al. [25] leverage the high distance resolution of UWB radars to separate different subjects’ respiration and use image processing techniques to detect sleep apnea. ViMo [26] leverages the high spatial resolution (distance, azimuth and elevation) of 60GHz millimeter wave antenna array to extract both the respiration rate and the heart rate of multiple subjects. Although wireless devices can provide strong signal with good quality, they are expensive and the monitoring process might interfere with normal data transmissions. Respiration Monitoring with Acoustic Signals: Acoustic signal travels much slower than wireless signals. The sampling rate of 48 kHz from COTS microphones provides a fine range resolution of 0.7 cm, while similar resolution on RF-based systems requires Gigahertz of bandwidth. So, recent works exploit acoustic signal to perform device-free breath sensing. Apneapp [10] transmits 18∼20 kHz FMCW sound signals to estimate breathing frequency and detect sleep apnea passively. Wang et al. [13] expand the frequency band of acoustic signal by transforming audible white noises into FMCW signals and enhance the signal SNR with receiving-end beamforming to conduct infant respiration monitoring. Unlike traditional processing methods for FMCW signal, CFMCW [14] uses cross-correlation to increase the accuracy of the acoustic￾based breath sensing. Xu et al. [27] leverage Energy Spectrum Density (ESD) of a single-frequency acoustic signal, Ensemble Empirical Mode Decomposition, and Generative Adversarial Network to reconstruct the breathing signal in a driving scenario. The vital problem of acoustic based approach is the attenuation of sound in the air, which makes it hard to expand the sensing range to more than two meters. Beamforming of Wireless Signals: Beamforming is an important technique in wireless communication, as it can enhance signal strength from/to different direction with receiv￾ing/transmitting array. In the wireless transmission scenario, Phaser [28] enables phase array signal processing on COTS device which increases spatial resolution, decreases phase error and suppressed the multipath interference. Wang et al. [29] use a blind distributed beamforming on both uplink and downlink to increase the backscatter sensing distance to 64 meters. For wireless sensing and tracking, mdTrack [30] gives a multi-dimensional Wi-Fi localization model and drastically increases passive localization divisibility. Vasisht et al. [31] analyze time of flight in different frequency band and between different TX/RX pairs to reach decimeter-level localization accuracy. WiDar 2.0 [32] and FreeSense [33] calculate the AoA and ToF to match and localize the moving subject. In sound signal processing, Roy et al. [34] set up a speaker array to achieve long-range ultrasound attacks on voice assistants. Moutinho et al. [35] address the inverse problem of localizing microphones with speaker arrays that are playing predefined sounds. Shen et al. [36] leverage microphone array and reflections from the wall to localize sound source. Most of the existing beamforming techniques ignore the possibility of utilizing the multipath effect to enhance the received signal. VII. CONCLUSION In this paper, we present new insights on how to tackle the design challenges for long-range, multiple users domestic respiration tracking systems. We propose to exploit the mul￾tipath effect to recombine the reflections in order to improve system sensitivity and robustness. In this way, we expand the sensing range of acoustic respiration patterns from the 0.7 to 1.0 meters in previous works to a room-scale of 3.0 to 4.0 meters. We believe our new insights could bring new opportunity for domestic sensing application. VIII. ACKNOWLEDGEMENT We would like to thank our anonymous reviewers for their valuable comments. This work is partially supported by National Natural Science Foundation of China under Numbers 61872173, 61902177, 61972254 and 61832005, Natural Sci￾ence Foundation of Jiangsu Province of China under number BK20190298 and Collaborative Innovation Center of Novel Software Technology