Rotation Speed Sensing with mmWave Radar

Rong Ding, Haiming Jin† , Dingman Shen

Shanghai Jiao Tong University, Shanghai, China
Email: {dingrong, jinhaiming, 631322231}@sjtu.edu.cn
IEEE INFOCOM 2023 - IEEE Conference on Computer Communications | 979-8-3503-3414-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/INFOCOM53939.2023.10229066

Abstract—Machines with rotary parts are prevalent in indus- for the illumination conditions. The laser tachometers [4] fail
trial systems and our daily lives. Rotation speed measurement to work, if the target object is highly reflective. In such
is a crucial task for monitoring machinery health. Previous case, the laser tachometers cannot distinguish the laser pulse
approaches for rotation speed sensing are constrained by limited
operation distance, strict requirement for illumination, or strong reflected by its reflective tag from those reflected by the object.
dependency on the target object’s light reflectivity. In this work, In this paper, we propose mRotate, a mmWave radar-based
we propose mRotate, a practical mmWave radar-based rotation system that can accurately sense the rotation speed of the target
speed sensing system liberated from all the above constraints. object from a safe distance, and is robust to the illumination
Specifically, mRotate separates the target signal reflected by the conditions and the target object’s light reflectivity. Among all
rotating object from the mixed reflection signals, extracts the
high-quality rotation related features, and accurately obtains the mmWave radars, we adopt the commercial TI IWR1443
the rotation speed through the customized radar sensing mode frequency-modulated continuous wave (FMCW) radar. Thanks
and algorithm design. We implement mRotate on a commercial to the over 77GHz high operation frequencies, the radar’s
mmWave radar and extensively evaluate it in both lab environ- signal is sensitive to the rotation incurred displacement on
ments and in a machining workshop for field tests. mRotate the target object’s surface. Besides, the over 1e3Hz high
achieves an MAPE of 0.24% in accuracy test, which is 38% lower
than that produced by the baseline device, a popular commercial chirp rate enables the radar to timely capture the rapid signal
laser tachometer. Besides, our experiments show that mRotate variations caused by high speed rotation. In what follows, we
can measure a spindle whose diameter is only 5mm, maintain elaborate on the challenges of designing mRotate, as well as
a high accuracy with a sensing distance as far as 2.5m, and our approaches that address them.
simultaneously measure the rotation speeds of multiple objects. In real practice, the rotating object is surrounded by other
I. I NTRODUCTION reflectors in the environment, and the received signal of the
radar is a mixture of the signals reflected by both the rotating
Machines with rotary parts play significant roles in not
object and other reflectors. Thus, the first challenge is to
only our daily lives, but also various industrial fields (e.g.,
extract our target signal reflected by the rotating object from
machining, automobile, aviation). Accurate and safe rotation
the complicated mixed signals. To address this challenge, we
speed sensing is of great importance for monitoring machinery
utilize the unique property of the target signal that varies
health. Operators in machining workshops are required to
periodically as the object rotates. Based on that observation,
regularly check the rotation speeds of the spindles to monitor
we propose the rotating object discovery mechanism, which
their operation statuses [1]. Checking the rotation speeds of
first separates the signals based on their propagation distances
the wheels has become a standard item for car inspection [2].
through the Range-FFT operation, and then accurately selects
Apart from the industrial applications, the rotation speeds of
our target signal by discovering the one with the periodic
the electric motors in washing machines and air conditioners
varying property through checking the existence of distinct
are used to diagnose the causes of anomalies [3].
peaks in its autocorrelation series.
Thus far, a series of sensing techniques have been exploited
The second challenge is that, when we measure the rotation
to develop different types of rotation speed sensors (i.e.,
speed of roughly centrosymmetric objects (e.g., wheels, fans),
tachometers), which can be categorized into contact-based and
the aliasing problem could mislead us to get the wrong rotation
contactless ones. Specifically, the contact-based tachometers
speed. Specifically, the signals reflected by the centrosymmet-
[4] are physically attached to the spinning shaft of the target
ric rotating parts are highly similar and show an aliasing period
object, which restrains the distance between the operator and
which is easy to be confused with the rotation period. We
the rotating object and incurs safety risks. Existing contactless
resolve such challenge by designing the aliasing elimination
tachometers utilize electromagnetic or optical signals to sense
mechanism, which employs chirp decomposition to obtain
the rotation speed. Among them, the electromagnetic-based
multiple descriptions of the rotating object, and carefully
methods [5, 6] apply the electrostatic or hall-effect sensors
aggregates the descriptions through feature aggregation to
to detect the changes of the electromagnetic field caused by
generate aliasing free signal feature group sum value (GSV).
the rotation. However, these approaches have very limited
Our design of the GSV ensures that it significantly amplifies
sensing distances which are typically less than 10cm. The
the difference between the centrosymmetric rotating parts, and
optical-based approaches [7–13] utilize optical sensors, such
thus makes the aliasing period disappear in the GSV series.
as cameras and laser receivers, for rotation speed sensing.
Meanwhile, note that the radar commonly generates chirps
However, camera-based tachometers have strict requirements
following a fixed chirp generation frequency in most of the
† Correspondingauthor. common sensing applications. However, we realize that only
This work was supported by NSF China (No. U20A20181, U21A20519). relying on the radar’s output data collected under such a

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 reflecting surface (ERS). The signal reflected by the ERS is


SRx (t) = α(t)ST x t − 2R(t)/c , (2)
where R(t) is the ERS-radar distance, and α(t) = β(R(t))γ(t)
is the path loss with β(R(t)) denoting the propagation loss
and γ(t) denoting the reflection loss. Note that R(t) is not
only time-varying but also periodic, because the target object’s
(a) System overview. (b) Radar signals. rotation incurs a periodic change to the shape of the ERS, and
Figure 1: Rotation speed measurement with mmWave radar.
further to R(t). Besides, α(t) is also periodic, because β(R(t))
fixed chirp generation frequency will yield multiple rotation changes with R(t), and γ(t) is periodic as well. After receiving
speed candidates rather than a single accurate solution. To SRx (t), the radar applies a mixer on ST x (t) and SRx (t) to
solve this challenge, we propose the chirp frequency hopping obtain the beat frequency (BF) signal
mechanism that sets the radar to transmit chirps under multiple
s(t) = ST∗ x (t)SRx (t) ≈ α(t)exp j4π(fc + Lt)R(t)/c , (3)


chirp generation frequencies to obtain multiple observations of
the rotating object. Our meticulously designed rotation speed and outputs the raw ADC data s[n] which is a sampled version
search algorithm then carefully explores the rotation speed of the original BF signal s(t).
search space and accurately pinpoints the rotation speed which After obtaining the sampled BF signals output by the radar,
is the one that aligns with the most observations. we then aim to extract from them the signal features related to
In summary, this paper makes the following contributions. the rotation speed. In Sec. III, we will introduce our proposed
• We design and implement a novel mmWave radar-based signal feature index of the first peak (IFP) that bridges the
rotation speed sensing system, named as mRotate, which radar signal characteristics and the rotation speed.
enables accurate rotation speed sensing from a safe distance, III. ROTATION R ELATED F EATURE : IFP
and is robust to the illumination conditions and the target
object’s light reflectivity. In this section, we introduce the notion of our proposed
• Technically, we propose a series of novel approaches that signal feature IFP, whose relationship with the target object’s
(i) separate the signal reflected by the rotating object from rotation speed acts as the theoretical foundation of mRotate.
the complicated mixed reflection signal, (ii) generate the A. IFP Extraction
aliasing free signal features to enable accurate rotation speed After obtaining the raw ADC data s[n] of the k-th BF signal,
sensing for centrosymmetric objects, and (iii) pinpoint the we conduct the Range-FFT operation on s[n], which takes
accurate rotation speed through customized radar sensing FFT on consecutive samples of s[n] to obtain its frequency
mode and algorithm design. spectrum, and maps each frequency spectrum component to
• We conduct extensive experiments to evaluate mRotate on the discrete range bins. The frequency spectrum component
a variety of rotating objects in both lab environments and corresponding to the range bin that the ERS-radar distance
field tests. mRotate achieves an MAPE of only 0.24% for belongs to is referred to as the ERS component Fk , where1
accuracy test in lab environments, which is 38% lower than



that produced by the baseline device, a popular commercial Fk = α (k − 1)TI exp j4πfc R (k − 1)TI /c . (4)
laser tachometer. Meanwhile, mRotate can measure the
Clearly, Fk is a sample of a continuous signal F (t) with
rotation speed of a small spindle whose diameter is only

5mm, achieve low errors from a sensing distance as far F (t) = α(t)exp j4πfc R(t)/c . (5)
as 2.5m, and simultaneously obtain the rotation speeds of
Since the period of α(t) and R(t) equals to the target object’s
multiple objects in the field tests.
rotation period T , F (t) is also periodic with period T .
II. P RELIMINARIES To investigate the temporal properties of F (t), we set the
radar to consecutively transmit N chirps and obtain the N ERS
In this section, we first introduce the basic signal propaga- components2 {Fk }k∈[N ] , which compose a sampling series of
tion model of using mmWave radar for rotation speed sensing. F (t) with a sampling frequency 1/TI . We refer to such a
The radar transmits FMCW chirps to the target object to sense series as the rotation signal series F ′ . After obtaining F ′ , we
its rotation speed. The transmitted signal of the k-th chirp is then calculate its autocorrelation series X. Specifically, X is a
discrete series of length N −1, whose p-th value X[p] satisfies
ST x (t) = exp j(2πfc t + πLt2 ) ,


(1)
N −p
where fc is the chirp starting frequency, L is the chirp slope, 1 X
X[p] = F ′ [n]F ′ [n + p], (6)
and t ∈ [(k − 1)TI , (k − 1)TI + Tc ] with TI denoting the time N − p n=1
interval between the generation of two consecutive chirps and
1 Since the duration of a chirp is only around 0.1ms in our experiments, the
Tc denoting the duration of one chirp.
rotation of the target object can be neglected during this period, and R(t) can
As illustrated in Fig. 1a, we refer to the part of the target be regarded always equal R((k−1)TI ) with t ∈ [(k−1)TI , (k−1)TI +Tc ].
object’s surface that reflects the radar signal as the effective 2 We denote [A] = {1, 2, · · · , A} for any positive integer A.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 Figure 2: The framework of mRotate.

where F ′ [n+p] is the conjugate of F ′ [n+p]. According to the ALGORITHM 1: Chirp Frequency Hopping.
characteristic of the autocorrelation series, a maximum value Input: A group C = {1/Ti }i∈[H] of chirp generation
of X occurs at index p∗ with frequencies;
Output: The H × N groups of raw ADC data
p∗ TI = mT, (7) {si,j }i∈[H],j∈[N ] of BF signals;
1 foreach i ∈ [H] do
where m referred to as the rotation period coefficient is any 2 foreach j ∈ [N ] do
positive integer. We define the smallest p∗ that satisfies Eq. 3 The radar transmits the j-th chirp following the
(7) as the index of the first peak (IFP). chirp generation frequency 1/Ti ;
4 The radar obtains the raw ADC data si,j of the
B. IFP and mRotate corresponding BF signal;

Through the IFP value, we bridge the chirp generation in- 5 return {si,j }i∈[H],j∈[N ] ;
terval TI and the target object’s rotation speed 1/T according
to Eq. (7). However, in real applications, obtaining the rotation In the rest of this section, we focus on illustrating the
speed from the IFP is anything but trivial. We enumerate the rotation speed sensing for a single object. In Sec. V-D2, we
three key challenges of such task as follows. will show that mRotate is able to simultaneously measure
multiple objects after straightforward augmentations .
• In real applications, the range bin that the ERS-radar dis-
tance belongs to is a priori unknown. How can we correctly A. Data Collection
find such range bin to build the autocorrelation series X?
In most of the common sensing applications (e.g., motion
• As we will show in Sec. IV-B, the aliasing problem makes
speed, distance, and material sensing), the mmWave radar
the IFP overwhelmed in the interferences. How can we
transmits a fixed number of chirps in each unit time interval.
robustly extract the true IFP in such scenario?
However, such a chirp generation mechanism is unsuitable for
• Even if we obtain the correct value of IFP, according to
rotation speed sensing for the following reason. We can only
Eq. (7), we can not directly solve the rotation speed, since
extract one IFP value from the chirps generated following a
the value of the rotation period coefficient m is unknown.
fixed generation frequency. In such case, according to Eq. (7),
How can we accurately get the rotation speed from the IFP
we cannot directly acquire the T because the value of the
values under multiple choices of rotation period coefficient?
rotation period coefficient m is unknown. Actually, we need
In Sec. IV, we will introduce the details of mRotate, which to extract multiple IFP values under different chirp generation
solves the above challenges and exploits the properties of IFP frequencies to build multiple equations in the form of Eq. (7)
to accurately sense the rotation speed. to solve T . To this end, in the DC module, we propose the
CFH mechanism described in Alg. 1.
IV. M ROTATE D ESIGN
At first, the algorithm takes a set C of chirp generation
In this section, we introduce our design details of mRotate, a frequencies as input. After that, the radar generates consecutive
mmWave radar-based rotation speed sensing system. As shown N chirps following each chirp generation frequency in C and
in Fig. 2, mRotate consists of the following modules. outputs the corresponding raw ADC data of the BF signals
• Data Collection (DC). In the DC module, the mmWave (line 3-4). Finally, the algorithm outputs the H × N groups
radar transmits chirps according to our carefully designed of raw ADC data {si,j }i∈[H],j∈[N ] (line 5). The obtained raw
Chirp Frequency Hopping (CFH) mechanism. After that, the ADC data will be further utilized by the FE module to extract
radar outputs the raw ADC data of the BF signals. H IFP values as elaborated in Sec. IV-B.
• Feature Extraction (FE). The FE module takes the raw
ADC data as inputs, and outputs the IFP values. The B. Feature Extraction
IFP values are extracted from the raw ADC data by the 1) Challenges of Extracting IFP: In Sec. III, we introduce
Aliasing Elimination (AE) sub-module and Rotation Object the basic steps to extract IFP from the raw ADC data.
Discovery (ROD) sub-module. However, in real applications, two key challenges prevent us
• Rotation Speed Search (RSS). The RSS module takes in from acquiring the accurate IFP.
the IFP values extracted by the FE module and obtains the The first challenge is that the aliasing problem makes it
rotation speed through the rotation speed search algorithm. difficult to obtain the true value of IFP. To illustrate it, we

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 (a) The in-phase components of the rotation signal series F ′ . (a) The in-phase components of GSV series.

(b) The amplitudes of the autocorrelation series of F ′ . (b) The amplitudes of the autocorrelation series corresponding to the GSV series.
Figure 3: The in-phase components of F ′ in our six-blade fan sensing Figure 4: The in-phase components of the GSV series in our six-blade fan
experiment and its autocorrelation series, with the red points indicating sensing experiment and its autocorrelation series, with the red points indicating
periodic rotation peaks and the yellow points indicating the aliasing peaks. the periodic rotation peaks.
conduct an experiment that applies the mmWave radar to sense extracted H IFP values. For each N groups of the raw ADC
the rotation speed of a six-blade fan. We collect the raw ADC data obtained under the same chirp generation frequency, the
data of 500 consecutive chirps under a fixed chirp generation FE module feeds them to the AE sub-module to acquire B
frequency, and obtain the rotation signal series F ′ shown in aliasing free autocorrelation series (line 2). After that, the
Fig. 3a. The autocorrelation series of F ′ is plotted in Fig. 3b, obtained B autocorrelation series will be fed into the ROD
which contains the following two kinds of peaks. sub-module to discover the range bin that the ERS-radar
• Periodic rotation peaks. These peaks are plotted in red dots distance belongs to and extract the IFP value (line 3). After
in Fig. 3b and are caused by the periodic rotation of the H iterations, the algorithm outputs the H IFP values (line 4).
target object. The index of the smallest periodic rotation
peaks is the IFP and the indices of the remaining rotation ALGORITHM 2: Feature Extraction.
peaks are the multiples of the IFP that also satisfy Eq. (7). Input: The H × N groups of the raw ADC data
• Aliasing peaks. Between two consecutive periodic rotation {si,j }i∈[H],j∈[N ] of the BF signals;
peaks, there are five aliasing peaks plotted in yellow dots Output: The H IFP values {pi }i∈[H] ;
1 foreach i ∈ [H] do
in Fig. 3b. Specifically, autocorrelation is a similarity mea- // Aliasing Elimination.
surement of the original series and its time-lagged version. 2 {Xb }b∈[B] ← AE({si,j }j∈[N ] );
The centrosymmetric structure of the rotating fan makes the // Rotating Object Discovery.
pattern of the signal reflected by each of the fan’s six blades 3 pi ← ROD({Xb }b∈[B] );
highly similar. Such similarity also exists in F ′ , since it is 4 return {pi }i∈[H] ;
the description of the BF signals in the frequency domain.
Thus, the in-phase3 components of F ′ shows a “virtual
3) Aliasing Elimination: As elaborated in Alg. 3, the AE
period” which is six times shorter than its true period as
sub-module takes the N groups of raw ADC data {si,j }j∈[N ]
shown in Fig. 3a, and five aliasing peaks occur between the
collected under the chirp generation frequency 1/Ti as inputs
two consecutive periodic rotation peaks as shown in Fig. 3b.
and outputs the aliasing free autocorrelation series {Xb }b∈[B]
As shown in Fig. 3b, since the amplitudes of the aliasing by our Chirp Decomposition (CD) and Feature Aggregation
peaks are very close to those of the periodic rotation peaks, we (FA) mechanism. At first, the algorithm applies the CD mech-
may easily confuse the aliasing peaks with periodic rotation anism to decompose each of the sampled BF signals si,j
peaks and further obtain the wrong IFP values. Since many into K sampled sub-BF signals {ck,j }k∈[K] (line 2). After
rotating objects are roughly centrosymmetric (e.g., wheels, that, the algorithm applies the FA mechanism to generate the
drills, fans), the aliasing problem is a key challenge that group sum values (GSVs) from the sampled sub-BF signals for
influences the accuracy of rotation speed sensing. aliasing free autocorrelation series calculation (line 3-6). After
The second challenge is that in real practice the range bin the normalization process (line 7-8), the algorithm calculates
which the ERS-radar distance belongs to is a priori unknown, the autocorrelation series of the GSVs (line 9-10) and outputs
which makes it hard to distinguish the signal reflected by the the aliasing free autocorrelation series {Xb }b∈[B] (line 11).
rotating object from the mixed reflection signals to extract IFP.
Chirp Decomposition. As discussed in Sec. IV-B1, the ERS
In the rest of this section, we will introduce the details of
components of the chirps that measure the centrosymmetric
the FE module, which solves the above two challenges using
parts of the rotating object are similar, which incurs the
the AE and ROD sub-module, respectively.
aliasing peaks in the autocorrelation series. Our first step to
2) FE Module Overview: The workflow of the FE module
relieve this problem is to find more ERS components to provide
is illustrated in Alg. 2. The FE module takes the H ×N groups
richer information of the rotation through the CD mechanism.
of raw ADC data {si,j }i∈[H],j∈[N ] as inputs and outputs the
According to Eq. (4), the phase of the ERS component is
3 We omit the quadrature components of F ′ for conciseness because they determined by both the chirp starting frequency fc and ERS-
show similar trends as in-phase components. radar distance. If the radar could simultaneously generate mul-

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 ALGORITHM 3: Aliasing Elimination.
Input: The N groups of raw ADC data {si,j }j∈[N ]
collected under chirp generation frequency 1/Ti ;
Output: The B autocorrelation series {Xb }b∈[B] ;
1 foreach j ∈ [N ] do
// Chirp Decomposition.
2 {ck,j }k∈[K] ← Decompose(si,j ) ;
// Feature Aggregation.
3 foreach k ∈ [K] do
4 Tk,j ← Range-FFT(ck,j );
5 foreach b ∈ [B] do
Gb,j ← K
P
6 k=1 Tk,j [b];
7 foreach b ∈ [B] do
Gb,j ← Gb,j − 1/n N
P
8 h=1 Gb,h ;
Figure 5: The illustration of CD and FA mechanism. The red dots indicate the
// Generate autocorrelation series. samples of the BF signal. Here we decompose a chirp into three sub-chirps
9 foreach b ∈ [B] do and calculate the GSV for the second value in each Range-FFT series.
10 Xb ← Autocorrelation({Gb,j }j∈[N ] );
11 return {Xb }b∈[B] ;

tiple chirps with different starting frequencies, we could get

multiple ERS components whose phases are the descriptions
of the ERS-radar distance under different starting frequencies.
Thus, our CD mechanism decomposes each chirp into K sub-
chirps4 (line 2 in Alg. 3). As illustrated in Fig. 5, a sub- Figure 6: The maximum values in autocorrelation series of the GSVs in each
range bin, and the ERS-radar distance, in our six-blade fan sensing experiment.
chirp is a consecutive portion of the original chirp, whose
corresponding sub-BF signal contains a consecutive portion of their ERS components. This is because the difference of
of the original BF signal. Since the duration of a chirp is only the two GSVs Gr,j1 and Gr,j2 is the cumulative difference
around 0.1ms, we deem the object’s rotation negligible during of {Tk,j1 [r]}k∈[K] and {Tk,j2 [r]}k∈[K] . In Fig. 4a, we plot
this period and all the sub-chirps are simultaneously generated. the Gr,j series extracted from the raw ADC data collected
Actually, the hardware of our mmWave radar itself does in the six-blade fan sensing experiment mentioned in Sec.
not support chirp decomposition. The CD mechanism can be IV-B1. Compared with the series of ERS components shown
equivalently achieved by directly decomposing the original BF in Fig. 3a, the similarity between the GSVs corresponding to
signals into sub-BF signals, which in practice is realized by the rotating object’s different centrosymmetric parts is greatly
dividing the raw ADC data of the original BF signal into reduced. The aliasing peaks in the autocorrelation series of
multiple segments, as illustrated in Fig. 5. Gr,j are also eliminated as shown in Fig. 4b.
Feature Aggregation. After obtaining multiple sub-chirps, Since we have no prior knowledge about the range bin
we propose to aggregate the obtained signal features from the that the ERS-radar distance belongs to, Alg. 3 calculates the
sub-chirps to generate a new signal feature, which distinctly autocorrelation series of GSVs for all the B range bins (line
amplifies the tiny differences between the centrosymmetric 11) and leaves the ERS-radar distance’s range bin discovery
parts of the rotating object. Specifically, after decomposing task to the following ROD sub-module.
the sampled BF signal si,j into K sampled sub-BF signals 4) Rotating Object Discovery: The ROD sub-module takes
{ck,j }k∈[K] (line 2 in Alg. 3), we conduct Range-FFT on each the B autocorrelation series output by the AE sub-module as
ck,j and obtain K corresponding Range-FFT series denoted as inputs, and outputs the extracted IFP. The ROD sub-modue
{Tk,j }k∈[K] (line 4 in Alg. 3), where each Range-FFT series solves the key challenge of discovering the range bin that the
Tk,j is a B × 1 vector with B denoting the total number of the ERS-radar distance belongs to by utilizing the periodically
PK
range bins. The GSV Gb,j = k=1 Tk,j [b] for each b ∈ [B] is varying pattern of the rotating object’s reflected signal. Note
extracted by adding up the all the components with the same that such characteristics also remain in the extracted GSVs.
index5 b in {Tk,j }k∈[K] (line 6 in Alg. 3). Moreover, only the autocorrelation series of a periodic series
Actually, suppose the ERS-radar distance belongs to the r- have distinct large values. In Fig. 6, we plot the maximum
th range bin, then the elements in {Tk,j [r]}k∈[K] are the ERS values in the autocorrelation series of the GSVs in all the B
components of the corresponding K sub-chirps. For different range bins, which are extracted from the data collected in the
centrosymmetric parts of the rotating object, the differences six-blade fan sensing experiment mentioned in Sec. IV-B1. We
of their corresponding GSVs are larger than the differences notice that the maximum value of the autocorrelation series
corresponding to the range bin that the ERS-radar distance
4 A larger K leads to more ERS components with increased computational belongs to is distinctly larger than others.
cost, and we set K empirically in our experiments as given in Sec. V-C2. The detailed workflow of the ROD sub-module is elaborated
5 We denote the b-th value of a B × 1 vector Y as Y [b]. in Alg. 4. After initialization (line 1), the algorithm calculates

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 ALGORITHM 4: Rotation Object Discovery. checks whether T satisfies two groups of conditions (line 5-
Input: The B autocorrelation series {Xb }b∈[B] ; 14). The algorithm calculates the rotation period coefficients8
Output: The IFP value p; {rj′ }j∈[H] (line 5-6) under each chirp generation frequency in
// Initialization. C, and checks the first group of the conditions that whether the
1 I ← empty array, I ← ∅; elements in {pj }j∈[H] are still the IFPs, if the target object’s
2 foreach b ∈ [B] do
3 vb ← maxn∈[N −1] Xb [n];
rotation period is T (line 7-10). The algorithm introduces a
boolean variable u to indicate whether the current T passes the
// Find the rotating object’s range bin.
4 q ← arg maxb∈[B] vb ;
IFP check, and initializes u as True (line 7). If the condition9
// Calculate the IFP value. in line 9 holds, the IFP value under the chirp generation
5 foreach j ∈ [N − 1] do frequency 1/Tj should be pj /GCD(rj′ , pj ) rather than pj
6 if Xq [j] > ηvq then according to Eq. (8), where rj′ is the calculated rotation period
7 Insert j to the end of I; coefficient and pj is the extracted IFP value under the chirp
8 foreach i ∈S[length(I) − 1] do generation frequency 1/Tj . In such case, the current T does
9 I ← I {I[i + 1] − I[i]}; not pass the IFP check, and the algorithm sets u as False (line
10 p ← mode(I); 10). If T passes the IFP check, the algorithm further checks
11 return p; it for the second groups of conditions (line 12-14). According
to Eq. (8), the rotation period T should satisfy that
the maximum value of each autocorrelation series (line 2-
3), and picks the autocorrelation series Xq with the largest pi Ti mod T = 0, ∀i ∈ [H]. (9)
maximum value, whose corresponding range bin is the one If Eq. (9) holds, all the rj ’s calculated in line 6 should be
that the ERS-radar distance belongs to (line 4). After that, the integers and the value of w in line 12 will be 0. However,
algorithm collects in an array I the indices of the periodic due to the signal noise of the raw ADC data in real-world
rotation peaks in Xq (line 5-7). Specifically, an index j is measurements, Eq. (9) may not hold for all the H measure-
selected into I only if the indexed value Xq [j] is close to6 ments. Thus, the algorithm uses the variable d to record the
vq (line 6). The IFP value is calculated as the mode of minimum value of w so far, and updates the rotation speed
the distances between the consecutive indices of the periodic value as s = 1/T , if d is updated (line 13-14). The rotation
rotation peaks (line 8-10). Compared with setting the smallest period T with the smallest w is most likely to satisfy Eq. (9).
element in I as the IFP, such a mechanism improves the Finally, the algorithm outputs the rotation speed s (line 15).
reliability of the results because the IFP can be accurately
obtained as long as the majority of the indices in I are correct.
ALGORITHM 5: Rotation Speed Search.
Finally, the algorithm outputs the IFP p (line 11).
Input: The set of chirp generation frequencies
C = {1/Ti }i∈[H] and IFP values {pi }i∈[H] ;
C. Rotation Speed Search Output: The object’s rotation speed s;
After receiving the IFP values generated by the FE module, // Initialization.
1 s ← 0, d ← H + 1;
the RSS module elaborated in Alg. 5 outputs the object’s rota- // Decide the rotation period search space.
tion speed. At first, the algorithm initializes (line 1) the rotation 2 D ← mini∈[H] pi ;
speed s and the variable d which will be used for checking the // Search the rotation period.
reasonableness of the candidate rotation speeds. According to 3 foreach m ∈ [D] do
Eq. (7), the H IFP values {pi }i∈[H] and their corresponding 4 T ← DTi /m;
// Satisfaction check for IFPs.
chirp generation frequencies {1/Ti }i∈[H] satisfy 5 foreach j ∈ [H] do
pi Ti = mi T, ∀i ∈ [H], (8) 6 rj ← pj Tj /T , rj′ = round(rj );
7 u ← True;
where {mi }i∈[H] are the rotation period coefficients whose 8 foreach j ∈ [H] do
values are unknown positive integers. In other words, each 9 if GCD(rj′ , pj ) > 1 then
pi Ti must be an integral multiple of the rotation period 10 u ← False;
T . Moreover, since the chirp generation frequency of our // Satisfaction check for Eq. (9).
mmWave radar is higher than the target object’s rotation 11 if u = True then
speed7 1/T , the value of T must be in the set {pi Ti /m}m∈[pi ] . w← H ′
P
12 j=1 |rj − rj |;
Thus, the algorithm then finds the minimum search space for 13 if d < w then
the least computational cost (line 2). 14 d ← w, s ← 1/T ;
After that, for each calculated candidate rotation period T
15 return s;
in the search space {DTi /m}m∈[D] (line 4), the algorithm

6 Note
that η < 1 and η is close to 1. 8 The
round(x) in line 6 of Alg.5 is to obtain the nearest integer of x.
7 The
chirp generation frequency of our mmWave radar is as high as 1e3Hz, 9 The
GCD(·, ·) in line 9 is to calculate the greatest common divisor between
which is much higher than the rotation speed of common objects. two values.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 Figure 7: Servo motor. Figure 8: Six-blade fan on servo motor. Figure 9: Tiny motor spindle. Figure 10: Laser tachometer (left) and
electromagnetic tachometer (right).

Figure 11: CNC machining center. Figure 12: Two turning lathes.

V. E VALUATION The ground truth of the servo motor and the fan is obtained
A. Implementation and Experimental Setup by the accurate build-in motor electromagnetic tachometer
shown in Fig. 10. For the tiny motor spindle, we use the
1) Implementation: We implement mRotate with a com-
measurement result of the laser tachometer as the ground truth,
mercial mmWave radar TI IWR1443 BoosterPack [14]. The
since it shows good performance on high-speed sensing, as
raw ADC data is collected by a TI DCA1000EVM data
we will discuss in Sec. V-B. In field tests, the ground truth
acquisition board [15]. The data processing codes are written
is obtained by designating (for the CNC machining center) or
in Matlab and run on a computer with Intel Core i7-1160G7
reading from the nameplates (for the turning lathes).
processor and 16GB memory.
2) Experiment Setting: We conduct extensive experiments B. Rotation Speed Sensing Accuracy
to evaluate the performance of mRotate. The experiment
In setting 1, we set 25 levels of the servo motor rotation
settings in lab environments include the following setting
speeds evenly spanning from 100rpm to 2300rpm, and set the
1 with a servo motor whose rotation speed can be flexibly
motor spindle-radar distance as 50cm. The absolute percentage
designated shown in Fig. 7 to evaluate the sensing accuracy
errors10 (APE) of mRotate and laser tachometer under different
and sensing distance, setting 2 with a six-blade fan attached
rotation speeds are shown in Fig. 13. mRotate achieves a
on the servo motor shown in Fig. 8 to evaluate the centrosym-
lower mean absolute percentage error (MAPE) (0.24%) than
metric object sensing performance, and setting 3 with a tiny
the laser tachometer (0.39%). The laser tachometer obtains
motor spindle shown in Fig. 9 to evaluate the tiny object
high errors when the rotation speed is low, because in such
sensing performance. The experiment settings for field tests
case one bit error of the laser pulse counts will cause high
in a machining workshop include setting 4 on a computerized
APE. However, mRotate obtains low APE in such case thanks
numerical control (CNC) machining center shown in Fig. 11
to that the CFH mechanism provides multiple measurements
to validate that mRotate can accurately sense the rotation
and the RSS algorithm robustly searches the accurate rotation
speeds of a large mechanical device, and setting 5 on the main
speed by utilizing all the measurement results.
spindles of two turning lathes shown in Fig. 12 to validate that
mRotate can measure the rotation speeds of multiple devices. C. Impact of Practical Factors
3) Baseline and Ground Truth: The baseline which we
1) Measurement Distance: The measurement distance is of
compare mRotate with is a commercial laser tachometer UNI-
great importance for the safety of the operators. In setting
T UT372 [13] shown in Fig. 10. Note that if we operate the
1, we evaluate the impact of the measurement distance of
laser tachometer to sense the rotation speed of the servo motor
mRotate. We first place the mmWave radar 0.5m, 1m, 1.5m,
by sticking its reflective tag on the spindle as recommended
2m, and 2.5m away from the spindle of the servo motor. For
in its manual, the laser tachometer will fail, because the metal
each spindle-radar distance, we set the rotation speed as three
spindle provides strong reflection which overwhelms the laser
pulses reflected by the tag. Thus, we have to cover the spindle 10 Suppose v is the sensing value and v is the ground truth value, the
s g
with black paper and then stick the reflective tag on it. absolute percentage error is |vg − vs |/vg .

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 Table I: Comparison of mRotate with or without AE in setting 2. The unit of
the numbers in the table is rpm.
Ground Truth 250 493 733 968 1202 1428 1640 1872
with AE 250 494 732 968 1200 1429 1644 1875
without AE 1550 2927 4127 5714 7059 8577 9864 10909

Table II: Ground truths and the measurement results of mRotate for field tests.
The unit of the numbers in the table is rpm.
CNC Turning Turning
Devices
Machining Center Lathe 1 Lathe 2
Figure 13: The APEs of mRotate and the laser tachometer under different
rotation speeds for experiments in setting 1. Ground Truth 1000 2000 3000 4000 5000 560 1120
mRotate 1001 2000 3003 4000 5000 563 1111

3) Target Object Size: To investigate mRotate’s ability to

measure the rotation speed of tiny objects, in setting 3, we
evaluate mRotate on a tiny motor spindle. As illustrated in Fig.
9, the width of the spindle is only 5mm. We set the spindle-
radar distance as 10cm. The rotation speed measurement result
of mRotate is 12500rpm, and the APE is only 0.34%. This
experiment validates that mRotate can accurately measure the
Figure 14: The APEs of mRotate un- Figure 15: The PARs with different rotation speed of an object with (i) a small size (several
der different measurement distances sub-chirp numbers and rotation speeds
and rotation speeds in setting 1. for experiments in setting 2.
millimeters) and (ii) a rotation speed as high as 12500rpm.
speed levels 968rpm, 1428rpm, and 1872rpm which distribute D. Field Tests
evenly from around 1000rpm to 2000rpm. As shown in Fig.
1) Measurement for Multiple Speed Levels: In setting 4,
14, mRotate achieves an APE lower than 0.21% for all the
we evaluate mRotate on a high-end CNC machining center,
measurements, and the APE is stable against the increasing
as shown in Fig. 11. We set 5 levels of rotation speeds which
distances. This experiment validates that mRotate is able to
distribute evenly from 1000rpm to 5000rpm. Tab. II shows
accurately measure the rotation speed from as far as 2.5m.
that mRotate achieves high rotation speed sensing accuracy
2) Centrosymmetric Rotating Object Measurement: In set- (APE ≤ 0.1%) in all the 5 speed levels. The distance between
ting 2, we evaluate mRotate to sense a centrosymmetric the machine drill and the protection baffle is 0.75m, which
rotating object. As shown in Fig. 8, we attach a six-blade fan is far beyond the laser tachometer’s recommended operation
on the servo motor, and set eight levels of the rotation speeds distance (< 20cm). However, mRotate can accurately sense
which distribute evenly from around 250rpm to 2000rpm. The the machine drill from a distance of as far as around 1m.
fan-radar distance is 16cm. We compare the results of mRotate To further illustrate the effectiveness of the ROD mechanism
with or without the AE mechanism in Tab. I. Without the AE for discovering the rotating object’s range bin in complex
mechanism, mRotate is badly affected by the aliasing problem real-world environments, we plot the ERS-radar distance, the
and its sensing results are four to six times higher than the maximum values of autocorrelation series of GSVs in each
ground truth. In contrast, with the AE mechanism, mRotate range bin, and the Range-FFT series in Fig. 16 and Fig.
achieves an APE lower than 0.24% for all the measurements. 17. The traditional object detection approach which searches
To further investigate the influence of the number of sub- for the peaks in the Range-FFT series cannot distinguish the
chirps on the functionality of the AE mechanism, we plot rotating object. However, the ROD mechanism can accurately
in Fig. 15 the PARs11 of the autocorrelation series, which discover the ERS-radar distance by utilizing the distinctly
are generated from the GSVs series aggregated from different large maximum value of the rotating object’s corresponding
number of sub-chirps. We show the results under four rotation autocorrelation series of the GSVs.
speed levels (i.e., 494rpm, 968rpm, 1428rpm, 1872rpm) for 2) Multi-object Measurement: Thanks to the wide field of
conciseness because the results under the other four speed the view (FoV) of our mmWave radar12 , mRotate can simul-
levels in Tab. I show similar trends. We notice that without taneously measure the rotation speeds of multiple objects. In
the AE mechanism (i.e., the number of sub-chirps is 1), the setting 5, we evaluate the performance of mRotate for sensing
PAR is very close to 1. As we use more sub-chirps, the PAR the rotating speeds of the main spindles of two turning lathes
is becoming larger and the differences between the periodic as illustrated in Fig. 12.
rotation peaks and aliasing peaks are amplified. In all the We first introduce how mRotate simultaneously senses the
experiments, we set the number of sub-chirps as 30, because rotation speeds of multiple objects. We plot in Fig. 18 and
such a configuration is both computationally efficient and easy Fig. 19 the ERS-radar distances, the maximum values of
to distinguish the periodic rotation peaks and aliasing peaks. autocorrelation series of GSVs in each range bin, and the
Range-FFT series in setting 5. As shown in Fig. 19, it is
11 PAR is the ratio of the periodic rotation peaks’ mean amplitude and the
aliasing peaks’ mean amplitude. 12 110◦ azimuth FoV and 30◦ elevation FoV.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 Table III: Comparison with other rotation speed sensing methods.
Safe Sensing Object Illumination Require
Method
Range Independent Independent Training
Mechanical [4] no yes yes no
Electromagnetic [5, 6] no no yes no
Camera [7–12] yes yes no no
Figure 16: The maximum values of Figure 17: The Range-FFT series of Laser [13] no no yes no
autocorrelation series of GSVs in each sampled sub-BF signal for experi- SpiralSpy [16] yes no yes yes
range bin for experiments in setting 4. ments in setting 4. RFTacho [17] no yes yes yes
mRotate yes yes yes no

from being broadly adopted. Besides, all the above camera-

based approaches fail to operate under poor illuminations. The
laser tachometers [13] are robust to illumination conditions.
However, the laser tachometer fail if the target object itself
Figure 18: The maximum values of Figure 19: The Range-FFT series of is highly reflective. In such case, the laser tachometer cannot
autocorrelation series of GSVs in each sampled sub-BF signal for experi- distinguish the laser pulse reflected by its reflective tag from
range bin for experiments in setting 5. ments in setting 5. those reflected by the object. Besides, the operation distance
difficult to distinguish the rotating objects by searching for of the laser tachometer (around 20cm [13]) is also inadequate
the peaks in Range-FFT series, while in Fig. 18 the maximum for many applications that require a longer safe distance.
value of the autocorrelation series corresponding to the range Compared with the above contactless approaches, mRotate
bins that the two rotating objects’ ERS-radar distances belong employs the RF signal for rotation speed sensing from a
to are distinctly larger than the others. To enable multiple reasonable distance, and is robust to the illumination condi-
objects’ rotation speed sensing, we thus slightly change the tions and the target object’s light reflectivity. Note that [16]
ROD sub-module introduced in Sec. IV-B4 to pick out the also utilizes mmWave radar for fan rotation speed sensing.
multiple autocorrelation series by setting a threshold on their However, the method proposed in [16] requires training, and
maximum values. After that, we extract the IFP values and does not support sensing other objects and the rotation speed
search the rotation speeds of the multiple rotating objects. out of the labels in training dataset. The method proposed in
As shown in Tab. II, mRotate can simultaneously sense the [17] has a maximum sensing distance of around 65cm, which
rotation speed of the two main spindles in setting 5 with a is inadequate for safely measuring the rotation speed of many
high accuracy (APE ≤ 0.8% for both the two main spindles). types of industrial machines. We comprehensively compare
mRotate with the above methods in Tab. III.
VI. R ELATED W ORK mmWave-based sensing. Apart from rotation speed sens-
Rotation speed sensing. Existing works on rotation speed ing, mmWave-based sensing has shown a rising popularity
sensing can be categorized as contact-based and contactless among a lot of other applications, such as micro-displacement
approaches. The contact-based sensors [4] are physically at- sensing [18, 19], vital signs recovery [20–22], posture and
tached to the shaft of the target object. Such a measurement gesture sensing [23–28], eavesdropping [29], physical layer
mode constrains their operation distances, which is likely to attacking [30], and many others [31–41]. Our work distin-
cause safety risks to the operator. The contactless approaches guishes from the above works, as we leverage the mmWave
relieve this problem and have become promising paradigms radar for a different task of rotation speed sensing.
for rotation speed sensing. The contactless approaches can be
categorized as electromagnetic-based [5, 6], optical-based [7– VII. C ONCLUSION
13], and RF-based ones [16, 17]. In this paper, we propose mRotate, a mmWave radar-based
The electromagnetic-based methods utilize electrostatic [5] rotation speed sensing system. Specifically, we design a series
or hall-effect [6] sensors to detect the change of the elec- of novel approaches that enable mRotate to extract the signal
tromagnetic field caused by the object’s rotation. Although reflected by the rotating object from the mixed reflection
the electromagnetic-based methods enable contactless mea- signals, generate the aliasing free signal features to enable
surement, their operation distances are less than 10cm. accurate rotation speed sensing for centrosymmetric objects,
The optical-based approaches utilize a variety of optical and accurately pinpoint the rotation speed through customized
sensors for rotation speed sensing. Among them, [7, 8] utilize radar sensing mode and algorithm design. We extensively
RGB cameras to capture the images of the rotating object and evaluate mRotate in both lab environments and field tests.
estimate the object’s rotation speed by analysing the extracted mRotate achieves an MAPE of only 0.24% for accuracy test,
features of the images. However, these approaches can only which is 38% lower than that produced by the baseline device,
measure rotation speeds less than 1500rpm, constrained by a popular commercial laser tachometer. Meanwhile, mRotate
the relatively low frame rate (30 fps) of the RGB cameras. can measure the rotation speed of a small spindle whose
Other works also propose to use high-speed cameras [9– diameter is only 5mm, achieve low errors from a distance as
12] to enlarge the measurement range of the rotation speed. far as 2.5m, and simultaneously obtain the accurate rotation
However, the prohibitive prices of these cameras prevent them speeds of multiple objects in field tests.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.
 R EFERENCES [24] H. Xue, Y. Ju, C. Miao, Y. Wang, S. Wang, A. Zhang, and L. Su,
“mmmesh: towards 3d real-time dynamic human mesh construction
[1] Uwoodcraft, “Turning lathe speeds,” https://uwoodcraft.com/woodturni using millimeter-wave,” in The 19th Annual International Conference
ng-lathe-speeds-a-complete-guide/. on Mobile Systems (MobiSys), 2021.
[2] Haynes, “Vehicle inspections,” https://haynes.com/en-us/tips-tutorials/w [25] S. Palipana, D. Salami, L. A. Leiva, and S. Sigg, “Pantomime: Mid-
hat-know-about-vehicle-inspections-all-50-states. air gesture recognition with sparse millimeter-wave radar point clouds,”
[3] “Rpm in washing machine,” https://theportablelaundry.com/washing-m Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiqui-
achine-rpm/. tous Technologies, vol. 5, no. 1, pp. 27:1–27:27, 2021.
[4] REED Instruments, “Photo-contact tachometer instruction man- [26] C. Shi, L. Lu, J. Liu, Y. Wang, Y. Chen, and J. Yu, “mpose: Environment-
ual,” https://www.reedinstruments.com/pdfs/cache/www.reedinstrumen and subject-agnostic 3d skeleton posture reconstruction leveraging a
ts.com/r7100/manual/r7100-manual.pdf. single mmwave device,” Smart Health, vol. 23, p. 100228, 2022.
[5] L. Li, H. Hu, Y. Qin, and K. Tang, “Digital approach to rotational speed [27] A. Adhikari, S. Avula, and S. Sur, “Mmsleep: Monitoring sleep posture
measurement using an electrostatic sensor,” Sensors, vol. 19, no. 11, from commodity millimeter-wave devices,” in The 20th Annual Interna-
2019. tional Conference on Mobile Systems (MobiSys), 2022.
[6] K. J, B. Varghese, U. A. Ponnary, F. Kamar, and R. T. Jacob, “Hall effect [28] E. M. Sitar, M. S. Saadat, and S. Sur, “A millimeter-wave wireless
sensor based portable tachometer for rpm measurement,” International sensing approach for at-home exercise recognition,” in The 20th Annual
Journal of Computer Science and Engineering Communications, vol. 2, International Conference on Mobile Systems (MobiSys), 2022.
no. 1, pp. 100–105, 2014. [29] P. Hu, Y. Ma, P. S. Santhalingam, P. H. Pathak, and X. Cheng, “MIL-
[7] Y. Wang, L. Wang, and Y. Yan, “Rotational speed measurement through LIEAR: millimeter-wave acoustic eavesdropping with unconstrained
digital imaging and image processing,” in 2017 IEEE International vocabulary,” in The 41st International IEEE Conference on Computer
Instrumentation and Measurement Technology Conference (I2MTC), Communications (INFOCOM), 2022.
2017. [30] Z. Sun, S. Balakrishnan, L. Su, A. Bhuyan, P. Wang, and C. Qiao,
[8] T. Wang, Y. Yan, L. Wang, and Y. Hu, “Rotational speed measurement “Who is in control? practical physical layer attack and defense for
through image similarity evaluation and spectral analysis,” IEEE Access, mmwave-based sensing in autonomous vehicles,” IEEE Transactions on
vol. 6, pp. 46 718–46 730, 2018. Information Forensics and Security, vol. 16, pp. 3199–3214, 2021.
[9] H. Kim, Y. Yamakawa, T. Senoo, and M. Ishikawa, “Visual encoder: [31] C. X. Lu, S. Rosa, P. Zhao, B. Wang, C. Chen, J. A. Stankovic,
robust and precise measurement method of rotation angle via high-speed N. Trigoni, and A. Markham, “See through smoke: Robust indoor map-
rgb vision,” Optics Express, vol. 24, no. 12, pp. 13 375–86, 2016. ping with low-cost mmwave radar,” in The 18th Annual International
[10] F. Natili, F. Castellani, D. Astolfi, and M. Becchetti, “Video-tachometer Conference on Mobile Systems (MobiSys), 2020.
methodology for wind turbine rotor speed measurement,” Sensors, [32] C. Xu, Z. Li, H. Zhang, A. S. Rathore, H. Li, C. Song, K. Wang, and
vol. 20, no. 24, 2020. W. Xu, “Waveear: Exploring a mmwave-based noise-resistant speech
[11] X. Zhu and S. Yu, “Measurement angular velocity based on video sensing for voice-user interface,” in The 17th Annual International
technology,” in The 4th International Congress on Image and Signal Conference on Mobile Systems (MobiSys), 2019.
Processing (ICSIP), 2011. [33] C. X. Lu, M. R. U. Saputra, P. Zhao, Y. Almalioglu, P. P. B. de Gusmao,
[12] J. Zhong, S. Zhong, Q. Zhang, and Z. Peng, “Measurement of instanta- C. Chen, K. Sun, N. Trigoni, and A. Markham, “Milliego: Single-chip
neous rotational speed using double-sine-varying-density fringe pattern,” mmwave radar aided egomotion estimation via deep sensor fusion,” in
Mechanical Systems and Signal Processing, vol. 103, pp. 117–130, 2018. The 18th Conference on Embedded Networked Sensor Systems (SenSys),
[13] U.-T. Technology, “Laser tachometer operating manual,” https://meters 2020.
.uni-trend.com/product/ut370-series/#Docs/. [34] J. Pegoraro, J. O. Lacruz, M. Rossi, and J. Widmer, “SPARCS: A sparse
[14] Texas Instruments Incorporated, “Single-chip 76-ghz to 81-ghz mmwave recovery approach for integrated communication and human sensing in
sensor integrating dsp and mcu,” https://www.ti.com/product/IWR1642/. mmwave systems,” in The 21st International Conference on Information
[15] Texas Instruments Incorporated, “Real-time data-capture adapter for Processing in Sensor Networks (IPSN), 2022.
radar sensing evaluation module,” http://www.ti.com/tool/DCA1000E [35] A. B. Pizarro, P. J. Mateo, F. Gringoli, and J. Widmer, “Augmenting
VM/. mmwave localization accuracy through sub-6 ghz on off-the-shelf de-
[16] Z. Li, B. Chen, X. Chen, H. Li, C. Xu, F. Lin, C. X. Lu, K. Ren, and vices,” in The 20th Annual International Conference on Mobile Systems
W. Xu, “Spiralspy: Exploring a stealthy and practical covert channel (MobiSys), 2022.
to attack air-gapped computing devices via mmwave sensing,” in The [36] S. Ding, Z. Chen, T. Zheng, and J. Luo, “Rf-net: a unified meta-learning
Network and Distributed System Security (NDSS), 2022. framework for rf-enabled one-shot human activity recognition,” in The
[17] M. Heggo, L. Bhatia, and J. A. McCann, “Rftacho: Non-intrusive RF 18th Conference on Embedded Networked Sensor Systems (SenSys),
monitoring of rotating machines,” in The 21st International Conference 2020.
on Information Processing in Sensor Networks (IPSN), 2022. [37] T. Zheng, Z. Chen, J. Luo, L. Ke, C. Zhao, and Y. Yang, “Siwa: see
[18] J. Guo, M. Jin, Y. He, W. Wang, and Y. Liu, “Dancing waltz with ghosts: into walls via deep UWB radar,” in The 27th Annual International
Measuring sub-mm-level 2d rotor orbit with a single mmwave radar,” in Conference on Mobile Computing and Networking (MobiCom), 2021.
The 20th International Conference on Information Processing in Sensor [38] R. Tavasoli, S. Sur, and S. Nelakuditi, “Sscense: A millimeter-wave
Networks (IPSN), 2021. sensing approach for estimating soluble sugar content of fruits,” in The
[19] C. Jiang, J. Guo, Y. He, M. Jin, S. Li, and Y. Liu, “mmvib: micrometer- 20th Annual International Conference on Mobile Systems (MobiSys),
level vibration measurement with mmwave radar,” in The 26th Annual 2022.
International Conference on Mobile Computing and Networking (Mobi- [39] J. M. Schellberg and S. Sur, “Accurate device self-tracking for robust
Com), 2020. millimeter-wave imaging on handheld smart devices,” in The 20th
[20] S. Zhang, T. Zheng, Z. Chen, and J. Luo, “Can we obtain fine-grained Annual International Conference on Mobile Systems (MobiSys), 2022.
heartbeat waveform via contact-free rf-sensing?” in The 41st Interna- [40] T. Liu, M. Gao, F. Lin, C. Wang, Z. Ba, J. Han, W. Xu, and K. Ren,
tional IEEE Conference on Computer Communications (INFOCOM), “Wavoice: A noise-resistant multi-modal speech recognition system fus-
2022. ing mmwave and audio signals,” in The 19th Conference on Embedded
[21] Z. Chen, T. Zheng, C. Cai, and J. Luo, “Movi-fi: motion-robust vital Networked Sensor Systems (SenSys), 2021.
signs waveform recovery via deep interpreted RF sensing,” in The 27th [41] E. Soltanaghaei, A. Prabhakara, A. Balanuta, M. Anderson, J. M.
Annual International Conference on Mobile Computing and Networking Rabaey, S. Kumar, and A. Rowe, “Millimetro: Mmwave retro-reflective
(MobiCom), 2021. tags for accurate, long range localization,” in The 27th Annual Interna-
[22] X. Xu, J. Yu, C. Ma, Y. Ren, H. Liu, Y. Zhu, Y.-C. Chen, and F. Tang, tional Conference on Mobile Computing and Networking (MobiCom),
“mmECG: monitoring human cardiac cycle in driving environments 2021.
leveraging millimeter wave,” in The 41st International IEEE Conference
on Computer Communications (INFOCOM), 2022.
[23] H. Kong, X. Xu, J. Yu, Q. Chen, C. Ma, Y. Chen, Y. Chen, and L. Kong,
“m3 track: mmwave-based multi-user 3d posture tracking,” in The 20th
Annual International Conference on Mobile Systems (MobiSys), 2022.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on October 23,2023 at 10:34:47 UTC from IEEE Xplore. Restrictions apply.