EVLeSen: In-Vehicle Sensing with EV-Leaked Signal

While out-vehicle sensing has achieved great success with the development of vehicle radar and Lidar systems, invehicle sensing attracts a lot of attention recently. However, the popular camera-based solutions raise privacy concerns and pose requirement on lighting conditions. Researchers recently utilize wireless signals for sensing. However, besides requiring dedicated hardware, the rich multipath in a small cabin space causes severe interference, degrading the sensing reliability. In this paper, we propose a new sensing modality for in-vehicle sensing, leveraging the leaked EM signals from electric vehicles. The key observation is that the human body can capture the leaked signals, and body motions affect the signal variation patterns. Our solution involves designing conductive cloth tags on the seat to effectively collect body-captured signals and adopting a reference tag to deal with interference. Through extensive experiments conducted over 100 hours, covering a driving distance of 4000 kilometers on various real roads, our system, EVLeSen, can achieve over 90% accuracy in recognizing body motions utilizing just the leaked ambient signals.


INTRODUCTION
Motor vehicles have become an indispensable component of our modern society.While out-vehicle sensing has become mature in the last few years with the rapid development of vehicle radar [53], Lidar [38] and camera [40], a lot of attention is now paid to in-vehicle sensing [62].Camera-based solutions [3] are popular for in-vehicle sensing.However, these solutions still face issues, including privacy concerns [47] and performance degradation under poor lighting conditions such as at night [36].Researchers also explore the latest wireless sensing technologies such as WiFi sensing [45] and acoustic sensing [31] for in-vehicle sensing.However, dedicated hardware deployment [68], extremely rich multipath due to the small cabin space [55], and the inherent inability to sense multiple targets [59], still pose challenges for these solutions to be adopted.
In recent years, we observe an increasing trend of promoting Electric Vehicle (EV) internationally due to its emissionfree nature [46].Governments worldwide have developed different incentive programs to encourage or even force the use of electric vehicles [56].For example, Norway set a national goal that all new vehicle sales after 2025 should be zero-emission vehicles [4].In Asian countries, electric vehicles comprise approximately 30% of new sales [25], and we envision more electric vehicles on the road in the future.
Along with this trend, we observe an interesting phenomenon: a lot of components inside electric vehicles such as batteries, powerlines, and power inverters [1] leak electromagnetic (EM) signals during their operations [21,51].While leakage is usually considered as bad, we ask the question: Is it viable to leverage the leaked signals from electric vehicles for in-vehicle sensing?If these leaked signals can be utilized for sensing, we do not need to deploy any dedicated transmitters to emit signals for sensing.Take popular WiFi sensing as the example-we need to deploy a WiFi transmitter in the vehicle cabin and transmit dedicated packets at a rate of 100-1000 packets per second for sensing [60].
However, we quickly realize utilizing the EV-leaked signal for sensing is a challenging task.This is because the leaked signal is very different from conventional RF signals (e.g., WiFi and UWB) utilized for sensing.Compared to those signals in the frequency band of megahertz or gigahertz, the frequency of the leaked signal (i.e., 60 Hz and 120 Hz) is much lower.Human motions can induce significant variations on WiFi signals but not on EV-leaked signals.For example, a displacement of 1 cm can induce a phase change of 120 degrees on WiFi signals but only an extremely small change of 7.2 × 10 −6 degrees on EV-leaked signals.Therefore, we can not utilize the effect of human motions on signal propagation for sensing and existing RF sensing modality does not apply.
In this work, we propose a new sensing modality based on the this low-frequency signal leaked from electric vehicles for in-vehicle sensing.Instead of capturing the effect of human motion on signal propagation, we involve human target in the transceiver loop. 1 Specifically, we consider the human target as part of the receiver.When the human target moves, the characteristic of the receiver varies, causing clear signal variations which can be utilized for sensing.Another observation is that when we link the human body to any ports such as a USB port in vehicle cabin using a simple tag (e.g., a small piece of conductive textile), we observe an even clearer signal variation.We believe this is because such connection enhances the coupling between the human body and electronic components inside the electric vehicle.Also we notice that the tag-body connection is very flexible: the textile serving as the tag can be attached to any parts of the body and the size of the textile tag can be very small.For example, we can deploy a small piece of conductive textile on the vehicle seat.Our finalized design, illustrated in Figure 1, adopts a small piece of adhesive conductive cloth fabric as the tag.
After the signal variations are captured, the next question is how can we infer target motions from the signal variations?To infer human motions (e.g., hand gestures) from the leaked signals, we need to model the relationship between target motions and corresponding signal variations.After a thorough investigation, we realize that in this new sensing modality, target motions affect the leaked signals through two distinct models.The first model pertains to the human body capacitance [5].Here, the human body functions as a conductor, creating an electric capacitor.Body movement changes the capacitance, inducing signal variations in the low-frequency range, i.e., DC to several Hertz.The second sensing model is related to human body's inherent bio-characteristic.Specifically, human postures with larger physical surface areas capture more leaked signals [12].When a person performs distinct motions, the captured leaked signal varies accordingly.These changes manifest in the relatively higher frequency range (approximately tens of Hertz).Leveraging these two 1 All the experiments conducted are IRB-approved by the host institute.

Without target
With target Sta�c Target sensing models, we can extract body motion information from the induced signal variations.However, signal variations are not solely affected by human motions.The vehicle's operational status also varies the signal.For example, vehicle acceleration can induce a larger electric current in the electric motor, leading to a variation in the leaked signal.This variation is mixed with motioninduced signal variation, making target sensing challenging.So the next question is how to deal with the real-world interference to achieve robust sensing performance?The straightforward solution is to obtain the effect of each acceleration and remove the interference by accurately measuring the acceleration.However, this method does not work due to the ultra-dynamic nature of acceleration in complicated reallife situations.To address this issue, we introduce an extra reference tag.This tag is attached close to the existing tag but does not make contact with the target body.Therefore, this tag captures the signal variation induced by all other factors (e.g., vehicle acceleration) except target motion.In contrast, the tag in contact with the target captures signal variation caused by both target motion and other factors.We can remove the effect of other factors to recognize the target motions based on the collected signals from these two tags.
Utilizing the above principles, we propose EVLeSen, the first in-vehicle sensing system exploiting the leaked signals of the electric vehicles.It offers the following advantages: • The proposed system does not require any dedicated transmitter to emit dedicated signals, but utilizes small tags to collect ambient leakage from electric vehicles for sensing.• Different from conventional wireless sensing which relies on the signal variation caused by influencing signal propagation, the proposed sensing modality relies on the signal variation caused by influencing the transceivers directly.
Owing to the completely different sensing principle, the proposed system can simultaneously sense multiple targets, which is a known challenge for conventional wireless sensing.Also, dynamic interference such as motions from a nearby person, which is a big issue for conventional wireless sensing, has little effect on our system.• As we adopt a textile-based tag, the tag can be placed on the seat without causing discomfort.Although the contact point is just between the tag and butt, the proposed system is able to sense motions of the whole body (e.g., arm/head/foot motions).Note that the contact point can be any body part and we place it on the seat mainly for comfortableness and convenience.
We conduct experiments under various real-world conditions to evaluate the system performance.We test EVLeSen over 100 hours with a driving distance of 4000 kilometers under a variety of road conditions.For motions including three hand gestures, three leg motions, and three head motions, EVLeSen can achieve a recognition accuracy of 95% when the vehicle is on the highway.When the vehicle is on downtown roads with a lot of acceleration and deceleration, we propose signal processing method to mitigate the effect.We show that interfering factors such as turning on/off air condition/Bluetooth in the vehicle do not affect the performance.We summarize our major contributions below.
• We propose a novel idea to utilize the leaked signals from electric vehicles for in-vehicle sensing for the first time.
We envision this new sensing modality can trigger followup research to utilize ambient signals in vehicles instead of dedicated signals for sensing and attract more attention to in-vehicle sensing.• Through benchmark experiments and analysis, we model the relationship between the induced signal variation and human motions for the proposed new sensing modality.
We believe the obtained models can help researchers understand the underlying principle, laying the theoretical foundation for the proposed sensing modality.• With the reference tag design, we effectively remove the effect of diverse interference, making the proposed sensing modality practical in real-world settings.• We conduct comprehensive field studies on real roads with three different electric vehicles to evaluate the performance of the proposed system under diverse conditions.
The rest of the paper is organized as follows.We introduce the EM leakage from the electric vehicles and human body's bio-electric features in §2 and elaborate the design of EVLeSen in §3.Then we detail the solution for sensing under varying speeds in §4.The implementation details and evaluations are presented in §5 and §6, respectively, followed by discussions ( §7), related works ( §8) and conclusions ( §9).

BACKGROUND
Electromagnetic leakage in electric vehicles.To meet the health and safety regulations, a lot of measures such as shielding, system grounding, and cable routing have been applied to electric vehicles to reduce the amount of electromagnetic leakage [23,29].Nevertheless, the operation of electric vehicles still results in unavoidable emission of electromagnetic signals from their electronic components, including electric motors, powerlines, and power inverters [1], as dictated by Maxwell's Equations [57].Typically, these leaked signals occupy the frequency range from several hertz to hundreds of hertz, with predominant components in the band of tens of hertz [17].Several studies have substantiated the presence of such electromagnetic leakage in contemporary commercial electric vehicles [21,51].Fortunately, these signals have been verified to be innocuous for human.This research aims to present a fresh perspective on the unfavorable electromagnetic leakage and utilize these ambient leakages for in-vehicle sensing.
Body-captured electromagnetic signals.The human body comprises approximately 60% water, encompassing intracellular fluid, interstitial fluid and plasma.These fluids contain various ions including chlorine, potassium and sodium [33], which contribute to the electrical conductivity of human body [44].This fact forms the foundation of bioelectrical impedance analysis.Specifically, by passing smallamplitude electric current through the body, commercial smart body composition scales assess body composition percentages, including muscle, fat, water, and bone, by analyzing their distinct electrical conductivities [27].Despite its rarity in daily contexts, the human body can be considered as an antenna capable of capturing electromagnetic signals, especially low-frequency ones, owing to the body's relatively large size compared with commercial antennas [24,34].This body characteristic enables us to perform in-vehicle sensing by leveraging the human body as the antenna to capture EV-leaked electromagnetic signals and then analyze the variation of the body-captured signals caused by body movements.

EVLESEN DESIGN
We first present the rationale behind utilizing human body to capture the EV-leaked signals in our EVLeSen system and then introduce its hardware and software designs.

Capturing the Leakage Signal
Conventional wireless sensing systems use RF signals, such as WiFi [63] and RFID [70] for sensing.These signals are in the high-frequency range with short signal wavelengths, making them susceptible to motions in the environment (e.g., hand gestures) during the process of propagation in the air.In contrast, the leaked signals from EV have extremely low frequency and super long wavelengths (on the order of millions of meters).These long wavelengths make the signal propagation not influenced by target motions.
To sense body motions from EV-leaked signals, we first attach a conductive tag to the target to extract body-captured signals, as shown in Figure 2(a).Specifically, a small piece (6  cm × 2 cm) of conductive material serves as the tag.Note that we use a copper tag, which is not our final design, to illustrate the concept and for our preliminary studies.Besides, the tag can be linked to the target in other ways, i.e., placed on the seat and the target sits on the seat, which will be detailed in Section 3.2.The benchmark experiments are conducted in a 2023 Tesla Model 3 [52].During the measurement process, the vehicle is driven on a countryside road with speeds between 35 and 45 mph.
To quantify the energy of the signals captured at the tag, we calculate the energy within a fixed window of time denoted as  using the following equation where  denotes the energy of the signals within the time window,   is the value (voltage) of the th sample within the interval, and   is the sampling frequency.Typically, when calculating signal energy, the resistance (denoted as ) should be considered.However, we follow the recent work [30] by using an open-circuit design at the tag to enhance the received voltage, so the value of  is essentially infinite.Thus, we omit it from consideration in Equation 1.The measured energy value is further scaled in the range of [0, 1], where 0 represents the zero value, and 1 denotes the maximum value.
The signal energy captured from the tag in contact with the human body is highlighted in blue in Figure 2(b).We can observe energy fluctuations over time due to vehicle operation condition changes, which is further elucidated in Section 4. Besides, we observe that the leaked signals captured by the human body are still very weak, making it challenging to utilize them for effective sensing.This is because a lot of measures have been adopted in electric vehicle design to reduce the amount of leakage to meet the safety and health regulations [1].While beneficial for health, they present challenges for the proposed in-vehicle sensing.
Enhancing the captured signal.To increase the amplitude of body-captured signals from the EV, we observe that establishing a "connection" between the tag in contact with human body and the electric vehicle significantly enhances the energy of captured signals.This connection does not require connecting the tag to any inner electronic components of    the vehicle.Instead, we just need to connect the tag to a standard port, such as USB port that is widely available in the car cabin.Although this setup does not directly connect the tag to the EV's electronic components, electromagnetic signals can propagate through the tag to leak stronger electromagnetic fields around the target human body.Thus, such a "connection" increases the amplitude of the body-captured signals by strengthening the coupling between the human body and the vehicle.With this simple connection, the energy of the signals captured by the tag, depicted in red in Figure 2(b), significantly increases.Note that this connection in practice can be made unnoticeable without causing any discomforts.

Hardware Design
A limitation of existing system design [6,30] is its reliance on target wearing the tag to ensure the tag is in direct contact with the skin.Researchers have demonstrated that such a wearable design is inconvenient and uncomfortable for targets [15,48].In response, we propose our approach, i.e., deploying the tag on the vehicle seat, eliminating the need for direct skin contact.When a target sits in the seat, the tag on the seat makes contact with the target with cloth in between, still capable of enabling strong coupling between target and the electronic components.
We first evaluate the effectiveness of this deployment.We place a copper tag on the seat, as depicted in Figure 3(a), and follow the same experiment procedure described in the early sections.The measured energy of the received signals from the tag on the seat is highlighted in yellow in Figure 3(b).We can see that the energy collected from the tag on the seat is approximately 50% of the energy when the tag is directly   touching the body skin because of the clothes between the tag and human body.Thus, we need to improve the tag's receiving performance without compromising the comfortableness of the system.To achieve this goal, we investigate the effect of tag parameters as follows: Tag size.Intuitively, we first study the impact of tag size on the received signals.In addition to the default size (6 cm × 2 cm), we include five additional copper sheets with varying dimensions: 3 cm × 1 cm, 9 cm × 3 cm, 12 cm × 6 cm, 15 cm × 9 cm, and 15 cm × 15 cm, as shown in Figure 4(a).To ensure experiment reliability, all tag centers are placed at the same position of the seat.Subsequently, we ask the target to sit in the seat and stay stationary while the vehicle is moving.
The measured energy values of the signals collected from tags of different sizes are presented in Figure 4(b).The results show that larger tag sizes yield higher signal power.This is because larger copper sheet can create a stronger coupling with the human body.However, there exists a saturation size (15 cm × 9 cm) for this increasing trend.This is because after the tag size is increased to a particular value (i.e., reaching the saturation point), the bottleneck of the coupling becomes the human body.
Tag material.We further investigate the influence of tag material on the collected signal.For this study, we employ five metal sheets made of copper, nickel, titanium, lead, and steel, all with the same size of 15 cm × 15 cm as illustrated in Figure 5(a).In consideration of comfort, we also explore the use of conductive fabric in our system design.The results are presented in Figure 5(b).Remarkably, the tags made of varying materials show similar performance on signal reception.This outcome is attributable to the sufficiently high conductivity exhibited by all these materials, enabling effective collection of electromagnetic signals from the human body through coupling.Given the comparable efficiency across the tested materials, we opt for the conductive fabric as the tag material for our system to provide better user experience.
Tag position.Lastly, the tag position is also considered as a parameter influencing the signal receiving performance.We tape a conductive fabric tag with a size of 6 cm × 2 cm at six different positions on the seat as shown in Figure 6(a) and conduct the experiment.Figure 6(b) details the collected energy levels from the tag located at six different positions.We can see that similar energy levels are measured.We believe the small difference is due to the target's sitting posture.
Final design: cloth tag placed on the seat.Considering all the aforementioned factors, we choose to position the conductive cloth tag on the seat to collect body-captured signals within the electric vehicle, as depicted earlier in Fig- ure 1.This placement ensures not only convenience but also continuous contact between the tag and target body.The main part of the tag has a size of 9 cm × 3 cm and is linked by a 1 cm-wide strip for connection.This strip serves as part of the tag and also as the transmission line to transfer the collected signals to the processing board.The cost of the overall tag design is less than 10 cents.We employ the Arduino Nano Board to capture signals from the tag at a sampling rate of 2000 samples per second and subsequently store the collected data.The sensing performance can be slightly increased with a larger tag size, but the improvement is limited.Our current choice is to keep the tag size compact and still maintain a good sensing performance.

Software Design
Given the distinctive attributes of the leaked signals in electric vehicles, traditional wireless sensing models are inapplicable.Consequently, we must develop new sensing models for EVLeSen.We attach the cloth tag described in Section 3.2 to the seat of the Tesla Model 3. The human target sits naturally in a comfortable manner in the seat.The vehicle is driven along a country road, adhering to the speed limits.
Different from previous benchmark experiments in which the target stays stationary, in this experiment, the target performs movements such as raising one hand.To eliminate the interference caused by varying vehicle speeds, which will be detailed in Section 4, we keep a constant car speed during the process.Based on our analysis, the induced signal variations caused by body motions are related to two models.
Sensing Model I: Body-capacitor model.Due to the body's composition, human body can be considered as a good electrical conductor in our sensing system.The body, along with the electric vehicle (a large conductor), forms a    large capacitor with a capacitance of  body as depicted in Figure 7. Additionally, the body always carries a certain amount of charge,  body , due to triboelectric effects [41], which is the underlying reason for static electricity.Consequently, a potential difference  body exists in the target body, given by where the body charge  body is assumed not changing, while the capacitance  body between the body and electric vehicle changes due to different body motions, according to the equation where  0 is the electric constant,   is the relative permittivity of the air,  is the area of overlap between the body and the electric vehicle, and  is the distance between them.Different motions cause changes in  and , thus altering  body .The changing capacitance further influences the body's electric potential  body , which is measured through the cloth tag.Therefore, when the target performs a specific motion, the voltage detected by the tag changes as illustrated in Figure 7.
Since the frequency of body movements is low, the signal variations induced by motions based on this model only happen in the low frequency range.
Sensing Model II: Body-antenna model.The second model takes a different perspective on the human body, viewing it as an electrical conductor that functions as an "antenna" to receive the electromagnetic signals in the electric vehicle.As an "antenna", the body's postures alter the antenna's physical size and surface area, causing variations of the received signal.Furthermore, distinct body postures also vary the physical distance between the receiving "antenna" (body) and the signal source ("electric vehicle"), leading to changes in the received signal amplitude.Consequently, postures such as raising one's hands tend to gather more signals compared to postures where the hands are down at the sides, as shown in Figure 8.Since the leaked signals emitted within the electric vehicle predominantly occur in the frequency range of tens of Hertz (e.g., 60 Hz in Tesla Model 3), this motion-induced influence manifests in the slightly higher frequency range.
Signal processing based on two sensing models.The above two sensing models show that a body motion simultaneously influences body-captured signals in two distinct frequency ranges.To analyze these effects, we initially apply the Discrete Fourier Transform (DFT) to the collected signal samples, resulting in frequency domain representations as shown in Figure 9(b) and Figure 9(f).These frequency spectra align with the observation that vehicle leaked signals span from several Hertz to hundreds of Hertz, with significant components typically below one hundred Hertz.
In light of this, we utilize both low-pass and band-pass filters to extract the low-frequency component (inspired by Sensing Mode I) and high-frequency component (inspired by Sensing Model II) of these signal samples, as explained earlier.The outcomes are illustrated in Figure 9(c), 9(d), and Figure 9(g), 9(h) respectively.By analyzing the two frequency components of the samples, it becomes evident the motion's effects manifest when the target raises one hand, as observed through the lens of the two sensing models.These frequency components will be employed in the subsequent section for gesture recognition purposes.

SENSING UNDER THE VARYING SPEED OF ELECTRIC VEHICLES
The

Impact of EV's Speed
To present the relationship between the leaked signals and the vehicle speed, we ask the target to sit in the seat naturally and drive the vehicle on several roads with different speed limits.The measured energy of the leaked signals with changing speed is shown in Figure 10(a), from which we can see that an acceleration leads to larger leaked signals.
Besides the vehicle speed change, we also notice that the road condition, especially the gradient, also influences the working status of the car engine and thus affects the leaked signal.As shown in Figure 10(a), we can see that despite relatively stable vehicle speed during the interval from 22 s   to 28 s, the energy of the leaked signals varies.This is attributed to the impact of the road gradient on the engine's behavior.Another intriguing observation is that the leaked signals also exhibit an increase when the vehicle decelerates, as seen in the time interval from 0 s to 5 s in Figure 10(a).This phenomenon can be attributed to the regenerative braking system employed in electric vehicles, which captures the kinetic energy during braking and converts it into electrical power while simultaneously slowing down the vehicle.
To further analyze the signal changes correlated with vehicle speed, we apply the band-pass filter to extract signal component in the high frequency range, i.e., around 60 Hz. Figure 10(a) displays the energy of the filtered signal and the energy of the original signal across different vehicle speeds, from which we can see that signal patterns for them are similar.This means the predominant component of the signal changes caused by vehicle speed change is located in the high-frequency range.Figure 10(b) showcases a more detailed zoom-in view of the original signal and the filtered signal.It presents additional evidence that the alterations in the high-frequency range are the primary contributors to the signal changes induced by vehicle speed change.
The signal changes induced by varying vehicle speed can significantly affect the sensing performance of EVLeSen.These changes become entangled with the signal fluctuations arising from the target movement.Moreover, these signal changes caused by vehicle operation are beyond our control and difficult to predict.Dealing with such interference is challenging.One approach is to record changes in the vehicle's speed and deduce the corresponding signal alterations for removal.However, constructing a precise model that captures the complex relationship between the alterations in leaked signals and the diverse factors associated with vehicle operations is challenging.

Reference Tag Design
Based on the above analysis of the interference within the vehicle, we propose to deploy a reference tag as illustrated in Figure 11(a) to mitigate such interference.The key design is that the reference tag captures the leaked signals from EV without connections with the human body.It means that the reference tag only exhibits the signal variation induced by EV, while the tag on the seat (sensing tag) exhibits both the signal variation caused by EV and the signal variation induced by body motion.By exploiting the signal streams from the two tags, our system can eliminate the interference caused by the vehicle.
We also use two separate processing boards to independently sample the signals from the reference tag and the sensing tag, to mitigate the coupling between these two tags caused by the ground loop effect [13].By employing an open-circuit design to enhance the signal amplitude, the tags are virtually linked to the ground pin on the board via air coupling.If a single processing board is used, the two tags share the same ground pin through air coupling and the the received signals interfere each other.Thus, we adopted two boards to minimize this common ground effect.

Extracting the Signals
After collecting the data from the sensing tag and reference tag, we need to further process them to get a clear signal pattern caused by the body motion.We exploit the observation that the signal changes predominantly manifest within a specific frequency range.Leveraging this insight, we employ a digital band-pass filter to extract signals acquired by the reference tag within the target frequency range.The thermal noise spans across the whole frequency spectrum, and the environmental noise, which the electromagnetic signals emitted by the surrounding devices only occupies particular ranges.Since our focus is on the close-to DC band and the 60 Hz frequency band, the impact of noise and interference can be mitigated through filtering.As depicted in Figure 11(b), the filtered signal from the reference tag encompasses the signal change induced by the vehicle's operation (highlighted in the red rectangular region) but not the signal change stemming from body motion (highlighted in the blue rectangular region).
The signal variation caused by body motions are still unclear in the filtered signals as shown in Figure 9(h), since the 60 Hz leaked signals dominates.Thus, we need to further extract the patterns introduced by the body and the vehicle ( () and  (), respectively) from the 60 Hz leaked signals.
We can treat the patterns as information modulated on a constant sine wave cos() (i.e., 60 Hz sine wave).Consequently, the received signals at the reference tag,  ref (), and the sensing tag,  sense (), can be expressed as follows: sense () =  () () cos().
(4) By applying self-multiplication, we have the processed signals denoted as  ref () and  sense (), respectively as follows: By applying a digital low-pass filter on the above signals, we can extract the signal patterns attributed to the body ( 2 ()) and the vehicle ( 2 ()).The processed signals from the sensing tag comprise both the influence of body motion and vehicle speed.On the other hand, the processed signals from the reference tag only encapsulate the vehicle speed influence, which can then be used to counteract the speedinduced interference in the signals of the sensing tag.

Multi-Channel Training Network
After investigating the two sensing models based on the EV-leaked signals and the strategy to mitigate the impact of the varying vehicle speeds, we can now incorporate these three distinct information channels to recognize the target motions.These three channels consist of: We discover that while Channel 1 and Channel 2 capture body motions using two distinct models, it is ineffective to perform motion recognition by using separate models for each channel and then combining their results.This inefficiency arises because the motion information in these two channels, obtained from different sensing models, remains interconnected.Besides, we also notice that directly mitigating    the vehicle-induced interference by subtracting the signal of channel 3 from the signal of channel 2 is challenging.This difficulty stems from the difference in the vehicle-induced signal changes recorded at the two tags due to hardware imperfections.Thus, to integrate these three channels more effectively, we employ them as inputs for a neural network model, ultimately leading to the final output of motion recognition.The Neural Network comprises two feed-forward layers with sigmoid hidden neurons and softmax output neurons, which are well-suited for the classification task of different motions.The signals from these three channels undergo downsampling and concatenation to create the inputs for the neural network.This simplified neural network effectively leverages the interconnections among the three information channels, enhancing the accuracy of motion recognition.The overall signal processing pipeline is illustrated in Figure 12.

IMPLEMENTATION
Now we introduce the implementation details of EVLeSen.As shown in Figure 13(a), we deploy two tags: one sensing tag and one reference tag.The sensing tag is deployed on the seat to capture the EV-leaked signals through the human body, while the reference tag is deployed on the side face of the vehicle's center console to collect EV-leaked signals without body contact.These two tags are made of flexible conductive fabric (cloth), and each has a size of 9 cm × 3 cm.Each tag costs less than 10 cents.To avoid the coupling effect between the sensing tag and reference tag, we employ two Arduino Nano boards as shown in Figure 13(b), each of which is designed to collect data from one tag.The sampling rate of the MCU is set as 2000 samples/s, which is high enough to sample the EV-leaked signals whose frequency is mainly below 1000 Hz.We also program the two Arduino boards to ensure they concurrently run the same code for collecting the leaked signals.We employ a fabric line with a width of 1 cm to connect each tag to its MCU board for data collection.
To enhance the signal power from the tag, we employ an open-circuit design.
We collect data using Arduino software and implement the proposed signal processing schemes in MATLAB.The signal processing happens on a ThinkPad X1 laptop with an Intel Core i7 CPU and 16 GB RAM.
We believe our design can be integrated into future electric vehicles.The textile-based fabric tags can be incorporated into the seat without affecting the customers' experience, while the MCU board can be embedded inside the electric vehicle.More importantly, EVLeSen operates without adversely affecting human health.The open-circuit design ensures that the current flowing through the body is minimal.All the conducted experiments were IRB approved by our University.

EVALUATION
In this section, we first present the preliminary evaluation of EVLeSen.Then, we evaluate the system's robustness under different conditions.

Motion Recognition
The experiment setup is described below.We perform the experiment in a Tesla Model 3 vehicle, deploying EVLeSen on the passenger seat.The vehicle is driven on a highway with a speed limit of 65 mph.To evaluate the motion recognition accuracy of EVLeSen, we employ ten different body motions, including three types of hand movements, three types of foot movements, three types of head movements, and one default posture without any movements, as shown in Figure 14(a).The target naturally sits in the passenger seat equipped with EVLeSen and performs each movement one hundred times when the vehicle is driven on the highway.For the collected data samples in each scenario, we divide them randomly into training (70%), validation (15%), and test (15%).We adopt the sensing accuracy as the evaluation metrics, which is defined as the ratio of correctly classified samples to the total number of samples in the test data set.We utilize the scaled conjugate gradient [35] for the training process.
The experiment results are shown in Figure 14(b).First, we observe that EVLeSen is able to achieve an average motion recognition accuracy of 94.4%.Second, for the small-scale movements, such as foot motions (motion index 5 to 7) and head motions (motion index 8 to 10), which are challenging to simultaneously recognize in traditional RF signal-based sensing systems [20], our EVLeSen is able to achieve high recognition accuracy.This is mainly because EVLeSen benefits from leveraging the human body capacitance for sensing.These promising results show that EVLeSen could offer great potential to enhance user-vehicle interaction capabilities by incorporating foot and head motions in addition to traditional hand movements.Finally, we find that lifting the foot can be erroneously categorized as a movement of the head.This misclassification arises because only the amplitude information can be utilized.Actions like nodding the head and lifting the leg both result in part of the body detached from the vehicle.These dramatically different motions thus generate relatively similar patterns in signal amplitude, causing ambiguities.

Robustness to Road Conditions
To further evaluate the robustness of our EVLeSen in practice, we conduct experiments on four distinct road conditions, as depicted in Figure 15(a).These four roads have different conditions in terms of speed limit, traffic signs, pedestrian crosswalks, and speed bumps.During each experiment, the target performs four representative motions-sit normally without other movements, vertical hand movement, lifted foot, and nodding-each for one hundred times.To evaluate the effectiveness of our proposed signal processing methods, we also compare EVLeSen with two other signal processing methods: a) Raw: directly puts the collected samples from the sensing tag into the neural network; b) No Reference: processes the collected samples from the sensing tag through the complete pipeline shown in Figure 12, excluding the data from the reference tag for interference mitigation.Note that compared to these two methods, EVLeSen leverages the samples from both the sensing and reference tags following the complete pipeline shown in Figure 12.Below, we analyze the experiment results shown in Figure 15 obtained under these four road conditions separately.
(1) Highway.The highest sensing accuracy of EVLeSen is obtained when the vehicle is driven on the highway, among the four different road conditions.Furthermore, the difference in accuracy for the three signal processing methods is the smallest.This can be attributed to the relatively stable nature of the EV-leaked signals on highways, where the vehicle's speed variations are less frequent.As a result, the signal patterns caused by body movements become more distinguishable in this scenario, even without the extra information extracted from the reference tag.
(2) Freeway.Besides a lower speed limit compared with the highway, the vehicle still operates in a similar stable status without much speed change.The good performance in freeway scenario demonstrates that EVLeSen can accurately recognize body motions under different vehicle speeds.(3) Downtown.This road type poses the most challenges for EVLeSen, leading to the lowest sensing accuracy among the four road conditions, as it involves frequent and substantial changes in vehicle speeds due to factors like road turns, traffic signs, and pedestrian crossings.These rapid speed fluctuations can significantly disrupt the pattern of EV-leaked signals.Consequently, if the collected signals are directly used for motion recognition, the system's accuracy drops to 75%.Through processing the signals based on the proposed two sensing models, the accuracy is improved to 83.3%.By applying the proposed signal processing methods together with the reference tag design, the accuracy can be further increased to 90%.
(4) Village driveway.The driveways around houses and apartments offer driving conditions similar to those of the downtown road but at lower speeds.Consequently, while speed changes are still frequent, they are of smaller magnitude.Additionally, these driveways often feature speed bumps designed for pedestrian safety.The speed bumps lead to speed changes and introduce large vehicle vibrations, which can fail conventional RF sensing systems [22,68].This is because the distance between the sensing device and the human body experiences sudden changes.In contrast, owing to the unique characteristics of EV-leaked signals, the proposed system is immune to changes in device-body distance.The proposed EVLeSen is thus still able to achieve a sensing accuracy of 93.3% under this road condition.

Robustness to Vehicle/User Diversity
Vehicle status.Next, we evaluate the impact of EV's working status on the sensing performance of EVLeSen.In the experiments conducted to assess the robustness of the system in this section, we focus on four representative motions: sitting, vertical hand movement, lifting the foot, and nodding.We explore six different vehicle statuses as illustrated in Figure 16(a): default, air conditioner on, external lights on, windshield wiper on, streaming media via Bluetooth, and low-power mode of the car is on.The default status means that all the vehicle's electronic components described above are switched off.The experiments are conducted on the highway.The results are shown in Figure 16(b).We can see that the recognition accuracy remains consistently higher than 93% across the six different vehicle statuses.This demonstrates the robustness of EVLeSen under different working statuses of the EV, even during night driving (the vehicle interior is dark), which is a challenging scenario for camerabased sensing [3].
Vehicle diversity.To evaluate the robustness of EVLeSen under different vehicle models, we employ three vehicles of different models: Tesla Model 3, Chevrolet Bolt, and Hyundai Kona.Their features are summarized in Table 1.The invehicle sensing accuracy in different vehicles is shown in Figure 17.We collect the signal energy readings from each vehicle type and calculate the average energy based on Equation (1).When we collect the signal energy readings from each vehicle, we maintain the same speed and keep the road condition the same.We observe that the energy of the EVleaked signals differs across vehicles due to variations in their electronic components and internal Specifically, the leaked signal from Hyundai Kona is weaker compared to the signals leaked from the other two vehicles.Therefore, the recognition accuracy in Hyundai Kona is 3% lower than that of the other two vehicles, as shown in Figure 17.Nevertheless, the sensing accuracy remains consistently above 92% across different vehicle types, affirming the robustness of EVLeSen against vehicle diversity.
Clothing style.We then conduct experiment to investigate the influence of the target's clothing on the sensing performance of EVLeSen.The target was instructed to wear different clothes as depicted in Figure 18(a).Different from the default setup, we collect data from one clothing style for training and apply the model to test other clothing styles.The results are shown in Figure 18(b).These results confirm the system's robustness against variations in the target's clothing style, even when the target wears metal accessories.The reason is that the body has a much larger physical size compared with the clothe and the accessories.Thus, the changes in body motion dominate the signal pattern change.
User diversity.We also evaluate the effect of user diversity, including age, gender, weight, and height, as detailed in Table 2.This table also outlines the average energy of the received signals for each user in a stationary position  within the same vehicle.Notably, the energy does not exhibit a linear correlation with body weight and height due to variations in sitting postures and positions.Despite the differences in captured energy, from the experiment results shown in Figure 19, we can see that EVLeSen maintains consistent performance across different users.The sensing accuracy is always higher than 94%, demonstrating the robustness of EVLeSen against user diversity.

Multi-Target Motion Sensing
Multi-target motion sensing is a well-known challenge in contact-free RF sensing [59,64].EVLeSen is capable of sensing multiple targets' motions by deploying multiple cheap sensing tags.As shown in Figure 20, we conduct experiments with three targets to evaluate the performance of multi-target motion sensing.For each target, we deploy a sensing tag on the corresponding seat.Note that we only need to deploy one reference tag.In this experiment, participants are instructed to perform one of the four motions-sitting, vertical hand movement, lifting the foot, or nodding.The experiment results are also shown in Figure 20 (highlighted in white color).We can observe that the average recognition accuracies for the three targets are 94.1%, 93.3%, and 92.9%, respectively.This experiment result demonstrates the capability of sensing multiple targets simultaneously with EVLeSen.Due to the low frequency (<200 Hz), the leakage signal wavelength is extremely large.The surrounding participants have almost no influence on the wireless channel if they are not in direct contact with the tag.Thus, EVLeSen is robust against surrounding interference.

DISCUSSION
Potential applications.Our system has the potential to enable a large range of applications, such as user-vehicle interaction, fatigue driving detection, and occupancy detection.
Note that in this paper, to ensure safety, we only conduct experiments on the passengers' seats.If we deploy the sensing tag on the driver's seat, EVLeSen can also effectively monitor the motion of the driver, which can help foster good driving habits.The unique body-captured signals associated with the users could also be leveraged for in-vehicle user identification.
Difference with other sensing methods.The proposed method eliminates the need for dedicated devices to actively transmit sensing signals, such as mmWave, sound, and laser signals.Instead, it relies solely on existing leaked signals within the electric vehicle, resulting in cost saving and easier adoption.Compared to camera-based methods, the proposed method offers more privacy protection.Besides, unlike existing wearable sensors, which can only monitor the movements of the body contact points, EVLeSen can track fullbody motion regardless of its contact position on the body.Consequently, the seat tag not only increases the user's convenience by eliminating the need for additional wearables but also enables whole-body motion sensing.
Advanced machine learning model.EVLeSen employs a simple Neural Network for motion recognition.In future work, machine learning models with transfer learning capabilities can be utilized to handle scenarios involving diverse users and different vehicle types, to reduce the training burden.Furthermore, we anticipate that advanced deep learning models can be employed to recognize more complicated body motions that involve movements of multiple body parts.
Non-electric vehicles.Besides various electric vehicles, we also conducted experiments with non-electric vehicles, specifically the KIA Forte 2018.The results reveal that the energy leakage from KIA Forte 2018 is only 5% of the energy leakage in the Tesla Model 3. The classification accuracy for the four motions-sitting, vertical hand movement, lifting the foot, and nodding-is 85.2%, which is 10% lower than that achieved in electric vehicles.

RELATED WORK
EM leakage.Researchers have explored electromagnetic (EM) leakage from various sources such as power lines [18,30], electric appliances [19,43], computers [37,61], and even visible light communication [8,10].These leaked signal have been exploited for energy harvesting [18], indoor mapping and navigation [32], information sniffing [61], data communications [9], and gesture recognition [6].EVLeSen is the first work to explore leaked signals from electric vehicles as a new sensing modality that can empower various applications such as user-vehicle interaction.
In-vehicle sensing.In-vehicle sensing has gained significant attention from academia and industry [62] as it can detect risky driving behaviors and facilitate user-vehicle interface.Various methods using cameras [39,49], biological sensors [16,69], RF signals [55,67,68], and wearable sensors [26,50], are designed to capture human motions inside vehicles.Previous studies also employ pressure sensor [7], thermal sensor [14], capacitance sensor [2] to infer vehicle occupancy.In contrast, EVLeSen introduces a new in-vehicle sensing modality, leveraging leaked signals from electric vehicles for sensing.
Wireless sensing.In the last few years, various dedicated wireless signals, such as WiFi [42,64], LoRa [58,59], and UWB [28,65], have been utilized for sensing.A large range of applications have been enabled, e.g., fall detection [54], vital sign sensing [66], and sleep monitoring [28].Owing to the ultra low frequency of leaked signals, the sensing approach based on leaked signals differs significantly from those developed for dedicated signals in terms of sensing model and system design.There also exist studies which utilize the ultra-low frequency leaked signals for human sensing [6,11].EVLeSen differs in the following aspects: a) Signal pattern: the signals from electric vehicles are more dynamic compared with the leaked signals from powerlines [6,11]; b) Receiver design: different from prior methods requiring users to wear devices on their bodies [6], EVLeSen places a tag on the seat which is less intrusive; c) Sensing model: EVLeSen employs two distinct models, i.e., body-capacitor model and body-antenna model to explain the influence of body motion on leaked signals from electric vehicles.

CONCLUSION
In this paper, we introduce a new sensing modality for invehicle sensing by exploiting the leaked electromagnetic signals from Electric Vehicles (EV).Leveraging the body's bio-electric properties, the human body is treated as both a capacitor and an antenna, leading to concurrent influences on the body-captured leakage signals.Besides, a reference tag is incorporated to counteract interference such as varying vehicle speeds.Utilizing cloth tags positioned on the seat without compromising comfortableness, our proposed system can accurately identify various body motions.We envision the proposed sensing modality can trigger new invehicle sensing applications to enhance driving safety and enable smarter user-vehicle interactions.

Figure 1 :
Figure 1: Proposed EVLeSen: capturing with body the electric vehicle's leakage signal for in-vehicle sensing.
connection to EV W/ connection to EV (b) Energy vs. configurations

Figure 2 :
Figure 2: Preliminary capturing of the leakage signal.
Energy vs. tag placements

Figure 3 :
Figure 3: Capturing signal with a tag placed on the seat.
Energy vs. tag sizes

Figure 4 :
Figure 4: Tag sizes influence on the signal energy.
Energy vs. tag materials

Figure 5 :
Figure 5: Tag materials influence on the signal energy.

Figure 6 :
Figure 6: Tag positions affect the signal energy.
(a) Signals with static target (b) Signals with target motion

Figure 7 :
Figure 7: Body-capacitor model: target motion influences the capacitance, resulting in voltage change.
(a) Signals with static target (b) Signals with target motion

Figure 8 :Figure 9 :
Figure 8: Body-antenna model: posture with larger physical size receives more signals.
two aforementioned models have established meaningful theoretical connections between the received signals and corresponding motions.However, real-world application introduces additional complexities as the EV-leaked signals are mixed with various forms of interference.Since these signals are leaked from the electric vehicle components, the energy levels of the leaked signals are contingent upon the operational status of these components.Following comprehensive experiments encompassing diverse EV conditions, we discerned that the vehicle's speed variation (i.e., acceleration and deceleration) has the greatest influence on the leaked signals, outstripping other factors (will be elaborated in more detail in Section 6).For example, an acceleration of the electric vehicle indicates an elevated electric current in the electric motor and correspondingly larger leaked signals within the vehicle.

Figure 10 :
Figure 10: Energy of body-captured signals when the user is static but vehicle speed changes.

Figure 11 :
Figure 11: Exploiting a reference tag to address the vehicle-induced interference.

Channel 1 :
the close-to DC component of the signals captured by the sensing tag, derived from the body-capacitor sensing model (i.e., Sensing Model I presented in Section 3.3).• Channel 2: the pattern of the high-frequency component of the signals captured by the sensing tag, derived from the body-antenna sensing model (i.e., Sensing Model II).• Channel 3: the pattern of the high-frequency component of the signals captured by the reference tag to counteract interference.
Conductive cloth tags MCU boards for two tags (b) Arduino board as MCU

Figure 13
Figure 13: Implementation of the proposed EVLeSen.

Figure 14 :
Figure 14: Motion recognition: recognizing ten in-vehicle motions when the EV is driven on a highway.
Four common road conditions tested in our robustness evaluation of the proposed EVLeSen system

Figure 15 :
Figure 15: Evaluation of the robustness performance of EVLeSen to different road conditions.

Figure 18 :
Figure 18: Impact of the clothing styles.

Figure 19 :
Figure 19: Impact of user diversity.

Figure 20 :
Figure 20: The results of multi-target motion sensing.

Table 1 :
Parameters of the three used electric vehicles.

Table 2 :
Parameters of the users.