Fall Detection IoT Elderly

Fall Detection IoT Elderly — Topics for IoT Students

IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2024 1

Real-Time Fall Detection Using Smartphone

Accelerometers and WiFi Channel State Information Lingyun Wang, Deqi Su, Aohua Zhang, Yujun Zhu, Weiwei Jiang, Xin He, Member, IEEE, Panlong Yang, Senior Member, IEEE

Abstract— In recent years, as the population ages, falls have in-

arXiv:2412.09980v1 [cs.LG] 13 Dec 2024

creasingly posed a significant threat to the health of the elderly. We propose a real-time fall detection system that integrates the inertial measurement unit (IMU) of a smartphone with optimized Wi- 1st Stage

Fi channel state information (CSI) for secondary validation. Initially, IMU Data

IMU Model

the IMU distinguishes falls from routine daily activities with mini- 2nd Stage

mal computational demand. Subsequently, the CSI is employed for Fall

further assessment, which includes evaluating the individual’s post- CSI Data Have the Capacity to Act?

fall mobility. This methodology not only achieves high accuracy but also reduces energy consumption in the smartphone platform. An CSI Model

Android application developed specifically for the purpose issues

an emergency alert if the user experiences a fall and is unable to move. Experimental results indicate that the CSI model, based on convolutional neural networks (CNN), achieves a detection accuracy of 99%, surpassing comparable IMU-only models, and demonstrating significant resilience in distinguishing between falls and non-fall activities. Index Terms— Fall detection, accelerometer, gyroscope, Wi-Fi channel state information, machine learning.

I. I NTRODUCTION employed to detect falls through the analysis of motion pat-

terns. They offer high accuracy and function effectively across Falls constitute the second most common cause of unin- diverse environments. However, their efficacy is contingent tentional injury-related fatalities and are a significant factor upon user compliance, which poses a significant challenge, that causes injuries and deaths among older adults.

Such particularly among elderly users who may forget or neglect to incidents commonly result in fractures, head injuries, and other wear these devices [14]. In contrast, ambient sensors, including grave health issues. According to estimated data collected by infrared and thermal sensors, monitor environmental changes WHO [1], each year falls result in 684,000 fatalities, predom- to identify falls without user involvement, effectively preserv- inantly in low- and middle-income countries.

Annually, 37.3 ing privacy. However, their effectiveness is limited to areas million falls are severe enough to require medical attention, within sensor range, making extensive coverage potentially leading to significant health impacts, long-term care needs, costly [15]. Vision-based systems, which utilize cameras and and substantial financial costs [2]–[5]. computer vision techniques to detect falls by analyzing visual Fall detection hence is of significant importance.

Researches data, can achieve highly precise fall detection autonomously, carried out in this domain encompass a broad range of yet they pose significant privacy concerns and are generally technologies and methodologies aimed at identifying falls, par- fixed in specific locations, requiring extensive setup [16]. ticularly among the elderly [6], to prevent serious injuries and Recent advancements in fall detection have focused on fusion- fatalities.

The primary approaches include wearable sensors, based methods, which combine data from multiple sensors ambient sensors, and vision-based systems [7]–[13]. Wearable to improve accuracy and reliability, thus enhancing overall sensors, such as accelerometers and gyroscopes, are widely reliability [17]. This trend is evident in the integration of radio This work has been in part supported by the Natural Science Foun- frequency (RF) signals with traditional sensor data to enhance dation of China under grant No. 62072004.

Corresponding author: Xin detection capabilities [18]. He, Yujun Zhu. For instance, Ponce [19] et al. demonstrate the effectiveness X.

He, D. Su, L. Wang, A.

Zhang, and Y. Zhu are with the School of Computer and Information, Anhui Normal University, Wuhu, 241002, An- of combining wearable sensors and vision devices to achieve hui, China (e-mail: {lingyunwang, zhangaohua, sudeqi, xin.he, zhuyu- high accuracy in fall detection by analyzing sensor locations jun}@ahnu.edu.cn). and minimal deployment strategies. Sowmya and Pillai [20] W.

Jiang and P. Yang are with the School of Computer Science, Nanjing University of Information Science and Technology, 210044, highlight the potential of using machine learning algorithms Jiangsu, China (e-mail: {weiwei.jiang, plyang}@nuist.edu.cn). with wearable sensors to detect falls, showing that ensemble

2 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2024

Wi-Fi Base Station

techniques like random forest classifiers can achieve up to 98% accuracy. Additionally, Ramanujam and Padmavathi [21] offer non-invasive and cost-effective solutions, with their re- search indicating that these systems can outperform traditional sensor-based methods in activity recognition and fall detection. Khan [22] et al. came up with the idea of integrating RF signals with traditional methods to further enhance detec- Elderly Living Alone tion capabilities, addressing privacy concerns associated with wearable and vision-based technologies.

These advancements Community

highlight the ongoing efforts to improve fall detection systems Fig. 1: An example application scenario for fall detection through the integration of diverse sensor technologies and using Wi-Fi CSI is monitoring elderly individuals living alone, sophisticated algorithms. where our system can immediately detect a fall and issue Researchers in fall detection face several significant chal- warnings when necessary. lenges.

A major challenge in fall detection is the high rate of false positives, where non-fall activities such as sitting down quickly or abrupt movements are misinterpreted as falls. for real-time fall detection. Lo [23] et al. emphasize the critical need for sophisticated Several innovative contributions of our work to the field of algorithms in their FPGA-based study that can more effec- fall detection can be summarized as follows: tively differentiate between fall and non-fall events, including 1) We propose a real-time fall detection system that only situations where individuals lie down quickly or end in a requires a smartphone connected to home Wi-Fi, elimi- sitting position.

This necessity underscores the importance of nating the need for additional tools. Our developed app developing advanced detection algorithms to minimize false implements AI algorithms for fall detection, offering positives and enhance the reliability of fall detection systems. a convenient, low-power solution easily integrated into Real-time processing of fall detection data is critical for daily life, especially for the elderly or those at higher timely intervention but presents a significant technological risk of falls. challenge due to the computational demands involved.

The use 2) Our approach combines sensor technology (accelerome- of millimeter wave signals in systems like mmFall [24] exem- ters and gyroscopes) with wireless communication per- plifies a solution that achieves high accuracy while maintaining ception. The system uses a two-stage detection process: low computational complexity which utilizes spatial-temporal first, detecting falls via IMU sensors, and then double- processing alongside a lightweight convolutional neural net- checking the user’s ability to move post-fall.

Additionally, work (CNN). However, the high cost of such advanced systems the CSI-based detection ensures continuous monitoring remains a considerable barrier to widespread adoption. within indoor spaces, even without the user carrying User compliance is another problem, as the effectiveness of a device, enhancing safety in settings like elderly care wearable sensors depends on consistent use by the elderly, who facilities. may forget or refuse to wear the devices [25].

Furthermore, there is a lack of extensive real-world datasets that include We yield several important findings and concrete results. By actual falls by elderly individuals, which limits the ability integrating IMU and CSI data, we significantly reduced the to validate and improve detection algorithms effectively [26], number of false positives and improved the overall accuracy [27]. Ethical concerns also prevent the collection of such data, of fall detection.

Our system is able to accurately distinguish leading to reliance on simulated falls by younger individuals, between falls and other rapid movements, such as quickly which may not accurately represent real fall dynamics. picking up a phone, demonstrating its practical applicability. Unlike existing fall detection methods that primarily rely on The real-world testing of our system shows high accuracy a single sensor modality, this study introduces a cross-modal rates in detecting falls, validating the effectiveness of our fusion strategy for real-time fall verification.

Our system approach. Furthermore, the use of CSI data as a secondary val- combines data from both the inertial measurement unit (IMU) idation step further enhanced the system’s reliability, providing and channel state information (CSI) via smartphones. The an extra layer of assurance against false positives.

The overall system initiates a data collection phase, during which IMU recognition accuracy of our system is 99%, which highlights and CSI data from various fall scenarios are collected using the potential of our integrated approach to offer a reliable and smartphones. The collected data undergoes preprocessing to efficient solution for fall detection, with the capability to be eliminate noise and normalize the inputs, which ensures the deployed in real-world scenarios. quality and consistency of the data.

The refined data is The rest of this paper is organized as follows: Section subsequently inputted into two distinct models: a multi-layer II details the system design of the proposed fall detection perceptron (MLP) model processes the IMU data, and a CNN system, including data acquisition, data preprocessing, feature model handles the CSI data. These models are trained to extraction, and the learning model design. Section III exam- distinguish falls by identifying the unique characteristics of ines the performance of our designed system from multiple fall events as opposed to normal activities.

The outputs from perspectives, comparing it with other models. Section IV and these models are integrated to improve detection accuracy and Section V conclude and forecast the research direction. minimize false positives, thereby providing a reliable solution

AUTHOR et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (MAY 2024) 3

Smartphone Amplitude of CSI data Router

Get data

-0.11246266,5.4203415,6.902695 -0.48753762,5.1810594,8.335996 -0.13938192,4.8693943,9.450453 -0.92482597,4.8598228,8.165507 CSI after DWT -0.69332033,4.7246284,8.390432

Merge Visualize Multilayer Perceptron

x1 x2 Fall Convolution x3 Input Layer x4 x5 y1 Max pooling y2 Fully connect x6 Softmax … x7 y3 . x8 . x9 . x10 y10 . . Confirm Check 1x2

Attention

Dropout .

Flatten x20 Dense Layer 1x64 Input Layer Other 29x106 IMU Model CSI Model

Fig. 2: The framework of the real-time fall detection system.

II. S YSTEM D ESIGN that the person has fallen and is unable to move, triggers

This section outlines our system configuration utilizing an an emergency response. This dual-analysis approach not only 802.11 wireless network interface card (NIC) to capture CSI detects falls but also assesses the condition of the individual data. The employed smartphone is the OPPO Reno9 Pro, afterward, enabling timely and appropriate interventions based which features a MediaTek 8100-MAX CPU and a battery on the severity of the incident and the person’s immediate capacity of 4500 mAh, while the computer used for data needs.

The framework of the fall detection system is shown processing and model training is outfitted with an NVIDIA in Fig. 2. GeForce RTX 3060 GPU. B.

System setups A. Overview of the system In daily activities, human motion is generally smooth and We design a smartphone-based system for real-time fall continuous, with actions such as walking, sitting, and turning detection and alert generation that integrates data from the to exhibit regularity and stability. However, during a fall, built-in accelerometer and gyroscope, alongside CSI, to iden- the body’s movements display sudden, unstable characteristics tify and respond to falls.

As depicted in Fig. 1, our system such as rapid descent, loss of balance, and collapse. These is primarily designed for one application scenario in which abrupt changes in motion patterns contrast sharply with ev- elderly individuals living independently. eryday behaviors and are crucial features for identifying fall- Initially, raw data from the accelerometer and gyroscope are related actions. We divide the falling process into three phases: collected and subjected to normalization to ensure consistency descent phase, impact phase, and stationary phase, as shown in the data scale and facilitate accurate analysis.

Subsequent in Fig. 3. feature extraction processes identify characteristics specific to Descent phase. During the descent phase of a fall, the fall patterns. These features are then used in a classification body experiences rapid changes in acceleration as it transitions algorithm to determine the occurrence of a fall.

Concurrently, from any potential pre-fall state to a rapid downward fall. In the system processes CSI data by first calculating the ampli- addition to this, the body often undergoes rotational and tilting tude of the Wi-Fi signals to gauge environmental interactions, motions, especially after losing balance. These movements followed by computing the rate of change in this amplitude to are attempts to maintain posture or find a support point. assess the dynamics and severity of the movement.

Compared to normal walking or other activities, this phase Our system further evaluates the individual’s capacity to typically exhibits irregular motion patterns, characterized by act post-fall. If the person can move, the system issues a more sudden movements. Fig. 3(a) illustrates the descent phase reminder indicating that although a fall has been detected, of a fall. the individual remains active and does not require assistance.

Impact phase. As shown in Fig. 3(b), during the impact Oppositively, if the individual is incapacitated, a warning alert phase, when a person falls and contacts the ground, a collision

4 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2024

(a) descent phase (b) impact phase (c) stationary phase Fig. 3: Three phases of a fall event: descent, impact, and stationary.

TABLE I: Composition of the collected dataset. Door

Sofa Sofa

Activity Explanation Data Size Desk Chair

fall fall down 820 Desk Desk play play with a smartphone 748 Desk pickput pick up and put down a phone 870 North sit sit down and stand up 876 Chair

squat squat down and stand up 818 SmartPhone Chair static stay still 896 walk walk around 759 Desk

Air-Condition West East up down put down and pick up 832 Desk toward down swing a phone quickly 866 Chair South throwing throwing a phone 820

Fig. 4: The environment of fall detection using IMU and Wi-

occurs between the body and the surface, which generates Fi CSI built in the smartphones. stress and propagates shock waves. At this moment, the change in acceleration is extremely abrupt. After hitting the ground, a rebound might occur.

This rebound is due to some of the and gyroscope features compared to the common everyday energy absorbed during the fall being released in some form actions, containing up and down, i.e., putting down and after the collision, causing the body to momentarily lift off picking up a phone quickly, toward down, i.e., swinging a the ground again. This phase is critical for detecting falls, as phone quickly, and throwing, i.e., throwing a phone onto a the acceleration patterns observed differ markedly from those sofa or bed. observed during normal activities or less severe falls.

The methodology for collecting data on specific actions Stationary phase. After a fall occurs, it can be cate- is structured as follows: Firstly, we gather data on falls gorized into two types: falls from which one can recover and common indoor actions, including the first and second independently, and falls from which one cannot. We primarily categories mentioned above.

We then use this data to design focus on non-recoverable falls, which are characterized by and train a machine-learning model, which is subsequently the individual lying on the ground, unable to move or seek tested in two parts. help on their own. It is only in these situations that a fall is The first part tests actions already known to the model to finally classified as a non-recoverable fall, which is the specific assess its ability to distinguish these from falls.

The second type of fall behavior that our research aims to monitor. The part tests actions not included in the model, such as taking stationary phase is shown in Fig. 3(c). a smartphone out of a pocket to see how this action is We collected data on ten different types of actions. Table I categorized.

If this action is easily misclassified as a fall, then lists the types of actions and the number of data samples we specifically collect data on this type of action and retrain collected for each. The goal is to accurately detect falls from the model. The newly trained model is then used to retest this among these various actions. action to see if it can be distinctly separated from falls in the These ten types of movements can be divided into three updated model. main categories.

The first category is falls. The second cate- The overall strategy involves training the model and then gory consists of actions commonly seen indoors in daily life, testing it against any action that might occur in daily life while including play, i.e., playing with a smartphone, pick and put, carrying a smartphone, to identify actions that are prone to i.e., picking up and putting down a phone, sit, i.e., sitting down being misjudged as falls.

Once a misjudged action is identified, and standing up, squat, i.e., squatting down and standing up, we collect data solely on that action, label it separately, and static, and walk. The third category comprises some less com- retrain the model. mon actions but, when analyzed through accelerometer and gyroscope data, these movements’ data tend to resemble the characteristics of falls more closely in terms of accelerometer

AUTHOR et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (MAY 2024) 5

y 20 20

10 10

Acc value

Acc value

x 0 0

-10 -10

-20 -20 0 10 20 30 40 0 10 20 30 40 Sampling point Sampling point z (a) before stage (b) after stage

Fig. 6: Comparison of the acceleration values before and after

the removal of the gravity. Fig. 5: Orientation of built-in sensors of smartphones.

C. Stage I

The first stage of a fall is defined as a combination of start fall over fall after the descent phase, the impact phase, and the onset of the stationary phase, which constitutes the main body of the entire t=0 t≈1.5 t=3 t/s falling process and is characterized by the greatest range of Fig. 7: The set-up window of a fall event. To cover the whole motion. Compared to everyday activities such as walking, duration of a typical fall event, we choose 3 s as a window sitting down, standing up, and squatting, the stage exhibits size. significant differences in the conditions of acceleration and body rotation during movement.

During the data collection phase, we conduct experiments An exponential weighted moving average filter is used using a mobile phone in the scenario shown in Fig. 4. The here to smooth the output of the accelerometer, reducing the sampling rate of the accelerometer and gyroscope is set to impact of high-frequency noise and vibrations. The parameter 10 Hz.

To ensure data diversity and experiment accuracy, the α represents the weighting relationship between the current experiments are carried out in an environment that both sim- and past data, where α is a constant between 0 and 1. If α ulates daily life and considers various possible fall situations. is close to 1, the current data has a greater weight, resulting With an OPPO smartphone carried in their pocket, our in a stronger filtering effect.

Conversely, if α is closer to 0, participants are asked to simulate various fall behaviors, in- the historical data has more weight, and the filtering effect is cluding slips, trips, and fainting in the four cardinal directions, weaker. For this case, α is set to 0.85. i.e. east, south, west, and north, represented by red dots in To compute the linear acceleration, we ascertain the object’s Fig. 4. This multi-directional simulation method ensures that true acceleration in space after the removal of gravitational the dataset covers all possible fall scenarios, thereby enhancing effects: the reliability of the system in practical applications. laccxt = accx − gx , (4) As illustrated in Fig. 5, when a smartphone is placed in different orientations, the accelerations recorded along the x, laccyt = accy − gy , (5) y, and z axes vary due to the presence of inherent gravity.

In lacczt = accz − gz , (6) reality, the position and orientation of a smartphone carried in a pocket differ each time, resulting in varied raw acceleration where laccxt , laccyt , and lacczt denote the actual acceleration data even for the same movement. Consequently, it is neces- values of a mobile phone moving in the x, y, and z axes, sary to remove the influence of gravity on the accelerometer respectively. The comparison of the acceleration values before to ascertain the actual acceleration values along the x, y, and and after the removal of gravity is shown in Fig. 6. z axes.

Subsequently, the acceleration values from these three Our data indicates that it takes approximately 1.5 seconds axes are vectorially combined to derive the acceleration value from the initiation of a fall to the impact stage. Therefore, we corresponding to the actual direction of motion. define a fall event as occurring within a 3-second window, as To remove the gravity component from the raw accelerom- shown in Fig. 7. eter data, a low-pass filter is employed, as: Due to the unknown orientation of smartphone placement on the body during human movement, it is necessary to obtain gx t = α × gx t−1 + (1 − α) × accxt (1) acceleration data independent of the placement direction. gy t = α × gy t−1 + (1 − α) × accy t , (2) Therefore, by adopting the method of vector composition, the accelerations of these three axes are combined to represent gz t = α × gz t−1 + (1 − α) × accz t .

(3) the actual acceleration value of the mobile phone, treated where gx t denotes the x-axis acceleration data at time t after as a point mass, as it moves with the human body. This removing the effects of gravity, while gx t−1 refers to the x-axis value reflects the acceleration magnitude in the direction of acceleration data from the previous moment after gravity has the smartphone’s movement. Fig. 8 exhibits the values of the been removed. accxt represents the x-axis acceleration data acceleration after synthesis along three axes. collected by the sensor at time t, which includes the influence According to the characteristics of the falling action, the of gravity. data from the first stage of the fall is further subdivided

6 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2024

Original First part 6000 After DWT 0-2s

Amplitude of CSI

Second part 2-3s

0 10 20 30 40 50 60 70 Sampling point

Fig. 10: CSI after DWT. Fig. 8: Values of the acceleration synthesized along three axes, further subdivided into two parts for the initial stage of the fall. continues with a second fully connected layer, taking the Amplitude of CSI

Fall output of the previous layer and processing it through a ReLU activation function to produce an array of 64 values. A second dropout layer follows, randomly dropping 5% of the neurons to further prevent overfitting. The flow then moves to a third Phase of CSI

fully connected layer, which takes the output from the previous layer, a 64-value one-dimensional array, connects it to 32 neurons, processes it through a ReLU activation function, and Doppler shift of CSI

outputs a 32-value one-dimensional array. Finally, the output layer takes this 32-value array, connects it to 11 neurons, and activates them with a softmax function to produce an 11- Sampling point value one-dimensional array, which represents the probabilities Fig. 9: Phase, amplitude, and Doppler shift of CSI corresponds of the different actions. The usage of a cross-entropy loss to a fall event. function with the softmax function as the activation for neural network outputs transforms the raw numerical output of the model into a probability distribution, ensuring outputs for each into two parts, as illustrated in Fig. 8.

The first part broadly category are between [0,1] and their total equals 1. Thus, the encompasses the falling and impact phases, while the second outputs can be interpreted as category probabilities. The model part includes a small portion before the stationary phase. learns to map inputs to category probabilities and selects the The reason is that during the first phase of the approxi- one with the highest probability as the prediction.

By applying mately 3-second fall, there is a significant disparity between the backpropagation algorithm, the model adjusts its weights the accelerometer and gyroscope characteristics between 0- to minimize the loss, improving classification accuracy for 2 seconds and 2-3 seconds. Each of these segments exhibits each category. The parameters of IMU model is indicated in distinct features.

In particular, within 1-2 seconds, the value Table III. of the accelerometer rapidly increases from around zero and then quickly drops back to zero. Throughout this process, D. Stage II the average value of the accelerometer is relatively high, and the rate of change is substantial.

The maximum value of the The primary purpose of this stage is to check whether or accelerometer during these 3 seconds generally appears within not the user still has the ability to move in the stationary the initial 0-2 seconds. The subsequent phase, which involves phase, using the Wi-Fi CSI data. Based on the configuration lying on the ground for about 1 second after the fall, is of the Wi-Fi CSI tool, the beacon’s transmission frequency is characterized by minimal changes and average values in both typically set at 10 Hz, consequently, the sampling rate for the the accelerometer and gyroscope.

CSI data captured from beacon frames is also 10 Hz. As mentioned earlier, the whole dataset comprises ten CSI data is collected from two receiving antennas on our distinct actions, each split into a training set and a testing mobile device, with each antenna gathering information from set in a 7:3 ratio. Each input sample to the model is a 53 subcarriers.

The data for each subcarrier is further divided one-dimensional array containing 20 features, i.e., maximum, into the real and imaginary components of the CSI. mean, median, kurtosis, and variance of the first and second Thus, the corresponding dimensions of the raw CSI data are part of accelerometer and gyroscope data. 30×212. The 30 rows indicate that, due to the 10 Hz sampling The input is first fed into a fully connected layer, where each rate, a total of 30 CSI values are captured over 3 seconds, and feature is connected to 128 neurons and processed through a the 212 columns correspond to the real and imaginary parts of ReLU activation function, resulting in a one-dimensional array the CSI data captured, summing up to 212 components from of 128 values.

This is followed by the first Dropout layer, 106 subcarriers across two antennas. which randomly sets the weights of input neurons to zero, From the collected CSI data, both the real and imaginary with a dropout rate of 10% to prevent overfitting. The process parts can be used to calculate the amplitude and phase of the

AUTHOR et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (MAY 2024) 7

TABLE II: Composition of continuously collected CSI data

Single sampling Real part Imaginary part Sample time First antenna Second antenna First antenna Second antenna t 0-53 0-53 0-53 0-53

TABLE III: Parameters of MLP model used for processing TABLE IV: Parameters of CNN model used for processing CSI IMU data. data. Layer Output Shape Param Layer Output Shape Param Dense (None, 128) 2688 Conv1D (None, 27, 32) 10,208 Dropout (None, 128) 0 MaxPooling1D (None, 13, 32) 0 Dense (None, 64) 8256 Attention Layer (None, 32) 32 Dropout (None, 64) 0 Dropout (None, 32) 0 Dense (None, 32) 2080 Flatten (None, 32) 0 Dense (None, 11) 363 Dense (None, 64) 2,112 Dense (None, 2) 130

CSI data. CSI provides granular information about the signal

during transmission, including multipath effects and phase wavelet denoising can effectively remove high-frequency noise changes. Multipath effects refer to the phenomenon where Wi- while preserving these characteristics, as shown in Fig. 10, it Fi signals reach the receiver through multiple paths due to allows for a clearer analysis of the underlying signal patterns, different reflections, refractions, and diffractions encountered which are crucial for interpreting the data accurately. along the propagation path.

These signals from various paths The second stage of fall detection is invoked after a fall cause phase superposition at the receiver, leading to changes is detected in the first phase. The purpose is to eliminate in the signal phase, which can be described as follow: instances where throwing a phone is mistakenly identified as a fall. It also distinguishes between falls from which one N X can recover independently and those from which one cannot. s(t) = Ai ej(ωt+ϕi ) (7) Therefore, following a fall detected by the IMU, the CSI i=1 dataset for this phase is futher divided into two categories: where s(t) is the received signal at time t, N is the number prolonged stillness and continued activity.

Consequently, our of paths the signal has taken, Ai is the amplitude of the signal CSI dataset comprises two types of movements: relatively from the ith path, ω is the angular frequency of the signal, ϕi stationary and those with significant amplitude, with 820 is the phase shift introduced in the signal from the ith path instances of the former and 800 of the latter collected. due to path length differences, j is the imaginary unit.

In These movements correspond to significantly different rates CSI data, multipath effects manifest as rapid phase changes, of change in CSI data, the model employs just a single making it challenging to extract phase information from CSI. layer of one-dimensional convolution followed by a fully This difficulty makes it hard to correlate CSI phase data with connected layer, with a softmax activation function in the the relative stillness and movement actions in the second phase output layer.

The input data dimension for the one-dimensional of fall detection. What’s more, in wireless communication, convolutional layer is (29, 106), utilizing 32 convolutional the Doppler effect manifests as a shift in the phase and kernels of size 3, and the output dimension equals the num- amplitude of the received signal, attributable to the relative ber of convolutional kernels, which is followed by a one- motion between the transmitter and receiver.

This phenomenon dimensional max pooling layer with a window size of 2. An is quantified as follow: attention mechanism layer is then added. During its use, the v model focuses more on important parts of the input sequence, ∆f = f0 cos(θ) (8) enhancing its ability to learn key information.

A dropout layer c follows with a 20% dropout rate to regularize the model and where ∆f represents the Doppler shift, v denotes the velocity prevent overfitting during training. Then comes a flattened of the observer relative to the source, c is the speed of light, layer, which transforms the multi-dimensional output into a f0 refers to the original transmitted frequency, and θ is the one-dimensional array. This is followed by a fully connected angle of movement relative to the line of sight.

In the realm layer with 64 neurons and ReLU activation, outputting a of CSI, these principles facilitate the detection of motion by dimension of 64. Finally, another fully connected layer is evaluating phase shifts over time, calculated as: added for task classification, using the softmax activation 2π∆f Ts function. The overall structure and parameters of the model ∆ϕ = (9) N are outlined in Table IV. where Ts denotes the symbol duration and N the number of subcarriers.

III. E VALUATION Hence, we select the amplitude of CSI as the feature. The In order to better compare the classification effect, we in- phase and amplitude waveform along with the Doppler shift troduce three evaluation indexes, i.e., accuracy rate, precision of CSI is shown in Fig. 9.

The amplitude waveform of the CSI rate and recall rate. For the precision rate, we consider two data contains numerous transient and sudden features. Discrete metrics: TP and FP.

8 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2024

1.0

0.8

0.6 Rate

0.4

0.2 MLP RandomForest 0.0 Accuracy Precision Recall

Fig. 11: Recognition rate of MLP and Random Forest models.

 3HUFHQWDJH

 (a) MLP 3UHFLVLRQ  5HFDOO OO

\

S 6LW

W WLF

ON

7R ZQ

+R UG WDO XD 3OD

N8

)D

:D

ZD

6WD

RQ

6T

R 3LF

G'

UL]

DQ 8S

Fig. 12: Precision and recall of MLP model on actual data

collected on ten different types of actions.

A. Performance Analysis of the IMU Model

Fig. 11 shows the performance of the dataset on MLP and random forest models. The MLP model built with TensorFlow and the random forest model using the sklearn library on the test set is quite similar across various metrics. The overall accuracy rates for all actions reached 96.31% and 93.89%, respectively.

To facilitate deployment on Android applications, the MLP was (b) Random Forest ultimately selected as the classifier to distinguish between falls and other daily activities, achieving a precision of 98.67% and Fig. 13: Confusion metrics of MLP and RF models on the a recall of 90.28% on fall actions. test set. To evaluate the accuracy of our IMU model, data on speci- fied action types were collected and analyzed. Analysis results, shown in Fig. 12, indicate that while the model generally accelerometer readings collected by the phone are very likely distinguishes falls from routine activities effectively, it may not to be zero.

For instance, after a fall, there might be no capture all fall events with full accuracy. High precision paired further activity, resulting in zero accelerometer data. Similarly, with lower recall for critical actions like “Fall” suggests that when a phone is quickly put down on a table, it likely the model can effectively reduces false positives but may miss remains stationary on the table afterward, also resulting in true positive events, potentially compromising its effectiveness zero accelerometer readings.

The same applies when a phone in real-world applications such as elderly fall detection. Lower is thrown onto a bed or sofa, the phone generally remains recall rates imply that the current model threshold lacks motionless thereafter. Therefore, accelerometer and gyroscope sufficient precision for detecting subtler movements essential data become ineffective for further analysis under these con- to reliable fall detection. ditions.

Fig. 13 reveals that if a fall action is misclassified, it is Consequently, subsequent judgments have introduced CSI generally mistaken for actions such as picking up or putting for wireless sensing, i.e., using WiFi signals to detect whether down a phone, or throwing the phone onto a sofa or bed. there is any movement in the environment, rather than relying Analyzing the actions themselves, these three types of actions solely on the phone’s built-in sensors.

This serves dual pur- exhibit similar changes in accelerometer and gyroscope data. poses. Primarily, it enhances the discrimination of erroneously When using these data for the same feature extraction, the ex- perceived actions that mimic falls, such as the act of throwing tracted features are also quite similar, hence misclassification a mobile phone. Typically, following the action of throwing is inevitable. a mobile device onto a surface such as a bed or sofa, an Moreover, following such misclassifications, the subsequent individual is unlikely to remain stationary.

Thus, subsequent

AUTHOR et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (MAY 2024) 9

ROC of amplitude of CSI

0 50 100 150 200 Sampling point

Fig. 14: The ROC of CSI amplitude indoors during periods

Fig. 16: Confusion metrics of CSI models. with or without human activity.

the CSI data as the person attempts to stand.

B. Performance Analysis of the CSI Model

Fig. 15 illustrates the training and validation performance Accuracy

Loss of our CSI model. The training accuracy swiftly ascends and

Loss

stabilizes near a value of 1.0, indicative of the model’s robust ability to classify training data correctly. Concurrently, the validation accuracy mirrors this trend, suggesting a consistent and generalizable performance across unseen data, without significant overfitting, as the curves for both training and validation accuracy closely align and stabilize at high values. Training epochs Training epochs Training epochs Loss Trainingmetrics epochs for both training and validation initially display a steep decline, leveling off to indicate minimal error as the (a) Training accuracy (b) Loss value model converges.

The closeness of the training and validation loss curves further substantiates the absence of overfitting, Fig. 15: The accuracy and loss metrics of CSI model. reinforcing the model’s effective generalization capabilities. Exhibiting outstanding convergence behavior, the CSI model consistently achieves high accuracy and maintains low movements can induce significant disturbances in the spatial loss across both training and validation sets.

This performance channel state, thereby causing notable fluctuations in the underscores its effectiveness and potential for real-world ap- collected CSI data. Conversely, in instances where a fall results plications, with minimal risk of overfitting. in an inability to self-rescue, such as fainting, there is minimal Fig. 16 shows that in a stationary state, the values of each subsequent movement, which results in a stable channel state, subcarrier may not be fixed, but as long as the position of the leading to minimal variation in the CSI data.

This stability phone remains unchanged and there is no human activity in is crucial in further reducing the probability of misjudgment. the space, the variation rate of each subcarrier remains very Fig. 14 shows the rate of change (ROC) of CSI amplitude low, typically below 50. However, human movement within indoors during periods with or without human activity, it can the space can cause changes in the amplitude of the CSI data, be seen that the first half of the period exhibits no activity, as leading to a variation rate much higher than when relatively indicated by the stable CSI waveform.

In contrast, the second static. In the test set, these two types of actions achieved a half shows activity, during which the amplitude of the CSI 99% accuracy rate on the CSI model designed. undergoes violent fluctuations. Our system achieves an accuracy that is 1.28% greater than Secondly, the observation of changes in CSI data post- that of the ANN reported in [28] when utilizing spectrogram incident enables the determination of whether a fall requiring features.

It also surpasses the fall detection system proposed immediate intervention has occurred. If a fall does occur, in [29] by 3.64%, and is 3.14% more accurate than the subsequent CSI data changes can be observed to determine research [30] conducted by Felipe et al. whether the fall is one that cannot be self-rescued. If it Furthermore, Our system not only enhances accuracy but is a non-self-rescuable fall, the subsequent CSI data will also resolves key issues such as reducing false positives and remain stable.

However, if it is a self-rescuable fall, significant maintaining user compliance without extra hardware, stream- disturbances will appear in the CSI data as the person gets up. lining deployment and increasing user adherence [31], [32]. In short, our system distinguishes between actual falls and similar movements by observing subsequent movements. True Above all, it is evident that our CSI model effectively distin- falls, especially those that cannot be self-rescued result in min- guishes between static and dynamic states.

The experimental imal movement, which is reflected in stable CSI data. On the results demonstrate that the IMU model can account for the contrary, self-rescuable falls lead to significant disturbances in majority of fall incidents. Consequently, by augmenting the

10 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2024

600 100.0 Instantaneous current value(mA) 550 99.5 500 99.0

Accuracy (%)

98.5 98.0 Not_trigger_CSI 97.5 Trigger_CSI 97.0 0 50 100 150 200 250 300 Left Pocket Right Pocket Carrying Bag Jacket Pocket Hand Time(s) Placement

Fig. 17: Power consumption comparison of whether to con- Fig. 18: The accuracy of the system in detecting falling tinuously trigger the CSI model. movements at five different locations: left pocket, right pocket, carring bag, jacket pocket, hand.

IMU model with the CSI model, it is possible to provide

warnings and call for assistance by assessing whether an the sensitivity of fall detection and reduce the probability of individual retains the ability to move after a fall. misclassifying other actions as falls, the real-time detection It is pertinent to mention that the exclusive use of the CSI results are stored in a buffer, which has a capacity of 20. model is not advisable primarily due to substantially higher Once the buffer is full, the oldest result is discarded each time power consumption.

Fig. 17 presents a comparison of power a new result is entered. The system then continuously monitors consumption when the CSI model is activated continuously, the number of falls among these 20 results in the buffer. In illustrating the inefficiencies of its standalone use.

The graph practical tests, setting the detection criteria to identify a fall shows the trigger condition for CSI activation (depicted in red), if 3 out of 20 results indicate a fall provided the best overall maintaining consistently higher values over time compared performance. to scenarios where the CSI model is not triggered (dotted Fig. 19(a) demonstrates the prompt for the IMU model to grey line). This clear difference in power usage highlights the call the CSI model after detecting a fall.

Fig. 19(b) depicts a necessity of integrating more power-efficient models such as real fall after which the person manages to get up, classified the IMU to enhance overall system efficiency. as a fall followed by an activity, indicating a self-rescuable Considering the potential impact of varying phone place- fall. Comparing Fig. 19(c) and Fig.19(d), in Fig. 19(c) there ments on model accuracy, as shown in Fig. 18, we conducted is little to no movement after the fall, which is confirmed by verification experiments to ascertain that environmental factors the accelerometer data visualization indicating no movement minimally affect the overall system detection capabilities.

This post-fall, classifying it as a non-self-rescuable fall, which is is due to the reliance on accelerometers and gyroscopes within the type of fall behavior our system aims to detect. the device to complete its functions. Furthermore, given that In contrast, Fig. 19(d) results from a misclassification the primary objective of CSI is to detect motion changes in caused by throwing the phone onto a sofa, whose accelerom- the environment, the presence of such changes ensures that eter data shows resulted in no subsequent movement of the our system will consistently generate a result.

On average, our phone. However, human activity still occurs in the space, system for fall detection achieved an accuracy rate of 99.1% causing significant fluctuations in subsequent CSI data, leading across five distinct directions. to the determination that this fall detection was a misjudgment of another action. This helps address the first stage IMU model’s shortcoming of insufficient accuracy in distinguishing C.

Deployment and Testing of the APP falls from fall-like actions. The fall detection APP is based on the Android 13 system and utilizes version 33 of the Android SDK to provide applica- IV. C ONCLUSION tion interfaces and libraries for development.

The source code This study presents an innovative real-time fall detection is set to be compatible with Java 8, meaning it is compiled system, utilizing built-in hardware, i.e., accelerometers and according to the syntax rules of Java 8. The Gradle build gyroscope sensors, and combines them with commercial WiFi tool is used to manage the project’s building, dependencies, on Android smartphones. We review the research background version control, resource management, and other tasks.

The and existing technologies addressing the challenges within fall IMU and CSI models deployed on the Android system have detection systems, drawing on studies from both domestic and been converted into the TensorFlow lite format. international contexts. A sensor-based method is explored in The fall detection process is monitored with a step size detail, involving the collection of data from accelerometers of 0.1 seconds.

During the entire sliding window process, and gyroscopes during movement. This approach proposes a approximately 5-10 fall detections are likely to occur. How- nuanced process for handling raw data of fall events tailored ever, for other actions similar to falling, such as picking up to different stages of a fall.

By doing so, it enhances the or placing a phone, it is rare for the detection to register 5- feature extraction from the fall process and introduces a sensor 10 falls throughout the sliding detection process. To enhance orientation-independent fall detection model.

AUTHOR et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (MAY 2024) 11

Remind Remind Warning Remind

The IMU detects that a fall has CSI has detected that you are You have fallen and are unable CSI has detected that you are occurred, after which the CSI will still moving after a fall. Do you to move. Help is being called! still moving after a fall.

Do you intervene for further verification. need assistance?. nihao need assistance?. Confirm Cancel Confirm Cancel Cancel Confirm

(a) IMU detected as a fall (b) self-rescuable fall (c) non-self-rescuable fall (d) misjudgment Fig. 19: Results of the fall detection by the app.

One of the intrinsic limitations of the smartphone-based fall of a fall as the center of balance is located in the cere- detection system presented in this study is its dependency bellum. Subsequent studies could investigate more effective on the device’s battery life. Since continuous monitoring is devices for monitoring head movements during falls, thereby essential for ensuring the safety of individuals, particularly improving the accuracy of fall detection.

Additionally, to the elderly, battery depletion could compromise the system’s enhance the model’s generalizability, statistical features such effectiveness. To address this concern, we propose the integra- as mean, variance, and kurtosis are extracted from the raw tion of an automated low-battery alert within the application. data. Nevertheless, these features may not fully encapsulate the This feature would actively monitor the device’s battery status complexity of the raw data, prompting a need for refined data and notify the user with an audible alarm or visual prompt processing techniques and feature extraction methodologies to when the charge level falls below a predetermined threshold, more effectively represent the pertinent information.

Further- urging them to recharge the device promptly. more, the existing CSI model incorporates data from all sub- The analysis identifies actions that are commonly misclas- carriers, resulting in extensive data volumes. Future research sified by sensor-based fall detection systems and their draw- could focus on selectively using subcarriers based on real-time backs. We further propose the use of CSI as a supplementary channel conditions to streamline data processing, potentially criterion in the model, effectively resolving the issue of the reducing computational demands and energy consumption. inability to collect sensor data when the phone is stationary, What is worth mentioning is that the system’s effectiveness thus enhancing the reliability and completeness of the entire depends on the assumption that users consistently carry their fall detection system.

To sum up, this article describes the smartphones and maintain a WiFi connection, conditions that deployment of IMU and CSI models within an Android-based may not be reliably met, particularly for elderly individuals fall detection App, detailing the system’s capability to collect who may not habitually carry their phones at home. This data, process it, and utilize both models in real-time fall limitation is even more pronounced for users with cognitive detection. impairments, such as dementia, who may forget to carry their phones.

V. D ISCUSSION AND F UTURE W ORK Recognizing this limitation, future works could involve sensor fusion with alternative sensing modalities. While our Leveraging the distinctive capabilities of sensors and CSI, primary focus in this study is on exploring the feasibility and we develop a novel real-time, mobile-based fall detection accuracy of smartphone-based technology, we acknowledge system, achieving preliminary results and laying a foundation that the inclusion of complementary sensors could enhance for further investigation.

Despite these advances, the project system reliability. Sensor fusion with wearable or ambient sen- is constrained by limited theoretical insights and research sors, for instance, could alleviate the dependency on consistent capacity, presenting numerous avenues for enhancement. smartphone use and provide a more comprehensive monitoring For instance, the current methodology employs a smart- solution. In the future, researchers could pay more attention phone positioned in a pocket to gather data pertinent to fall on the integration of these complementary sensing modalities detection, which may not accurately capture the dynamics to create a hybrid fall detection system.

12 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2024

Addressing these areas could substantially elevate the sys- [20] A. Sowmya and A. S.

Pillai, “Human fall detection with wearable tem’s efficacy and efficiency, rendering fall detection more sensors using ML algorithms,” in 2021 2nd International Conference on Smart Electronics and Communication (ICOSEC). IEEE, 2021, pp. viable and precise for real-world implementation. 1092–1095. [21] E. Ramanujam and S.

Padmavathi, “A vision-based posture monitoring system for the elderly using intelligent fall detection technique,” Guide R EFERENCES to Ambient Intelligence in the IoT Environment: Principles, Technologies and Applications, pp. 249–269, 2019. [1] “Falls,” 2021. [Online]. Available: https://www.who.int/news-room/ [22] A. A.

Khan, B. B. Johnson, and K.

Snyder, “Contactless fall detection fact-sheets/detail/falls with radio frequency sensors for the aging population,” Alzheimer’s & [2] N. Thakur and C. Y.

Han, “Country-specific interests towards fall detec- Dementia, vol. 19, p. e078644, 2023. tion from 2004–2021: An open access dataset and research questions,” [23] P. M. Lo and A.

Abd Aziz, “A FPGA threshold-based fall detection Data, vol. 6, no. 8, p. 92, 2021. algorithm for elderly fall monitoring with verilog,” Bulletin of Electrical [3] H. Koda, Y. Kai, S.

Murata, H. Osugi, K. Anami, T.

Fukumoto, and Engineering and Informatics, vol. 10, no. 5, pp. 2477–2487, 2021. H. Imagita, “Relationship between muscle strength asymmetry and body [24] W.

Li, D. Zhang, Y. Li, Z.

Wu, J. Chen, D. Zhang, Y.

Hu, Q. Sun, and sway in older adults,” Journal of Aging and Physical Activity, vol. 26, Y. Chen, “Real-time fall detection using mmwave radar,” in ICASSP no. 3, pp. 457–461, 2018. 2022 - 2022 IEEE International Conference on Acoustics, Speech and [4] R.

Khanmohammadi, S. Talebian, M. R.

Hadian, G. Olyaei, and Signal Processing (ICASSP), 2022, pp. 16–20. H.

Bagheri, “Time and frequency domain analysis of gait initiation [25] M. M. Baig, S.

Afifi, H. GholamHosseini, and F. Mirza, “A systematic in younger and older adults,” Journal of aging and physical activity, review of wearable sensors and IoT-based monitoring applications for vol. 25, no. 2, pp. 212–217, 2017. older adults–a focus on ageing population and independent living,” [5] W.

Ludwig, K.-H. Wolf, C. Duwenkamp, N.

Gusew, N. Hellrung, Journal of medical systems, vol. 43, pp. 1–11, 2019. M.

Marschollek, M. Wagner, and R. Haux, “Health-enabling technolo- [26] D.

Naranjo-Hernandez, L. M. Roa, J.

Reina-Tosina, and M. A. Estudillo- gies for the elderly–an overview of services based on a literature review,” Valderrama, “Personalization and adaptation to the medium and context Computer methods and programs in biomedicine, vol. 106, no. 2, pp. in a fall detection system,” IEEE transactions on information technology 70–78, 2012. in biomedicine, vol. 16, no. 2, pp. 264–271, 2012. [27] S.

Denkovski, S. S. Khan, B.

Malamis, S. Y. Moon, B.

Ye, and [6] Y. Hu, F. Zhang, C.

Wu, B. Wang, and K. J.

R. Liu, “DeFall: A. Mihailidis, “Multi visual modality fall detection dataset,” IEEE Environment-independent passive fall detection using WiFi,” IEEE In- Access, vol. 10, pp. 106 422–106 435, 2022. ternet of Things Journal, vol. 9, no. 11, pp. 8515–8530, 2022. [28] M.

Cheffena, “Fall detection using smartphone audio features,” IEEE [7] L. Ren and Y. Peng, “Research of fall detection and fall prevention Journal of Biomedical and Health Informatics, vol. 20, no. 4, pp. 1073– technologies: A systematic review,” IEEE Access, vol. 7, pp. 77 702– 1080, 2016. 77 722, 2019. [29] C.-L.

Lin, W.-C. Chiu, F.-H. Chen, Y.-H.

Ho, T.-C. Chu, and P.-H. Hsieh, [8] S.

Madansingh, T. A. Thrasher, C.

S. Layne, and B.-C. Lee, “Smartphone “Fall monitoring for the elderly using wearable inertial measurement based fall detection system,” in 2015 15th international conference on sensors on eyeglasses,” IEEE Sensors Letters, vol. 4, no. 6, pp. 1–4, control, automation and systems (ICCAS).

IEEE, 2015, pp. 370–374. 2020. [9] M. Tolkiehn, L. Atallah, B.

Lo, and G.-Z. Yang, “Direction sensitive fall [30] F. A.

S. Ferreira de Sousa, C. Escriba, E.

G. Avina Bravo, V. Brossa, J.- detection using a triaxial accelerometer and a barometric pressure sen- Y.

Fourniols, and C. Rossi, “Wearable pre-impact fall detection system sor,” in 2011 Annual international conference of the IEEE engineering based on 3d accelerometer and subject’s height,” IEEE Sensors Journal, in medicine and biology society. IEEE, 2011, pp. 369–372. vol. 22, no. 2, pp. 1738–1745, 2022. [10] I.

Basu, S. Ray, S. Sarkar, S.

Saha, and S. Panda, “Advancements in [31] Y. Li, K.

C. Ho, and M. Popescu, “A microphone array system for au- fall detection technologies: A comprehensive review and innovative so- tomatic fall detection,” IEEE Transactions on Biomedical Engineering, lutions,” in 2023 7th International Conference on Electronics, Materials vol. 59, no. 5, pp. 1291–1301, 2012.

Engineering & Nano-Technology (IEMENTech). IEEE, 2023, pp. 1–3. [32] S. Mobsite, N.

Alaoui, M. Boulmalf, and M. Ghogho, “A privacy- [11] T.

Nakamura, M. Bouazizi, K. Yamamoto, and T.

Ohtsuki, “Wi-Fi-CSI- preserving AIoT framework for fall detection and classification using based fall detection by spectrogram analysis with CNN,” in GLOBE- hierarchical learning with multilevel feature fusion,” IEEE Internet of COM 2020-2020 IEEE Global Communications Conference. IEEE, Things Journal, vol. 11, no. 15, pp. 26 531–26 547, 2024. 2020, pp. 1–6. [12] T. Z.

Chowdhury, “Using Wi-Fi channel state information (CSI) for human activity recognition and fall detection,” Ph.D. dissertation, Uni- versity of British Columbia, 2018. [13] R. Ramezani, Y. Xiao, and A.

Naeim, “Sensing-Fi: Wi-fi CSI and accelerometer fusion system for fall detection,” in 2018 IEEE EMBS International Conference on Biomedical & Health Informatics (BHI). IEEE, 2018, pp. 402–405. [14] O. D.

Lara and M. A. Labrador, “A survey on human activity recognition using wearable sensors,” IEEE communications surveys & tutorials, vol. 15, no. 3, pp. 1192–1209, 2012. [15] Q.

Zhang and M. Karunanithi, “Feasibility of unobstrusive ambient sensors for fall detections in home environment,” in 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2016, pp. 566–569. [16] F. X.

Gaya-Morey, C. Manresa-Yee, and J. M.

Buades-Rubio, “Deep learning for computer vision based activity recognition and fall detection of the elderly: a systematic review,” Applied Intelligence, vol. 54, no. 19, pp. 8982–9007, 2024. [17] S. Cagnoni, G. Matrella, M.

Mordonini, F. Sassi, and L. Ascari, “Sensor fusion-oriented fall detection for assistive technologies applications,” in 2009 Ninth International Conference on Intelligent Systems Design and Applications, 2009, pp. 673–678. [18] M.

Krovvidi, “Activity detection using rf and sensor data fusion,” Ph.D. dissertation, Carnegie Mellon University, 5 2024. [Online]. Available: https://kilthub.cmu.edu/articles/thesis/ Activity Detection Using RF and Sensor Data Fusion/25796344 [19] H. Ponce, L.

Martinez-Villasenor, and J. Nunez-Martinez, “Sensor location analysis and minimal deployment for fall detection system,” IEEE Access, vol. 8, pp. 166 678–166 691, 2020.

IMU Model

Get data

Attention

Dropout .

Sofa Sofa

y 20 20

10 10

Acc value

Acc value

x 0 0

-10 -10

C. Stage I

1.0

0.8

0.6 Rate

0.4

\

S 6LW

W WLF

ON

7R ZQ

N8

)D

:D

ZD

6WD

RQ

6T

R 3LF

G'

UL]

DQ 8S

Loss

Accuracy (%)