Real-Time Forward Head Posture Detection and Correction System Utilizing an Inertial Measurement Unit Sensor

Abstract

Forward head posture (FHP) has become a prevailing health issue in modern society as people spend more time on computers and smartphones. FHP is a posture where the head is forward and the anterior and posterior curvatures of the lower cervical and upper thoracic spines are both, respectively, exaggerated. FHP is often associated with neck pain, bad static balance, and hunched shoulders or back. To prevent this, consciously maintaining good posture is important. Therefore, in this study, we propose a system that gives users real-time, accurate information about their neck posture, and it also encourages them to maintain a good posture. This inexpensive system utilizes a single inertial measurement unit sensor and a Raspberry Pi system to detect the changes in state that can progress to an FHP. It retrieves data from the sensor attached to the user’s cervical spine to indicate their real-time posture. In a real-world office environment experiment with ten male participants, the system accurately detected the transition to the FHP state for more than 10 s, with a delay of less than 0.5 s, and it also provided personalized feedback to encourage them to maintain good posture. All ten participants recognized that their average craniovertebral angle had to be increased after receiving visual alerts regarding their poor postures in real time. The results indicate that the system has potential for widespread applications.

1. Introduction

In the last decade, the number of Internet users has grown significantly worldwide, increasing from ∼2.7 billion in 2012 to 5.3 billion in 2022 [1]. In addition, as modern society accesses the Internet via computers and smartphones, the average time spent on such electronic devices has also experienced a correlating increase. When a user uses a computer or smartphone for a long time, the user’s head bends forward, which results in a bad posture that induces physical health issues such as neck pain. Overtime, this may progress to the forward head posture (FHP) condition, in which the anterior and posterior curvatures of the lower cervical and upper thoracic spines, respectively, are exaggerated to maintain balance.

FHP is the most common cervical postural defect found in almost the entire human population regardless of race or age [2]. It is strongly associated with chronic neck pain in modern populations [3], and it also negatively impacts static balance control [2,4]. According to Dieleman et al. [5], out of 154 diseases, those related to neck pain had the highest healthcare expenditures in 2016 for the United States alone at approximately USD 134.5 billion. According to Chen et al. [6], the annualized prevalence of neck pain among white collar workers ranged from 42% to 63%, while that of neck diseases only ranged from 17% to 21%. Furthermore, studies have shown that the weight exerted on the spine increases drastically when the head is bent forward due to FHP, where ∼49 lbs of weight is exerted on the spine due to the forward 45$°$ head bend, which was indicated in a person in [7]. Therefore, the habitualization of bad posture adversely affects the spine, and it is also highly associated with round shoulders, thoracic kyphosis, and turtle neck [8].

To prevent or treat bad postures that can progress to FHP, it is important for people to consciously maintain good posture. Therefore, recognizing one’s neck condition in real time in daily life is important [9]. In modern times, the use of computers and smartphones has become a daily routine as people are continually immersed in digital content. However, because it is difficult for people to recognize their bad posture, they are unable to make a conscious effort to maintain a good posture for a long time. Particularly, bad posture has been reported to cause body pain in adolescents who sit and study for a long time, as well as in office workers who use computers in their occupation, thus making both daily life and study difficult [10].

In this study, we propose a system for real-time posture detection in daily life using an inertial measurement unit (IMU) sensor attached to the human body, which is used to detect bad posture and induce a correct posture for the prevention of diseases caused by FHP. The contributions of this paper are as follows:

• The identification of optimal location on the body to attach a wearable IMU sensor to accurately detect FHP status.
• The providing of an intuitive visual interface to enables users to easily recognize a bad posture and correct their posture.
• The proposal of a personalized system that can visually display the results of real-time (with a delay of less than 0.5 s) posture–data acquisition and processing, as well as an improvement in the effectiveness of practical correction and prevention of bad postures in daily life.

This paper is organized as follows: Section 2 introduces the studies related to posture judgment and FHP correction methods; Section 3 describes the proposed system structure and algorithm; Section 4 discusses the location selection for sensor attachment for the purposes of detecting and correcting posture when using a gyro acceleration sensor, as well as details the conducted experiments and result analyses; and Section 5 presents the conclusions of this study.

2. Related Works

The research on determining an individual’s posture has resulted in the suggestion of methods for analyzing captured images of people with sensors.

Lee et al. [11] placed a kinetic camera on a monitor to capture a human participant’s frontal view, wherein the depth image analysis was used to measure the distances between the camera and the forehead, as well as between the camera and the torso. The difference between these two distances was then used to determine the FHP status. Meanwhile, Matuska et al. [12] used a smartphone to take the frontal image of an individual, and the neck tilt was measured based on the tilt of the smartphone and the tilt of the face (which was obtained by analyzing the frontal image). Recently, some researchers have attempted to classify posture using machine learning. Lawanont et al. [13] took a picture of a person with a smartphone, and they then used the tilt of the smartphone and the slope of the face obtained through image analysis for measuring the tilt of the neck. Kulikajevas et al. [14] cut a video recording of a person’s sitting posture into frames to collect images, and they then used a MobileNetV2-based model to classify the sitting posture into three categories (straight, backward, and forward), wherein an accuracy of 91.5% was achieved. Chen [15] utilized Openpose [16] to extract body joints and lines connecting joints, while a convolutional neural network (CNN)-based model was used to distinguish between good and bad postures, and it achieved a 90% accuracy. However, when analyzing posture through camera images, classification or detection accuracy varies greatly depending on the camera position, and considerable training data are required to accordingly determine human posture given a particular human body structure. In addition, there is a risk of privacy violation because the image analysis process exposes either the front or side view of an individual.

Meanwhile, studies that utilize sensor data to determine posture are as follows. In a study by Matuska et al. [12], sensors were attached to chairs and cushions to determine the sitting posture, and participants were required to sit on a chair with six pressure sensors to measure the force applied to each sensor. Based on this, a system was proposed that classifies the sitting posture and notifies the individual whether the posture is good or bad via a mobile application. Meanwhile, in the study conducted by Jeong et al. [17] for determining the sitting posture using machine learning, six pressure sensors were placed on a chair cushion and six distance sensors were attached to the backrest to collect data while a k-nearest neighbors algorithm was used to determine the participant’s sitting posture. Hu et al. [18] attached six flex sensors to a chair and used an artificial neural network to determine the sitting posture. However, in the aforementioned studies, the number of sensors required to predict the sitting posture was high, thus making the approach costly. The indirect location of sensors on the chair cushion or backrest limited the sensors’ ability to accurately distinguish between normal and FHP states.

Because of these disadvantages, studies have reported the direct attachment of wearable sensors to the human body to determine the user’s posture for effective posture correction [19]. For instance, Tlili et al. [20] attached three IMU sensors to a participant’s back and shoulders in the form of a belt; therefore, a system that can determine the user’s back posture by detecting body tilt via sensors was created; in addition, accordingly, sound was also used by the system for the purposes of an alarm. Meanwhile, Radhakrishnan et al. [21] attached an IMU sensor to the ear of a participant to obtain gyroscope values, as well as for determining whether the user was in an FHP state. Gyroscopes use the Coriolis force, the inertial force of a rotating object, to measure angular velocity, and they also integrate it to obtain the neck tilt [22]. However, owing to the nature of gyroscopic sensors, errors can occur [23], and these errors become large when they are integrated, thereby causing angular drifts [24]. Therefore, accurately determining the degree of tilt is difficult using the gyroscope data obtained from an IMU sensor. To resolve this, in this paper, we used gravity acceleration data instead of gyroscopic data to accurately determine the degree of neck tilt.

The existing studies on correcting FHP conditions are as follows: Mohamed et al. [25] used corrective exercises with biofeedback, and—after four weeks of corrective exercises—the average craniovertebral angle (CVA) of 35 participants increased by ∼4$°$ (a low CVA angle reflects FHP); An et al. [26] used four weeks of neurodevelopmental therapy and cervical spine exercises to increase the mean CVA of 12 participants by 4$°$; and Suwaidi et al. [27] used Chiropractic BioPhysics [28], a stretching method, to increase the average CVA of 33 participants by more than 10$°$ over three months. While corrective exercises and stretches are effective, these are usually difficult to perform independently without professional help. Additionally, exercises require sustained commitment and considerable time to physically correct bad posture. However, previous studies have proven that wearable sensors can help improve posture in everyday life to easily achieve similar results as are obtained with specialized exercises and in less time.

3. FHP Detection and Correction Using an IMU Sensor

In this study, the proposed system for FHP detection and correction requires a user to simply attach one IMU sensor to their body for gravity acceleration–value collection. Then, the sensor determines the user’s FHP condition and alerts the user to consciously correct the FHP condition by asking to them change their posture. This section describes the proposed method for detecting FHP, as well as the process for the measured data to determine FHP.

3.1. Identifying the Criteria of FHP

A good metric for accurate FHP determination is measuring the CVA [29]. The CVA is the angle between the line connecting the tragus (center of the outer ear), C7 (seventh cervical vertebra) (Figure 1a), and the horizontal line passing through C7 [30,31]. Lee et al. [32] measured the CVA using the side photographs of 30 healthy individuals without cervical spine abnormalities, wherein a result of −53.2 ± 5.85$°$ was obtained. However, the CVA varies from person to person. Therefore, a clear reference for CVA has not been reported in the literature to determine FHP [33]. However, as shown in the actual CVA measurement presented in Figure 1b,c, the higher the CVA, the better the posture.

Other methods of measuring the CVA include using an inclinometer [34] or cervical range of motion (CROM) [35,36]. The inclinometer presented in Figure 2a can be tilted to create an incline equal to the imaginary line that connects the tragus and C7 to measure the CVA. When the CROM presented in Figure 2b is worn on the head, it displays the angle at which the head is tilted. However, because instrumental methods require such equipment, it is difficult to measure the real-time CVA for daily postures. Further, when taking side photos, CVA estimations vary depending on the position of the camera, and—even if the position of the camera is fixed—the angle at which the photo is taken will vary owing to the different heights of individuals.

Using the relationship between the CVA and sensing data collected by an IMU sensor, this study proposes a system to determine CVA with the sensing data retrieved by only a singular IMU sensor, which is then used to help prevent and treat FHP. Moreover, in this study, the average angle obtained from the user’s good posture in the first 10 s during the experiments is employed as the user’s characteristic standard angle of CVA ($baselineCVA$). Then, the FHP is determined based on this standard angle to build a personalized system that changes the standard angle according to an individual’s physical situation and build.

3.2. FHP Detection and Correction by Sensing Data Processing

Section 3.1 mentions that the $baselineCVA$ is different for each person; hence, it is difficult to find a clear standard for determining FHP. Therefore, at the start of the experiment, the participant was instructed to maintain a good posture for >∼10 s. In the meantime, this reference angle was set as their $baselineCVA$ using the average of the measured angles. Subsequently, as shown in Figure 3, the obtained sensing values received were continuously processed by transiting between Algorithm#1 and Algorithm#2, and the information regarding the current state was provided to participant.

Setting the standard angle as the user’s initial good posture after the start of the experiment starts will trigger Algorithm#1. If Algorithm#1 determines that it is a bad posture, then Algorithm#2 operates instead. If Algorithm#2 determines that the participant is out of a bad posture, then the system returns to Algorithm#1.

Figure 4 is an algorithm that checks if current users have bad posture (FHP).

According to Yip et al. [37], people with neck pain have an average CVA of approximately 5$°$ lower than those without neck pain. Therefore, neck pain is related to FHP [3]. The value obtained by subtracting 5$°$ from the CVA corresponding to the good posture of a user (standard CVA) was set as the $baselineCVA$ in Figure 4 and Figure 5; according to $Threshold1$ in Figure 4, it was determined whether the participant was in FHP when their wrong posture persisted for more than a few seconds. In this study, $Threshold1$ was set to 20. Because the time until a new angle was received was every 0.5 s, Algorithm#1 checks if the posture is wrong, i.e., if the posture is likely to be FHP, for more than 10 s. If Algorithm#1 recognizes it as a bad posture, Algorithm#2 works to induce the user to correct their posture on their own. Algorithm#2 determines the posture state by calculating how different the recently obtained CVA value is from previous ones. Finally, Algorithm#2 induces the user to escape the bad posture state.

As shown in Figure 5, $Threshold2$ is obtained by adding one to the average sensing values for 10 s before the new sensing value enters the immediately preceding sensing values. If the new angle is less than $Threshold2$, it is classified as $Case1$ and remarks that posture has not improved.

If the entry angle is greater than the $baselineCVA$ and less than $Threshold2$, it is classified as $Case2$ and that the posture is improving but not sufficiently enough. However, if it is greater than the $baselineCVA$, it is classified as $Case3$ and is considered to be out of the bad posture state.

3.3. Real-Time Posture Visualization for Bad Posture Corrections

To help correct bad posture, the current CVA and state of the neck are displayed on a human-shaped picture, as shown in Figure 6, thus allowing users to know their current condition more intuitively.

Users can see their current state through their posture images (Figure 6). On the right side of Figure 6, the CVA and $baselineCVA$ obtained in real time are displayed numerically. The algorithmically determined CVAs can be displayed on the screen of the system to prompt correction if the posture is bad.

4. Experiments and Analysis

After an experiment for determining the ideal IMU sensor location for FHP detection, the participants were asked to type with and without system feedback in front of a computer screen similar to that found in a real-world office environment. This was conducted to observe any difference between the two cases when the FHP detection and correction system process are applied.

4.1. Sensing Location Determination of the IMU Sensor

As it is an IMU sensor, the MPU6050 (TDK Electronics, TDK Worldwide, Tokyo, Japan) [38,39] can read gravity acceleration values along x, y, and z-axes, as well as three-axis gyroscope values. Figure 7a shows the three-dimensional axes and rotation of MPU6050. As shown in Figure 7b, the MPU6050 sensor is cost-effective, very lightweight, and small, making it a great fit for a wearable sensor in daily life. Among the sensing values provided by MPU6050—using the x, y, and z acceleration values—the current forward tilted angle ($\theta$) of the sensor can be obtained through Equation (1), as shown in Figure 7c [40]. In Equation (1), x, y, and z are the gravity acceleration values along each of the axes of the MPU6050.

$$forwardtiltedangle\left(\theta \right)=arctan(\frac{x}{\sqrt{y^2+z^2}}).$$

(1)

To verify the accuracy of the sensor, we tilted it close to certain angles (0$°$, 30$°$, 45$°$, 60$°$, and 90$°$), as shown in Figure 8, and we checked if the forward tilt angle ($\theta$) obtained from the sensor was similar to the angle of the sensor that is measured when taking a photo from the side.

As shown in

Table 1. Comparison of the forward tilted angle ($\theta$) with the actual angle.

	Figure 8a	Figure 8b	Figure 8c	Figure 8d	Figure 8e
Forward tilted angle ($\theta$)	87.8$°$	60.2$°$	44.3$°$	28.3$°$	3.0$°$
Angle measured from the side	87.2$°$	60.6$°$	44.7$°$	28.5$°$	3.1$°$

, the small difference (<0.5$°$) between the forward tilted angle ($\theta$) and actual angle sensing data can be useful for our approach.

First, the IMU sensor was attached to several parts of the upper body (as shown in Figure 9), while the waist and neck were kept straight, which is the standard for good posture (as shown in Figure 1b). Subsequently, the participant was instructed to gradually extend their neck forward to reach FHP. The CVA of the standard postures (standard CVA) and FHP states differed by up to 20$°$, wherein the optimal position was determined by testing various attachment positions to verify if the forward tilt angle ($\theta$) could be measured to represent the CVA value.

All 10 male participants were asked to adopt the same posture in the same experimental environment, and their CVAs were measured by taking a photo from the side. The information regarding the 10 male participants is shown in

Table 2. Participant information.

	Age (Years)	Height (cm)	FHP History	Time Spent on Cell Phone and Computer (per Day)
Person#1	24	178	X	7
Person#2	24	178	X	9
Person#3	22	176	X	9
Person#4	25	184	X	9
Person#5	23	170	X	8
Person#6	25	183	X	10
Person#7	23	185	X	10
Person#8	24	182	X	7
Person#9	24	170	X	9
Person#10	24	172	X	11

. In accordance with the guidelines detailed in the Helsinki Declaration [41], the experiment was conducted after obtaining the written consent of all participants. All participants were normal people with no history of FHP, and their use time of cellphones and computers is shown in Table 2. $\theta$. The CVA values shown in

Table 3. Relation between the forward tilted angle ($\theta$) and CVA according to sensor location.

Sensor Location	[1] Forehead	[2] Chest	[3] Back of Neck	[4] Cervical Spine
Forward tilted angle ($\theta$) of FHP ($°$)	−4.6 ± 2.0	−12.9 ± 4.7	−35.1 ± 2.6	−59.2 ± 1.9
Forward tilted angle ($\theta$) of the standard posture ($°$)	−18.5 ± 2.9	−23.1 ± 1.8	−16.3 ± 6.6	−40.0 ± 4.6
CVA of the FHP ($°$)	30.5 ± 4.0
CVA of the standard posture ($°$)	49.9 ± 5.2

are the average values obtained from adopting different sensor attachment positions for the 10 participants, and the mean and standard deviations of these values are also provided.

When sensors are placed at the [3] back of the neck and [4] cervical spine, the difference in $\theta$ between the FHP and the standard posture is ∼20$°$, which is approximately the same as the difference in CVA between them. However, when the sensor is placed on the [2] chest, the difference in $\theta$ is only 10$°$, while the CVA changes by 20$°$. In this case, it is difficult to predict CVA with the sensing values; hence, the [2] chest position was excluded.

Second, to reduce the number of FHP false positives, the same experiment was performed for a slightly tilted head posture, as shown in Figure 10. This is a common posture that people assume when looking at the bottom of a monitor or desk while working on a computer. In this case, although the CVA decreased slightly, it is not a FHP condition. The participants were then asked to demonstrate the posture shown in Figure 10, and their average $\theta$ were summarized, as shown in

Table 4. Relationship between the forward tilted angle ($\theta$) and CVA with a slight head tilting.

Sensor Location	[1] Forehead	[3] Back of Neck	[4] Cervical Spine
Forward tilted angle ($\theta$) of the pose with head slightly tilted ($°$)	−6.7	−32.1	−45.3
Forward tilted angle ($\theta$) of the standard posture ($°$)	−18.5	−16.3	−40.0
CVA of the pose with head slightly tilted ($°$)	44.0
CVA of the standard posture ($°$)	49.9

.

Because the results presented in Table 4 do not show the FHP, the CVA hardly changed; however, only when a sensor was attached to the [3] back of the neck or [1] forehead did the measured forward tilted angle ($\theta$) become similar to the $\theta$ in the FHP state. That is, if a sensor is attached to the [1] forehead or [3] back of the neck, a state that is not FHP may be mistakenly judged as FHP.

Based on these findings, we decided to attach the MPU6050 sensor to the cervical spine, and then the $\theta$ and CVA of the sensor changed linearly. Table 3 shows that, in the standard posture, the average $\theta$ of the sensor was −40.0$°$ and the average CVA was 49.9$°$. The forward stretching of the neck reduced the $\theta$ and CVA of the sensor. In the FHP state, the average $\theta$ of the sensor was −59.2$°$ and the average CVA was 30.5$°$. Figure 11 shows this result graphically, where the CVA and $\theta$ were observed to decrease with the same slope. Therefore, this can be expressed using Equation (2).

$$CVA=Forwardtiltedangle\left(\theta \right)+{89}^{\circ }.$$

(2)

4.2. Experimental Environment

As shown in Figure 12, the information received from MPU6050 entered the Raspberry Pi through I2C communication [42] using a connected wire. Afterward, the Raspberry Pi was configured in such a way that it judged the posture via the IMU sensor using Algorithm#1 and Algorithm#2, which were inside the Raspberry Pi, to output the visual results on the monitor. The communication time cycle of the raw data received through I2C communication was very short, taking about 10 s to connect the Raspberry Pi to the MPU6050 to obtain 1000 acceleration values; therefore, the time required to retrieve one value was about 0.01 s. Noise was detected in the raw data. In this experiment, the time sampling method was used to reduce the noise effect. The sampled value was the average of the incoming values with a time interval of 0.5 s. In addition, if a received value was significantly different from the immediately preceding value, it was classified as a missing value to reduce the sampling error. As a result, the value was updated every 0.5 s. To observe the difference between the raw and sampled data, the MPU6050 was tilted for about 10 s and the sensor angles were changed. As a result, 1000 raw data values were obtained, which were quite noisy, as seen in Figure 13. After time sampling every 0.5 s, it can be seen that the sensor $\theta$ versus time graph was smoother and less noisy.

All subsequent experiments were conducted by attaching MPU6050 to the cervical spine, as shown in Figure 14.

Table 5. Standard CVA.

	Standard CVA ($°$)
Person#1	58
Person#2	48
Person#3	56
Person#4	50
Person#5	46
Person#6	58
Person#7	49
Person#8	45
Person#9	46
Person#10	43

shows the accepted standard CVA for ∼10 s that was obtained while maintaining a good posture by each participant.

In Table 5, the CVAs of Person#2 and Person#4 was more than 5$°$ lower than those of Person#1 and Person#3. If the $baselineCVA$ was set to 50$°$, the system was determined to be in the FHP state even though Person#2 and Person#4 were in good posture. Therefore, the $baselineCVA$ should be set differently for each person.

4.3. Experiment 1: The FHP Detection and Correction System

Figure 15 captures a portion of the real-time CVA of Person#2, where the horizontal axis represents time and the vertical axis represents the CVAs.

The standard CVA for Person#2 was 48$°$, as shown in Table 5. Thus, Algorithm#1 subtracted 5 from the average angle and defined 43$°$ as the $baselineCVA$ for Person#2. Changes in the CVA and system judgment over time are shown in Figure 15. Times corresponding to the region shaded with blue indicate that the subject had a good posture, while those shaded with red indicate that the subject had an FHP.

At about 250 s after the start of the experiment, the participant’s CVA drops below the baseline. However, because Algorithm#1 is based on a bad posture sustained for more than 10 s, it only judges this to be a bad posture after about 10 s and then gives the participant a visual reminder. Algorithm#2 is triggered when this happens, and it tries to bring the CVA back to the baseline. If satisfied, it considers this a good posture. At about 310 s after the start of the experiment, if the CVA falls below baseline again and lasts for more than 10 s, then Algorithm #1 again considers the participant to have poor posture. Figure 16 shows the participant’s state changes through the posture images.

In the following section, how and why $Case1$ and $Case2$ were distinguished will be explained with respect to Algorithm#2 using real experimental data.

Figure 17 and Figure 18 shows some of the real-time CVA data obtained from Person#7, who underwent experiments, and the corresponding posture images. Algorithm#1 could determine that the participant was in a bad posture. However, after several experiments, we decided to use $Threshold2$ to distinguish between cases where the user belatedly realizes that they are assuming a bad posture, or cases where they try to maintain a good posture ($Case1$), as well as cases where the user tries to correct the bad posture ($Case2$). $Threshold2$ is the average of the data received in the previous 10 s plus 1. Because our system is a personalized system, users can also set $Threshold2$ to be based on the average of the previous values for 10 s plus a number such as 0.5 or 2 instead of 1. Therefore, if the user maintains a bad posture, the CVA is unlikely to pass over $Threshold2$. Figure 17 shows a case where the CVA slowly increased from 510 s, crossed $Threshold2$ at 517 s, and then transitioned to $Case2$. However, the CVA did not become greater than the baseline, and—at 522 s—it fell below $Threshold2$ again, thus transitioning to $Case1$. From then until 535 s, Person#7’s CVA remained at 42; as such, the system continued to alert their poor posture. After receiving the alert, Person#7 corrected their posture and their CVA increased, eventually bringing the CVA above the threshold and out of the bad posture. $Case2$ was for alerting the user when the user corrected their posture and their CVA increased but was still below the baseline; this helped the user escape their bad posture state.

4.4. Experiment 2: Measuring the Effectiveness of the Correction System in Real Life

To show that the proposed system was as effective as an FHP correction system, we conducted the following experiments: two experiments with and without personalized feedback. In the first experiment, participants sat indoors facing a computer monitor similar to a real office environment, and they used the computer for web surfing, typing, etc. The experiment lasted 30 min, and the 10 participants (Table 2) sat in a chair in the room and initially used the computer with only the sensors attached, such that the subjects were not aware of their current CVA. In the second experiment, we used the proposed system to allow the subjects to see their CVA and posture in real time on the monitor. In the absence of personalized feedback, participants often became engrossed in the computer screen, thus decreasing their CVAs.

Table 6. Average CVA value without personalized feedback up to 3 min after starting the experiment.

	Average CVA Immediately after the Start of the Experiment ($°$)	Average of CVA Values Dropped Due to Screen Immersion ($°$)	Difference ($°$)
Person#1	53	48	5.0
Person#2	43.2	39.8	3.4
Person#3	53.4	46.8	6.6
Person#4	46.4	43.6	2.8
Person#5	43.8	39.2	4.6
Person#6	55.7	51.2	4.5
Person#7	47.0	41.0	6.0
Person#8	41.0	34.7	6.3
Person#9	40.2	38.8	1.4
Person#10	38.5	32.9	5.6

shows the average CVA of each participant up to 3 min after starting the experiment, as well as the drop in their CVAs due to being immersed in the screen.

Table 6 shows that the CVA of the participants varied from person to person when they were immersed in a screen, and it was shown that it can drop by up to 6 degrees or more from the start. This naturally seems to progress to an FHP posture, but the participants were unable to recognize it.

Furthermore, under the same conditions, we used the proposed method to let the participants know their CVAs in real time, and we identified the times at which they were in an FHP in the two experiments, which was then tabulated. The FHP time was calculated as the time required for Algorithm#2 to run an experiment, which lasted about 900 s. Additionally, the average CVA for each experiment is shown in

Table 7. Average CVA value in a 15 min experiment.

	Baseline CVA ($°$)	FHP Time without Feedback (s)	FHP Time with Feedback (s)	Average CVA without Feedback ($°$)	Average CVA with Feedback ($°$)
Person#1	53	660	230	51.3	53.7
Person#2	43	620	280	41.7	44.4
Person#3	51	450	0	50.7	54.5
Person#4	45	540	300	44.8	46.7
Person#5	41	340	160	43.7	44.1
Person#6	53	490	150	53.3	55.9
Person#7	44	550	110	43.5	46.9
Person#8	40	1394	504	37.8	41.1
Person#9	41	1336	714	40.1	42.2
Person#10	38	1438	287	33.4	40.4

.

Table 7 shows that, in the 15-min experiment, the FHP time in the presence of the feedback from our system varied from participant to participant. However, these values decreased by more than 5 min on average, and the average CVA increased by about 3$°$. Some of the participants even tried to maintain their initial good posture to avoid the alert; therefore, the system was considered effective for correcting bad posture. It is also a personalized system because the baseline for determining the FHP status is set differently for each person. If the $baselineCVA$ was adjusted only in accordance with Person#1 and thus all the CVAs below 53$°$ were considered an FHP, then Person#2, Person#4, and Person#6 would have been considered to be exhibiting an FHP even though they were not actually showing an FHP.

In

Table 8. Comparison between the proposed approach and those proposed in recent studies.

	Detecting FHP	Providing Personalization	Using Wearable Sensor	Correcting Posture
Lee et al. [11]	O	X	X	X
Chen [15]	X	O	X	X
Matuska et al. [12]	X	X	X	X
Hu et al. [18]	X	O	X	X
Tlili et al. [20]	X	X	O	X
Radhakrishnan et al. [21]	X	X	O	X
Mohamed [25]	X	O	X	O
An et al. [26]	X	O	X	O
Our system	O	O	O	O

, several of the approaches for FHP detection reported in the literature are compared and corrected with the proposed method. First, the feasibility of the FHP detection, which is the neck forward posture, was checked. Other approaches such as the one proposed by Radhakrishnan et al. [21], which used sensors on chairs, and the approach proposed by Hu et al. [18], which used the Openpose software, (v. 1.7.0) can detect incorrect posture but not the FHP. Subsequently, the feasibility of customizing the system by considering the different physical conditions of each person was also checked. Studies that used depth cameras [15] or IMU sensors [36] utilized fixed threshold values; hence, they cannot detect an FHP depending on the user’s environment or physical condition.

In addition, existing research [11,12,15,18,20,21] on posture determination has focused on detecting incorrect posture and has not been concerned with correcting it. Meanwhile, the research on correcting bad posture [26,27] has focused on correcting people who are already in the FHP state and not on detecting whether a person is or not in the FHP state. However, the proposed system detects the FHP condition in a personalized way using different thresholds for each person. The proposed system is convenient as it employs only one wearable sensor, and it uses feedback to increase the average CVA of participants by 3–5$°$. Participants in a correction system using exercise or therapy for more than four weeks [25,26] exhibited an average increase in CVA of ∼4$°$. Therefore, the correction effect of the proposed system is similar to that of a sustained commitment to exercise and therapy.

5. Conclusions

In this study, we proposed a system to detect and correct FHP conditions in real time. Experimentally, we identified the cervical spine as the best location for CVA measurements using only one IMU sensor. The sensor attached to the cervical spine obtains CVA data in real time during daily life. Subsequently, we used two algorithms in the proposed system to determine the FHP status in real time, where the system alerted the participants about their postures using a visual interface. Therefore, the users could recognize bad posture and correct it effectively. The results of the experiments that were obtained from an environment similar to a real-world office setting or work environment showed that FHP conditions for >10 s were detected satisfactorily with a delay of ≤0.5 s, and that a similar level of CVA increase as that obtained using exercise or therapy was also yielded in the FHP correction.