1 Introduction

Walking is the most common human physical activity; especially for children or young people, it is a natural daily activity. However, for various reasons, elderly people may have difficulty walking. The gait of the elderly with gait problems is not only clearly different from that of healthy people but also easily observed with the naked eye. However, it may be difficult to identify gait abnormalities because there is no clear medical standard or definition of normal gait. Gait abnormalities in the elderly are accompanied by a risk of fracture or brain damage due to a fall, limiting their independent activities. Furthermore, they may have to rely on their family members or other caregivers for assistance.

One of the methods traditionally used to assess gait abnormalities is a direct examination of a person’s gait. The most representative way to evaluate gait abnormalities in medical fields is to evaluate gait with a method called “Timed up and go test.” In this method, a subject stands up from a chair, walks 3 m, turns around, and returns. An examiner records the time it took to sit back down on the chair and observes the gait. There are a few problems with this test. First, there is a high probability that the subjective opinion of the examiner will be involved. Second, there is no objective data for determining gait abnormality, and finally, there is no clear standard for concluding that there is a problem with the gait.

Researchers have been considering a variety of gait analysis methods that would solve the problem of subjective human observation in determining gait abnormalities. In a study of gait analysis, normal and abnormal gait were differentiated using an optical motion capture system in a controlled experimental environment. After attaching a number of markers to specific parts of the subject’s body, the subjects walked a certain distance. After capturing the subject’s gait using several infrared cameras installed in the laboratory, the collected gait data were analyzed to determine gait abnormality. The advantage of this method is that it shows high accuracy in determining gait abnormalities. However, considerable effort is required to collect and analyze human gait data this way. In addition, a disadvantage of laboratory environment is that it is not realistic because it is far removed from the daily environment in which humans walk. Moreover, markers are attached to test subjects for experimentation, which may cause inconvenience in walking, hindering the collection of natural gait data.

According to a recent report on healthcare wearable devices, their market share in the healthcare sector is expected to increase rapidly in the future. Based on the recent statistics on senior citizens released by the National Statistical Office, there are 70,700,000 people aged 65 or older, accounting for 13.8% of the total population. It is estimated that by 2060, the proportion will reach approximately 40% of the total population, and the growing proportion of older adults is a global trend. Interest in the use of smart health devices is increasing across all age groups. Furthermore, owing to the recent interest in the Internet of Things (IoT) and the rapid development of wearable sensor devices, studies conducting gait analysis using inertial sensors, such as three-axial accelerometers, gyroscopes, or insole pressure sensors, are actively underway. The use of sensors is widely seen as an alternative to address the shortcomings of human observation and the optical motion capture system mentioned above. Wearable sensors have low power consumption, are easy to carry, and do not require a controlled experimental environment. The exclusion of the subjective opinion of the experimenter may also be an advantage of using wearable sensors in gait analysis study.

The purpose of this study is to develop an algorithm that can measure unnatural human gait using wearable devices. To achieve this goal, the research is divided into three parts: investigation and analysis of the latest gait inspection methods, gait data collection and feature extraction using IoT wearable devices, and unnatural gait measurement algorithm development using an autocorrelation coefficient.

This paper is organized as follows. First, studies that measured gait abnormalities are summarized in Section 2, and specific methods for distinguishing natural and unnatural gait using an autocorrelation coefficient are described in Section 3. Next, Section 4 describes in detail the test subjects and experimental methods for the gait analysis and the experimental results of the gait analysis algorithm using the autocorrelation coefficient devised in this study. Finally, the research results are summarized and future research directions are presented in Section 5.

2 Related works

Unnatural gait is very common in the elderly: 63% of the elderly in nursing homes have a gait disorder, and 8–19% of those who are not hospitalized in a medical institution have a similar problem [22]. Figure 1 shows mediolateral sway (internal and external axis) and stride velocity (stride width) using wearable sensors (body-attached sensors) targeting healthy community-based groups of all ages. In Fig. 1, it can be seen that as the person ages, the regression of the gait function becomes clear [11]. Gait abnormalities in the elderly are related to diseases and act as a limiting factor in their quality of life. Therefore, the field of geriatric medicine should measure gait abnormalities in the elderly, identify the causes of the abnormal gait, and provide pharmaceutical or interventional treatment accordingly.

Fig. 1
figure 1

(A) mediolateral sway (ML) and (B) stride velocity measured using wearable sensors. The black dotted lines are linear regression, and the dark solid lines represent the data values determined by a non-parametric regression analysis [11]

For example, a study of dementia among the elderly indicated that one of the common antecedents of dementia were gait abnormalities, such as a change in the shape of the gait or slowing down, and the incidence of dementia actually increases [23]. Therefore, gait analysis can help predict future dementia. The graphs in Fig. 2 show the gaits of the healthy elderly and dementia patients, and it is clear that the two graphs are different [30].

Fig. 2
figure 2

Gait sequences of healthy person (upper graph) and dementia patients (lower graph) [30]

Recent gait analysis studies have assumed that human gait is typically a series of repetitive patterns. Therefore, after determining the patterns within the gait cycle, efforts have been made to observe and analyze temporal characteristics according to the patterns [5, 12, 26, 28]. The human gait cycle begins by lifting one foot away from the ground and moving forward; this is called toe off (TO). The next step occurs when the foot touches the ground; this is called heel strike (HS). In the case of TO, force to move forward is applied, so the speed increases. In contrast, in the case of HS, the speed decreases as the foot touches the ground. As soon as one foot touches the ground, the other foot performs TO. These two states occur in succession. One gait cycle starts with the TO state of one foot and ends with the next TO state of the same foot. Acceleration sensors are accepted as suitable for representing the change in speed in TO and HS; thus, many researchers are actively using accelerometers for gait analysis.

A peak detection algorithm was applied to detect specific events occurring within the gait cycle [20, 25]. In another study, hidden Markov models were adapted [14]. In these studies, the researchers calculated gait features related to time, such as swing time, double stance time, and stance phase time, and used them for gait analysis as gait features or characteristics. Swing time is defined as the time from TO to HS of one foot, double stance time refers to the time when both feet are on the ground, and the stance phase time is the time from HS to TO of one foot. According to gait analysis studies, obtaining the abovementioned gait characteristics using inertial sensors produces relatively accurate results.

Walking speed has also been studied as an important gait feature of humans. For people over 65 years of age, walking speed can be seen as a distinct physical change due to decreased physical capacity [21, 24, 27, 33]. Therefore, calculating walking speed accurately is a goal of many researchers in the gait analysis field. Studies investigating walking speed applied methods of human gait model [13], integration calculation [3], or machine learning approach [7, 8, 10, 31].

Another approach to extracting gait characteristics is to use the ground reaction forces as gait features measured by insole sensors [1, 6, 17, 18]. Before insole sensors were used, the ground reaction force generated each time the subject walked on a force plate was calculated. However, this method is expensive, and it is difficult to analyze gait if an experimental environment is not provided. The main disadvantage is that it is difficult to obtain information on natural gait in daily life. On the other hand, if the insole sensor is used, walking data can be obtained from the sensor by attaching it to the test subject’s shoe.

In gait analysis studies, gyro sensors tend to be used in addition to the acceleration and insole sensors. The inertial sensors are attached to the subject’s body to acquire gait data, and their positions vary from study to study. The most commonly used positions are waist, thighs, shanks, and feet. The number of sensors used in the experiments varies from one to six. The primarily measured gait features are temporal characteristics, such as stride length, stride time, swing time, stance time, cadence, step time, and step length. In addition, the pressure values from the insole sensor and the angle values from the gyro sensor were also extracted and used.

In most gait studies, data related to specific events within the gait cycle are used. A study was conducted with a pressure sensor to analyze the gait motion that occurs between toe off and heel strike, an acceleration sensor to measure the acceleration component, and a gyro sensor to determine the rotational speed [9]. In another study, four inertial measurement units were attached to the thigh and shank area of each leg. The joint angular trajectory and gait cycle during walking were obtained to analyze the gait [32]. As shown in these studies, to accurately obtain the characteristics of walking, at least three to four sensors were used to record various data. However, this is inconvenient, especially for the elderly or patients with gait abnormalities, and must be resolved for commercialization. To address this problem, we conducted research on gait abnormality measurement with one acceleration sensor. An acceleration sensor was chosen because it is often used in combination with user information for various behavior recognition tasks, such as calculating the amount of activity or finding out a life pattern.

The most recent research trend on gait pattern analysis is mainly applying machine learning or deep learning techniques [2, 4, 15, 16, 19, 29]. Beyond the conventional method that directly designed the gait analysis algorithm to obtain the desired result by detecting the gait cycle from the raw data collected from the sensor and analyzing the characteristics of the gait cycle for each individual or group, efforts to obtain meaningful results by using a large amount of walking data to learn through various machine learning and deep learning algorithms are in full swing. Accelerometers, gyro sensors, and insole sensors used in existing gait analysis studies are still used for gait data collection and analysis in machine learning oriented gait research. The main gait features obtained in these studies using machine learning techniques show different patterns depending on the final results to be obtained in the study. In addition, it can be seen that there are various types of machine learning and deep learning algorithms used in each study. Gait analysis using machine learning is expected to receive more and more researchers’ interest in the near future.

3 Proposed method

3.1 Definition of natural and unnatural walking

Walking refers to the process of moving the body forward while maintaining the center of the body. The start of the gait cycle is defined as the point at which the heel strikes the ground (initial contact in Fig. 3). During the first part (approximately 0% to 60%) of the gait cycle, one foot is touching the ground (the right foot in Fig. 3). Then, one foot is kept in contact with the ground while the other foot moves to move the upper body forward. In this process, the other foot is in the air. Finally, the foot touches the ground, and the foot that was previously supporting the ground is no longer in contact with the ground (initial swing in Fig. 3). When one foot first touches and then falls off the ground, it is called the stance phase. According to Fig. 3, the right foot first touched the ground. Therefore, the right foot is initially in the stance phase. The next step is called the swing phase because the foot that has fallen off the ground remains in the air. At this stage, the body moves forward, and acceleration occurs. Finally, the foot that was in the air lands on the ground, causing deceleration. In this way, both feet go through the stance phase and swing phase once periodically and continuously, and this process is called the gait cycle.

Fig. 3
figure 3

Human gait cycle (Figure from: Stöckel, Tino & Jacksteit, Robert & Behrens, Martin & Skripitz, Ralf & Bader, Rainer & Mau-Moeller, Anett. (2015). The mental representation of the human gait in young and older adults. Frontiers in Psychology. 6. 943. https://doi.org/10.3389/fpsyg.2015.00943)

From the definition of the gait cycle, we can infer that people who walk naturally will have a repetitive and constant gait pattern created by the left and right feet in the stance phase and swing phase. Furthermore, it is conceivable that a certain pattern of acceleration and deceleration, which occurs as the body moves, especially in the swing phase, will be created. Based on this, natural walking refers to a gait in which the medial side of the leg is in a straight line while maintaining the proper stride length in an upright position. On the other hand, unnatural walking is out of the range of natural walking and occurs in various combinations, such as decrease in walking speed, decrease in stride length, or foot drag. For example, gait abnormalities caused by arthritis or muscle disease break the center of gait by focusing only in one direction depending on the location of the damaged area. It also acts as a cause of disrupting the gait cycle defined earlier. Figures 4 and 5 show typical raw 3-axis acceleration values of two different human subjects. In these pictures, Accx stands for the acceleration of a sideway direction (left and right), Accy represents the one of an up-down motion, and Accz indicates that of forward and backward movement.

Fig. 4
figure 4

Raw acceleration data of subject I

Fig. 5
figure 5

Raw acceleration data of subject II

In this paper, instead of using the terms normal and abnormal gait, which has been commonly used in gait analysis studies, the terms corresponding to our study were newly defined. Walking without any problems was defined as natural walking (NW). The gait contrary to natural walking was defined as unnatural walking (UW). The reason for defining a new term for this study is that in the case of NW, it is not significantly different from the normal walking we talked about earlier, but in the case of UW, it is a walking method that we artificially devised for this study, thus avoiding confusion for researchers. UW will be described in detail in the next section.

3.2 Reasons for using autocorrelation coefficients

The correlation coefficient is defined as a numerical value for the correlation between two variables x and y. Since the correlation coefficient was proposed by Karl Pearson, it is also referred to as Pearson’s correlation coefficient. In our study, walking data of participants were collected while they walked a given distance for a certain amount of time. The walking data were collected using an acceleration sensor according to a predetermined sampling rate. Therefore, the walking data are defined as the value of the 3-axis accelerometer data collected according to the change in time. The formula for the correlation coefficient according to the sample of walking data collected using the acceleration sensor is as follows:

$$ {r}_{xy}=\frac{\sum_{i=1}^n\left({x}_i-\overline{x}\right)\left({y}_i-\overline{y}\right)}{\sqrt{\sum_{i=1}^n{\left({x}_i-\overline{x}\right)}^2}\sqrt{\sum_{i=1}^n{\left({y}_i-\overline{y}\right)}^2}}, $$



is a total sample size in a given experimental time

xi, yi:

are the individual acceleration samples of gait at time t

$$ \overline{x}=\frac{\sum_{i=1}^n{x}_i}{n} $$
$$ \overline{y}=\frac{\sum_{i=1}^n{y}_i}{n}. $$

According to Eq. (1), two variables x and y are required. In other words, the above formula is necessary when investigating the correlation between two distinct variables x and y, but the desired correlation in this study should be based on the correlation between the same acceleration values ​​over time, rather than between two different variables. Therefore, it is difficult to apply the general correlation formula presented above to our study. Unlike general correlation coefficients, autocorrelation coefficients are suitable for elucidating correlations between the same variables over time. In this study, the correlation between the acceleration values from the left and right feet is investigated to determine whether walking is natural or artificial. Thus, it was judged that it is more appropriate to use an autocorrelation coefficient, which can obtain a correlation between the same data (acceleration values in our research) rather than a general correlation coefficient that identifies the correlation between different variables x and y.

The correlation between the value \( {x}_{t_1} \)at time t1 and the value \( {x}_{t_2} \) at time t2 is called autocorrelation with xt as an arbitrary process. The autocorrelation function defines this relationship. The value obtained through it is the autocorrelation coefficient, which indicates the degree of autocorrelation between different data, and is expressed as a value between −1 and 1. If this value is close to 1, the two data points are very close with the same characteristics. Otherwise, there is no meaningful relationship between two data points.

As a result, in this study, a gait analysis was conducted using the autocorrelation coefficient defined in Eq. (2) without applying the general correlation coefficient presented in (1). The autocorrelation coefficient rh applied in this study is defined as follows:

$$ {r}_h=\frac{\sum_{t=h+1}^n\left({x}_t-\overline{x}\right)\left({x}_{t-h}-\overline{x}\right)}{\sum_{t=1}^n{\left({x}_t-\overline{x}\right)}^2},0\le h\le n-1, $$



is a total sample size in a given experimental time.


is the individual acceleration sample of gait at time t.

\( \overline{x} \):

the average of x1, x2, …, xn .

The autocorrelation function is effective in calculating whether data values at a specific time are related to each other or not when compared to previous data values. In the formula given in (2), the autocorrelation coefficient rh is calculated by the autocorrelation function over time. This function returns a value between −1 and 1 for a given input sample. Therefore, the autocorrelation coefficient for a specific sample h is calculated by formula (2), and we use this autocorrelation coefficient value to distinguish between NW and UW.

3.3 Algorithm for determining natural and unnatural gaits using autocorrelation coefficients

The proposed algorithm consists of five steps. In the first step, the values from the acceleration sensor built into the smartphone are collected using a mobile application created by us for this study. People participating in the experiment first run the mobile application to check whether the acceleration sensor value is displayed properly. Then they put the smartphone in the right pocket of their pants and start the experiment according to the instructions of the experimenter. Details related to gait data collection are described in Section 4.

The second step is to refine the gait data collected in the first step. In most cases, the initial acceleration value contains noise due to various environmental factors. Therefore, it is necessary to remove this noise. In this study, it was removed using the low pass filter (LPF) algorithm. LPF is applied as shown in Fig. 6 to remove abnormally high or low values among the 3-axis acceleration sensor data. In particular, Fig. 6 only shows the acceleration value of a forward and backward direction (Z-axis). As shown in the graph, we move on to the next step after removing the noise. Figure 6 shows the noise removal results for the z-axis, not all acceleration sensor data on the x, y, and z axes. The reason for this is that in Section 3.1, the act of walking is explained as an activity that makes the whole body move forward with the force that occurs when two legs alternately touch the ground. The x, y, and z axes of the 3-axis acceleration sensor are each determined in a direction for generating acceleration. In particular, the z-axis represents acceleration produced through movement in the forward and backward directions. Therefore, the acceleration created in the z-axis can best represent the walking behavior of a person. For this reason, in this study, the gait analysis was only conducted using the acceleration value from the z-axis.

Fig. 6
figure 6

Low pass filter (LPF) applied to acceleration data

After the filtering process, the gait data of all test subjects were normalized to a value between 0 and 1. The reason for this is that the ranges of acceleration values ​​of different subjects are diverse, as shown in Figs. 10 and 11. By performing normalization, it is possible to analyze the gait data within a common range. In addition, to compare the data, they were normalized by substituting the data distribution with a value between 0 and 1.

After normalization is performed, the next step applies the autocorrelation coefficient to the normalized gait data to differentiate NW and UW. As a criterion for determining natural and unnatural gait, the autocorrelation coefficients of the left and right footsteps were calculated, and if they were above a certain threshold, it was determined as NW. If they were lower than the threshold value, the walking was determined to be UW. To this end, data were distributed for each step of the left and right feet, and after removing the data from the front and back, which measured unnecessary actions, 800 data samples of 8 s were used. To determine the optimal amount of data per one step, the accuracy was obtained by extracting the results for 1 to 200 data samples for one step.

For instance, if the sampling rate of an acceleration sensor is 50 ms, it can be possible to get 20 data samples per second. We identified the most appropriate number of data samples to represent a single step for all of the participants in the experiment. Figure 7 shows that, as a result of analyzing the gait of all participants in 200 data samples, when the experiment was conducted by defining 67 data samples as one step, the accuracy of distinguishing between NW and UW was the highest. Based on this result, the original data collected from the experimenters’ walking was analyzed by grouping 67 data samples into one step.

Fig. 7
figure 7

F1-score based on the number of data samples

The final step was to determine the accuracy of the proposed algorithm. All gait analysis processes were terminated with measuring the accuracy using the F1-score method. Figure 8 depicts the process of gait analysis by processing gait data in a CSV format, and Fig. 9. shows the pseudocode of our proposed algorithm. We discuss the estimation of the accuracy using the F1-score method in Section 4.

Fig. 8
figure 8

Gait analysis procedure

Fig. 9
figure 9

Proposed gait detection algorithm

4 Experiment

4.1 Experimental settings

Gait data were collected from 25 healthy men and women in their twenties without trauma or disease. They each performed NW and UW five times. We conducted experiments with the approval of the Institutional Review Board (IRB, GWNUIRB-2020-33) for human experiments, and explained all possible risks to the participants. Each participant signed an informed consent form approved by the IRB. The experiment was conducted by installing a splint approximately 5 cm high on the sole of the participant’s foot for UW. By doing this, the participants in the experiment had no choice but to walk unnaturally and differently from their NW. When the subjects walk in this manner, the balance of their body will be unnatural, and accordingly, it is expected that they will be slower than their usual pace. As a result, it was predicted that the values recorded by the acceleration sensor would be smaller than those recorded during NW. To collect NW data, the participants were instructed to walk comfortably at their usual walking speed. Data were collected while the participants walked on a flat surface at their usual walking pace. As can be seen in Figs. 10 and 11, the acceleration value measured by the acceleration sensor in the case of NW was greater than that in the case of UW. Figures 10 and 11 show the NW and UW of two different participants. During NW, there are no restrictions, so the acceleration value reflects the speed of one’s walking. On the other hand, during UW, with a splint on the sole of the foot, normal acceleration is not produced because the walking is unnatural. Therefore, as shown in the figure, the acceleration value is lower than during NW. In the initial setting, all the participants wore tight jeans, sneakers, and the smartphone was placed in the right jean pocket. The experiment was conducted in an indoor corridor, and the walking distance was 10 m. The smartphone application collected acceleration data at 100 Hz. To measure the natural gait, the participant walked the 10 m distance five times with the smartphone application running. To measure unnatural gait, approximately 5-cm splint was attached to the sole of the shoe by wrapping a compression bandage. Then, the smartphone application was running while the participants walked the 10 m a total of five times.

Fig. 10
figure 10

Natural (a) and unnatural(b) raw acceleration data of subject A

Fig. 11
figure 11

Natural (a) and unnatural (b) raw acceleration data of subject B

4.2 Gait data collection

4.2.1 Mobile applications

Walking time and values ​​in the x (left-right), y (up-down), and z (forward-backward) directions of the 3-axis acceleration sensor built into the smartphone were saved together. The smartphone used in this study was a Samsung Galaxy Note 9 (SM-N960F), and the walking data were collected using the acceleration sensor built into this smartphone. For data collection, we made our own mobile application using Android Studio. Figure 12 shows the running screen of our application designed to collect the participants’ walking data in real time. The collected data were saved using Google’s Firebase real-time database so that they were saved whenever the sensor value changed. In addition, Firebase authentication was used to authenticate data identity. Therefore, the UID of the participants was obtained according to the authentication request, and their data values corresponding to the child node are stored. Here, the UID is a value that is used as the user’s unique identifier every time the application is installed, so that the subject corresponding to the data can be identified. However, this method required the participants to install or subscribe to the application. Therefore, as shown in Fig. 12, the experiment step was simplified by changing the method to receive an ID from each participant before collecting their data.

Fig. 12
figure 12

Values of 3-axis acceleration from the acceleration sensor in the smartphone as shown in the application

4.2.2 Data collection and analysis

The process for collecting and analyzing gait data consisted of four steps. First, by using the self-produced mobile application mentioned in Section 4.2.1, the walking data of the participants were collected in real time. Second, the collected walking data were automatically stored in the Firebase database in Jason format. Third, we extracted the data stored in the Jason format in Firebase for gait data analysis. Finally, by using the API provided by the Pandas library, the gait data format in Jason format is changed to the CSV format. The gait analysis algorithm is applied to the gait data in CSV format to obtain the autocorrelation coefficient mentioned in Section 3 and the data are classified into normal gait and abnormal gait. The gait analysis is completed by calculating the accuracy of the classified result using F1-score. Figure 13 shows the configuration of the system and the procedure for data collection.

Fig. 13
figure 13

System organization for data collection and analysis

4.3 Experimental results

4.3.1 F1-score

F1-score was used to measure the accuracy of the algorithm for distinguishing between normal and abnormal walking. For example, suppose there are a total of 100 gait samples. Of these, 50 samples are of normal walking and 50 for abnormal walking. If we provide a total of 100 samples as input values to our proposed algorithm, and if the algorithm classifies all 50 normal walking samples as normal walking and all 50 abnormal walking samples as abnormal walking, the accuracy of the algorithm is 100%. F1-score is calculated using the formula in (3). In Eq. (3), precision is the ratio of what is actually classified as true in the model, and recall is the ratio of what the model predicts as true among what is actually true. Precision and recall are shown in Eqs. (4) and (5). In the formulas of precision and recall, TP stands for “true positive,” FP for “false positive,” and FN for “false negative.Footnote 1

$$ F1- SCORE=\frac{2}{recall^{-1}+{precision}^{-1}} $$
$$ precison=\frac{TP}{TP+ FP} $$
$$ recall=\frac{TP}{TP+ FN} $$

4.3.2 Accuracy calculation

To calculate the accuracy of the algorithm proposed in this study, the accuracy was computed by dividing one step to up to 200 points of data. Using 67 points showed the highest accuracy at 61.5% (Fig. 7). Therefore, based on a total of 800 data samples, 67 were grouped into left and right foot data, and the autocorrelation coefficient between the data divided into left and right feet was calculated and used as input of the algorithm. The same process was performed for normal and abnormal walking data. Through this process, the classification of walking determined by the proposed algorithm was calculated. The next step was to use the F1-score method to calculate the accuracy between the actual value and the results of the algorithm. Table 1 shows the results of this calculation as a confusion matrix, and Fig.14 shows the pie chart results for accuracy.

Table 1 Confusion matrix
Fig. 14
figure 14

Accuracy of the proposed algorithm

When normal walking data are provided as input to the algorithm, the probability of recognizing normal walking is 80% and the probability of determining abnormal walking is 20% (Fig. 14A). When abnormal data are entered into the algorithm, the probability of determining abnormal walking is 60%, and the probability of determining normal walking is 40% (Fig. 14B). The probability that the proposed algorithm judges normal walking as normal is 80%, whereas the probability of determining abnormal walking as abnormal is 60%, which is relatively low accuracy. The reason for the low accuracy in determining abnormal walking is that there may be various environmental and other factors that we did not predict in advance. However, we learned through experiments that both normal and abnormal walking show a certain pattern. In addition, especially in the case of abnormal walking, due to the splint attached to only one foot, there may be an element that prevents the other foot from being able to walk normally. If so, it is possible to speculate that the accuracy in determining the abnormal walking was relatively low because a periodic walking pattern was preserved. This could cause the autocorrelation coefficient to be above the threshold, even if the pattern was different than the pattern in normal walking.

5 Conclusion

Walking is the most common and repetitive physical activity that a person performs in daily life, and human gait contains much information. Various attempts have been made to develop technologies that can be used in the healthcare field by using ultra-small sensor devices equipped with advanced information technologies, such as IoT, big data, cloud computing, and artificial intelligence. The era has arrived in which information on human gait can be obtained without using an experimental environment equipped with expensive imaging equipment and special facilities. If sensor devices are used effectively, it is expected that in the future, instead of administering treatments after the disease occurs, the disease would be predicted in advance and the disease would be prevented before it becomes serious. Three conditions need to be satisfied to achieve this goal. First, micro-sensor devices usually operate at low power and have limited computing capacity. Therefore, gait analysis should be possible while consuming minimal power with a small amount of calculation. Second, it is necessary to collect walking data using a minimum number of sensors so that there is no inconvenience in daily life. Lastly, the results of gait analysis should be accurate and reflect the current state of health. In this study, we created an algorithm to discriminate between normal and abnormal walking using the autocorrelation coefficient. Smartphones used by ordinary people every day were used, and walking data were collected using the acceleration sensor built into the smartphone. The gait analysis consisted of three steps (filtering, normalization, and calculation of autocorrelation coefficient), and the proposed algorithm is designed to obtain accurate results with the least possible computation. F1-score was used to calculate the accuracy of the algorithm. Normal walking data were identified with 80% accuracy and abnormal walking data with 60% accuracy. As mentioned in the experimental results section, in the case of abnormal walking, the foot with a splint prevented normal walking, affecting the walking of the other foot, resulting in relatively low accuracy. Compared with other gait analysis studies, our study tried to distinguish between normal and abnormal walking, the most basic task for gait analysis, with a relatively simple methodology. However, a clear definition of normal and abnormal gait must be established as an important task to be preceded first in the gait research field.

As a future research task in gait analysis, it is meaningful to conduct a study to discriminate between normal and abnormal gait using machine learning or deep learning techniques, which are currently receiving much attention. Similar to other algorithms or programs, artificial intelligence techniques require clean input data. In other words, the extent to which the input data reflect the phenomenon is the key to success. In addition, research to extract features that well represent the features of human walking should also be conducted.