A Single Smart Ring for Monitoring 20 Kinds of Multi-Intensity Daily Activities––From Kitchen Work to Fierce Exercises

Abstract

Statistical analysis of human daily activities contributes to time planning and health management. It also helps people make healthcare judgments and disease predictions when combined with big data technology. However, traditional methods for recognition of human daily activities based on vision and multisensors have limitations due to poor portability and weak capability in multiple activities detection. Herein, the development of a wearable human activity recognition (HAR) smart ring capable of recognizing at least 20 multi-intensity activities, i.e., motions ranging from clean-the-desk to playing basketball, based on a microinertial measurement unit and a hierarchical decision algorithm is proposed. Users only need to wear the smart ring on the index finger of the right hand to accurately identify 20 common activities that are classified as light, vigorous, and fierce. A novel hierarchical decision algorithm using 70 features is proposed to improve the accuracy and speed of recognizing common human activities and provides a final recognition accuracy of 98.10% for 20 types of activities. This extremely portable, reliable, and high-accuracy HAR solution is a significant advancement in providing real-time quantitative data for personal lifestyle supervision, time planning, and healthcare management.

Introduction

Human activity recognition (HAR) technology is widely used in health care, smart home, security monitoring, and other fields because it can judge the current activity type, environment, and human behavior intention by analyzing various types of data generated during human activities.^{[ 1–3 ]} In particular, a series of common health problems faced by modern people due to unhealthy lifestyles can be effectively prevented through statistical analysis of human daily activities. However, the current mainstream study methods^{[ 4–7 ]} generally have problems (e.g., high equipment cost, large size, and limited types of recognizable activities). In this context, how to design a portable, highly accurate HAR system that can recognize various complex activities poses an important technical problem. In recent years, researchers have proposed various solutions to identify common human daily activities. The most mature and common technologies mainly include computer vision-based HAR^{[ 8 ]} and wearable sensor-based HAR.^{[ 9 ]}

Computer Vision-Based HAR

Computer vision-based HAR uses cameras to collect video or picture information of human daily activities for recognition. Some researchers have used distributed fixed cameras to monitor simple activities in daily life.^{[ 10–12 ]} For example, Babiker et al.^{[ 11 ]} used a single camera to record five daily activities of a person in the room, including walking, lying, waving, boxing, and sitting, with a recognition rate of 94%. Khan et al.^{[ 12 ]} proposed a geriatric care system based on video sensors, which can recognize six types of abnormal elderly activities (e.g., falling and fainting), with an average accuracy of 95.8%. However, only a limited number of simple activities can be recognized in the aforementioned schemes, which cannot meet the goal of high-precision recognition of multiple complex activities.

Some researchers have introduced new equipment and proposed new recognition schemes to identify a variety of more complex human activities.^{[ 13–15 ]} Bagate et al.^{[ 13 ]} captured people's 3D skeleton data using a Kinect depth camera and achieved average recognition accuracy of 85% in the identification of eight common daily activities (including approaching and hugging). Hossen et al.^{[ 14 ]} adopted thermal infrared cameras to recognize 15 types of soldiers’ outdoor activities (including walking, holding a gun, and carrying a backpack), and achieved a recognition accuracy of 74%. The aforementioned studies have shown that recognition accuracy is limited as the number of required recognition activities increases. In addition, computer vision-based HAR has obvious shortcomings in volume and power consumption, and its recognition accuracy is easily affected by the light and barriers of the external environment. The processing of pictures or videos also raises extremely high requirements for the computing power and response speed of the HAR system.^{[ 16–18 ]}

Wearable Sensor-Based HAR

In recent years, wearable devices, characterized by small size, low cost, and low power consumption, have been selected by researchers to recognize simple daily activities. Some researchers have used single-node inertial sensors to recognize multiple daily activities.^{[ 19–23 ]} For example, Shen et al.^{[ 21 ]} developed a hand-controlled activity identification system with a Microsoft band based on an inertial measurement unit (IMU), with the advantages of high compliance and long wearing time (≈48 h). It has an average recognition accuracy of 97.3% in identifying seven hand and arm movements. Anjum et al.^{[ 22 ]} recognized 7 simple sports activities using the accelerometer and gyroscope of a smartphone, with an accuracy of 95.2%. Although the aforementioned schemes reduce the volume and power consumption of the HAR device, the number of activities their device can recognize is little compared to other sensing solutions.

Daily activity recognition schemes based on multisensor nodes, which contain more comprehensive and higher dimensional motion information, have proved more effective in recognizing various complex activities.^{[ 24–28 ]} For example, Bao et al.^{[ 24 ]} wore five accelerometers on different parts of the human body (the right hip and four limbs) to collect motion data, and recognized 20 kinds of motions with an accuracy of 84%. Andreas et al.^{[ 26 ]} placed three IMUs on the subjects’ right hand, right lower arm, and right upper arm, and recognized 12 daily activities (including opening windows, drinking water, and watering flowers) with an accuracy of 94.1%. In essence, HAR systems based on multisensor nodes perform much better in multiactivity recognition but are inconvenient to carry or wear and cause discomfort to the subjects due to a large number of nodes.

Table  1 summarizes various HAR schemes based on wearable sensors and compares the different schemes in terms of the number of nodes, number of detectable activities, and their corresponding recognition accuracy. More sensor nodes are often required to recognize multiple activities, but the multisensor node HAR systems are often bulky and inconvenient to wear, adding to the users’ physical and psychological burden. Conversely, single-sensor node HAR systems are reduced in the overall system size and significantly improve comfort, but they are usually only used to identify a few simple activities, while their recognition accuracy also needs to be improved.

Table 1 Previous research work related to microinertial measurement unit-based activity recognition (Acc-Acceleration, Gyro-Gyroscope)

Type	Literature	Sensor type	Node	Position	Node size [mm3]	Number of activities	Accuracy
Single-sensor node	Fatemeh Tahavori[ 19 ] (2017)	Acc, Gyro	1	Waist	30 × 18 × 10	Only 6 daily activities, including walking, sitting to stand, standing, and so on.	92.3%
	Emily J. Huang[ 20 ] (2022)	Acc, Gyro	1	Body side and back	Smartphone	Only 6 daily activities, including walking, standing, ascending stairs, descending stairs, sitting, sit-to-stand transition, stand-to-sit transition, and going through a revolving door.	88.9%
	Chao Shen[ 21 ] (2020)	Acc, Gyro	1	Wrist	47 × 18 × 12	Only 14 hand and arm activities, including flip wrist to left, rotating wrists, finger stroke, and so on.	97.3%
	Alvina Anjum[ 22 ] (2013)	Acc, Gyro	1	Waist	Smartphone	Only 7 sports activities, including walking, running, driving, cycling, and so on.	95.2%
	Edmond Mitchell[ 23 ] (2013)	Acc	1	Back	Smartphone	Only 7 sports activities, including walking, jogging, hitting the ball, and so on.	87.0%
Multisensor node	Ling Bao[ 24 ] (2004)	Acc	5	Limbs, right hip	70 × 98 × 18	20 daily activities, including walking, running, brushing teeth, reading, bicycling, and so on.	84.0%
	Kerem Altun[ 25 ] (2010)	Acc, Gyro, Mag	5	Chest, arms, legs	15 × 15 × 6	19 daily and sports activities, including walking, sitting, standing, jumping, rowing, and so on.	98.7%
	Bulling Andreas[ 26 ] (2014)	Acc, Gyro	3	Arms	21 × 21 × 8	12 daily activities, including smash, forehand, backhand, opening a window, and so on.	94.1%
	Yu-Liang Hsu[ 27 ] (2018)	Acc, Gyro	2	Wrist, ankle	37 × 56 × 15	21 daily and sports activities, including upstairs, standing, shooting basketball, tennis, golf, and so on.	98.2%
	Holger Junker [ 28 ] (2008)	Orientation angle	5	Back, arms, wrists	49 × 49 × 9	10 daily activities, including a drink, phone up, spoon, cutlery, handheld, and so on.	97.9%
Single-sensor node	This work	Acc, Gyro	1	Right index finger	25 × 16 × 19 (25-mm height includes the shank of the ring)	20 daily and sports activities, such as brushing teeth, running, doing squats, sitting to read, sweeping the floor, walking, playing basketball, and so on.	98.1%

Some commercial wearable devices used for HAR are also listed below for comparison (Table  2 ). Our device is believed to be smaller and lighter than these products, while could also detect more activity types. The only exception is the Oura Ring Generation 3 has similar physical attributes in terms of size and weight and could recognize 40 different sports activities, which are considered vigorous or fierce motion activities. However, the smart ring of the current study is the only device currently available that can detect multiple motion types, i.e., ranging from light, vigorous, to fierce activities. The smart ring of the current study is the first device capable of recognizing multi-intensity physical activities, which range from light-intensity household activities to vigorous- or fierce-intensity exercise-related activities, e.g., sit-to-eat, brush-teeth, run, operate-computer, sit-to-read, sweep-the-floor, clean-the-desk, do-sit-ups, playing badminton, playing basketball, riding a bike, run, walk, and so on.

Table 2 Comparison of the smart ring of the current study with existing commercial smart wearable devices (Acc-Acceleration, Gyro-Gyroscope, Mag-Magnetometer)

Device name	Type	Weight	Size (mm3)	Sensor used*	Intensity of activity that can be detected	Number of activities	Activity detection duration requirement
Apple Watch Series 7	Smartwatch	32 g	41.0 × 35.0 × 10.7	3D-Acc, 3D-Gyro, 3D-Mag	Vigorous and fierce activities	9 sports activities, such as elliptical training, indoor run, indoor walk, open-water swim, outdoor cycling, outdoor run, outdoor walk, pool swim, power	Not mentioned
GARMIN Venu 2 Plus	Smartwatch	51 g	43.6 × 43.6 × 12.6	3D-Acc, 3D-Gyro, 3D-Mag	Fierce activities	5 sports activities, such as biking, elliptical training, running, swimming, and walking	At least 10 min
Amazfit GTR 3 Pro	Smartwatch	32 g	46.0 × 46.0 × 10.7	3D-Acc, 3D-Gyro, 3D-Mag	Vigorous and fierce activities	8 sports activities, such as elliptical training, indoor run, indoor walk, outdoor run, outdoor walk, swimming, treadmill, and rowing machine activity	Not mentioned
HUAWEI Watch GT 3	Smartwatch	43 g	45.9 × 45.9 × 11.0	3D-Acc, 3D-Gyro, 3D-Mag	Vigorous and fierce activities	6 sports activities, such as elliptical training, indoor run, indoor walk, outdoor run, outdoor walk, and rowing machine activity.	Approximately 10 min for walking and ≈3 min for other workouts
Withings ScanWatch	Smartwatch	58 g	38.4 × 38.4 × 13.2	3D-Acc	Fierce activities	13 sports activities, such as badminton, basketball, boxing, cycling, dance, running, soccer, squash, table tennis, tennis, volleyball, walking, weights, and so on.	Not mentioned
Fitbit Charge 5	Smart band	28 g	36.7 × 22.7 × 11.2	3D-Acc	Fierce activities	7 sports activities, such as aerobic workout, elliptical training, outdoor bike, running, sports, swimming, walking, and so on.	At least 10 min
Galaxy Fit2 Pro	Smart band	33 g	51.3 × 25.0 × 12.5	3D-Acc, 3D-Gyro	Fierce activities	6 sports activities, such as cycling, dynamic workout, elliptical training, rowing machine activity, swimming, walking/running, etc.	At least 10 min
Oura Ring Generation 3	Smart ring	5 g	18.0 × 18.0 × 8.0	3D-Acc	Vigorous and fierce activities	40 sports activities, such as basketball, boxing, cycling, dance, elliptical training, golf, hiking, housework, running, soccer, surfing, swimming, tennis, walking, and so on.	At least 10 min
This work	Smart ring	7.3 g (including battery)	25.0 × 16.0 × 19.0 (25 mm height includes the shank of the ring)	3D-Acc, 3D-Gyro	Light, vigorous and fierce activities	20 sports activities and daily activities, such as sit to eat, brushing teeth, running, operating computer, doing sit-ups, sitting to read, sweeping floor, cleaning the desk, playing badminton, walking, playing basketball, riding a bike, etc.	At least 8 min

To assess the current work with other studies and the existing commercial devices, radar maps of the key comparison parameters are shown in Figure  1 . The smart ring shows a significant advantage over other research devices in terms of portability, i.e., small overall size in general while requiring only one sensing node on a finger (Figure 1a). In terms of activity–recognition accuracy and the number of detectable activities, the smart ring also shows a clear advantage over other devices. Moreover, the smart ring has advantages in lower weight and size, while being capable of detecting more daily and exercise-related activities over existing commercialized devices (Figure 1b). Although the smart ring has currently the disadvantage of detecting fewer activities than the Oura ring, it can detect activities with three different degrees of intensity (i.e., light, vigorous, and fierce). However, the Oura ring can only detect two degrees of intensity (i.e., vigorous and fierce). This capability of the smart ring may open up a wider range of applications for personal activity monitoring in the future. Moreover, with further improvement of the algorithms, the smart ring can also improve its capability of detecting much more than 20 activities with mixed intensities in the future.

Figure 1. Comparison with other wearable devices aiming at daily activity recognition. a) Comparison with devices from research groups; the proposed device of the current study has the advantages of higher accuracy and detecting more activities while having a smaller number of nodes and lower overall sensor volume compared to other devices. b) Comparison with commercialized devices; the device used in this study has the advantages of lower weight and volume while having high detection capability and several detectable activities.

Current Work

A new scheme to identify 20 kinds of activities using just a single smart ring driven by a hierarchical decision algorithm was proposed. The main contributions of this article can be therefore summarized as three key aspects: 1) Development of a new portable and lightweight experimental system that uses a small smart ring as a single HAR collection node; 2) Applying a special two-stage feature selection process based on random forest and Pearson correlation coefficient for activity representation; 3) Applying a robust and accurate hierarchical decision algorithm to realize high-precision recognition of multiple complex activities.

Correspondingly, the experimental process of the current work is shown in Figure  2 . First, the IMU and Bluetooth low-energy microcontroller (MCU) are integrated and embedded in a ring carrier to construct a portable data acquisition and transmission platform. The ring is worn on the index finger of the subject's right hand to collect acceleration and angular velocity signals generated by 20 kinds of human activities. Second, data preprocessing is used to eliminate noise signals in the original data. Finally, a hierarchical decision classification algorithm is proposed to quickly and accurately identify 20 kinds of activities.

Figure 2. Schematic diagram of the smart ring human activity recognition (HAR) system for processing collected data and recognizing 20 different daily and exercise activities.

In this work, a two-layer implicit category hierarchy is used to solve this problem. In the first identification layer, 20 types of activities are divided into 3 activity groups following the amplitude and periodic difference between the activities. In the second layer, the feature selection algorithm is used to select appropriate features to identify similar activities in each activity group. In the activity recognition process of each layer, the appropriate classification algorithm is selected by weighing the running speed against the recognition accuracy. The overall recognition accuracy reaches 98.10%.

Hardware System

In recent years, various smart sensor technologies and materials have been developed due to the application needs of different fields.^{[ 29–32 ]} For example, the textile triboelectric nanogenerator (TENG) technology provides sensors with the capabilities of sensing physical inputs and active self-powering.^{[ 29,30 ]} However, with the need for small size, low power consumption, operation stability, and low cost, we adopted inertial sensors to design the smart ring for this work. Inertial sensors can monitor the motion characteristics of the human body due to their ability to measure multiple motion directions, as well as their excellent sensitivity to both low and high-frequency human motions.

The smart ring consists of an embedded sensor board and a customized ordinary ring and a switch. The circuit board includes an NRF52832 SOC module, an MPU9250 inertial sensor unit (which includes a triaxial accelerometer, a triaxial gyroscope, and a triaxial magnetometer), and a button battery. The NRF52832 is an integrated module, including an MCU unit and wireless Bluetooth module, to minimize the overall volume. The MPU9250 is used as a motion detection module to detect multiple motion information such as acceleration and angular velocity. The button battery has a voltage of 3.3 V, which supplies power to the whole circuit board.

The MPU9250 internally integrates three independent 3-axis gyroscopes, 3-axis accelerometers, and 3-axis magnetometers, all of which output 16-bit digital data. Each sensor provides a digital output through a specialized ADC (analog to digital converter), and then interacts with the MCU through the integrated circuit bus (IIC) interface. When a user wears a ring for movement detection, the inertial measurement unit will collect data on the 3-axis accelerations and 3-axis angular velocities of the finger movements. The data will then be sent to the controller NRF52832. Finally, the data will be packed and sent to computing platforms, such as laptops and mobile phones, via Bluetooth module for processing

The ultra-lightweight (7.3 g, including battery) and ultra-small chip size (15 mm × 16 mm × 1 mm) of the ring make it comfortable to wear without affecting daily activities. The specific parameters are shown in Table  3 . In the scheme of the current study, the subject only wears a smart ring on the index finger of the right hand for high-precision recognition of 20 daily activities. Mobile phones are used as data receivers so that the subjects’ normal activities during the data collection process would not be affected. The frequency of people's hand movements in daily life is generally lower than 6 Hz. The sampling rate of the sensor is set to 50 Hz, which can fully meet the needs of the experiment and capture more motion details.

Table 3 Performance specifications of the smart ring

Item	Parameter
Angle accuracy [°]	0.1
Accelerometer detection range [g]	±16
Gyro detection range [° s−1]	±2,000
Battery capacity [mAh]	35
Supply voltage [V]	3.3
Working current [mA]	3
Data output frequency [Hz]	50
Baud rate [bps]	115 200

Experimental Method Data Acquisition

In this experiment, 8 students (four males and four females) participated in data collection for 10 days under natural and unsupervised conditions. Twenty typical daily activities were selected as the activity types to be recognized by the smart ring, (Table 3). The experimental procedures were reviewed and approved by the Ethics Committee of Northeastern University (No. NEU-EC-2021B023S). The collection process of a piece of activity data was adopted for illustration. Each subject was asked to wear a smart ring on their index finger for data collection, and the data of daily activities will be sent and stored on the mobile phone. After data collection, each subject is required to record what they did in a particular period, which is used as activity labels.

By observing the original data and analyzing the actual characteristics of the activities, 20 daily activities were found to be preliminarily divided into three activity groups based on the degree of intensity of exercise motions. The results are shown in Table  4 .

Table 4 Twenty activities and their labels

Label	Activity	Label	Activity	Label	Activities
A1	Sit to eat	B1	Brush teeth	C1	Run
A2	Operate computer	B2	Sit to makeup	C2	Do sit-ups
A3	Using smartphone	B3	Change clothes	C3	Do squats
A4	Sit to read	B4	Sweep the floor	C4	Skip rope
A5	Sit to write	B5	Clean the desk	C5	Play badminton
A6	Ride in the car	B6	Walk	C6	Play basketball
A7	Ride a bike	B7	Kick the shuttlecock	–	–

Activities A1–A7 can be described as recreational activities, which can be regarded as light activities. Moreover, activities B1–B5 can be described as daily behavior activities, which can be regarded as relatively vigorous activities. Furthermore, C1–C6 can be described as sports activities, which can be regarded as extremely fierce activities. The raw data from 20 activities can be observed in Figure  3 .

Figure 3. Raw gyroscope data of different activities. GX, GY, and GZ represent gyroscope data on the x-axis, y-axis, and z-axis, respectively.

Data Preprocessing

The moving average filtering algorithm is used to eliminate high-frequency noise generated by users’ unconscious jitter and electromagnetic influence. The mathematical formula is as follows.[Image Omitted. See PDF]where $y(n)$ is the new data sequence obtained after filtering, $x(n)$ is the raw data sequence, and N is the number of points in average filtering. The experiment of the current study indicated that the best filtering effect was achieved when N = 8.

A fixed sliding window segmentation method is used to segment long-lasting actions into lengths suitable for calculation and recognition. To be more specific, the window length is set to 60 s, and the window overlap rate is 50%. The segmentation process is shown in Figure  4 .

Figure 4. Data segmentation process of sliding window with 50% overlap rate and fixed sliding step.

In addition to the three-axis acceleration data and three-axis angular velocity data directly collected by the sensor, the resultant acceleration and the resultant angular velocity describing the overall changes in hand movements were also calculated as follows[Image Omitted. See PDF] [Image Omitted. See PDF]where $A_k$ and $G_k$ denote the resultant acceleration and the resultant angular velocity, respectively. $A_x$,$A_y$, and $A_z$ represent the acceleration data of x-, y-, and z-axes respectively. $G_x$, $G_y$, and $G_z$ represent the angular velocity data of x-, y-, and z-axes, respectively.

Hierarchical Decision Algorithm

The hierarchical decision algorithm can decompose complex multiclassification problems into multiple small-scale subproblems, thereby reducing computational complexity and improving the interpretability of selected features. Figure  5 describes the overall schematic diagram of the hierarchical decision algorithm.

Figure 5. Schematic diagram of the structure of the hierarchical decision algorithm. A1–A7 activity numbers 1–7, B1–B7 activity numbers 8–14, and C1–C6 activity numbers 15–20 in Table 3.

In the first level, 20 kinds of activities are divided into 3 activity groups according to the similarities and differences in the characteristics of different activities. In the second level, activities in each activity group are accurately recognized by selecting appropriate features and machine learning models.

First-Level Activity Recognition

Based on the divided principle, groups A, B, and C can effectively distinguish the amplitude-related feature due to the differences in activity intensity. The amplitude area can be used to describe the sum of the movement amplitudes of the finger in all directions during the entire movement process, as defined by Formula (4).[Image Omitted. See PDF]where $x(i)$, $y(i)$, and $z(i)$ represent the value of the ith sampling point of the x-, y-, and z-axes acceleration or gyroscope signals, respectively.

In addition, Fourier transform^{[ 33 ]} is performed on the x-axis acceleration data, and the variance in the frequency domain within the range of 0.1–6 Hz is calculated to describe the periodicity. The frequency domain variance is calculated by Formulas (5) and (6).[Image Omitted. See PDF] [Image Omitted. See PDF]where $x(m)$ is a finite-length sequence of length M, and $N(N\ge M)$ is the interval length of the discrete Fourier transform.

The distribution box diagram of 20 kinds of activities on $A_{\mathrm{h}}\text{\_area}$, $G_{\mathrm{h}}\text{\_area}$, and $A_x\_\text{FV}$ is shown in Figure  6 . Differences in amplitude and period between different activity groups were observed by comparing the feature value distribution intervals. The activities in groups A and B are mainly manifested in the difference in amplitude. The activities in group A have small feature values on $A_{\mathrm{h}}\text{\_area}$ and $G_{\mathrm{h}}\text{\_area}$, while those in group B have large feature values. The feature values of group C are >0.8 on $A_x\_\text{FV}$ due to the obvious periodicity, while those of the other two groups are all <0.8. This is almost consistent with the aforementioned analysis of the original data and also proves the correctness and interpretability of the extracted features of this study.

Figure 6. Distribution of 20 types of activities on features of Ah_area, Gh_area, and Ax_FV.

Second-Level Activity Recognition

In the second-level recognition algorithm, features are first extracted to describe the activity characteristics. Feature contribution ranking and feature correlation analysis are then performed to reduce the dimensionality of features and eliminate redundant features (Figure  7 ). Finally, the best feature subset is used to identify activities in the three groups, and the appropriate classifier was selected following recognition accuracy.

Figure 7. Important evaluation results of 64 features in group A for motion recognition.

The time and frequency domain features are extracted from the eight dimensions $(A_x,A_y,A_z,A_{\mathrm{h}},G_x,G_y,G_z,G_{\mathrm{h}})$ of the sample data. For each dimension, 15 features (15 × 8 = 120D features) are extracted to characterize the activity data and identify human daily activities. The 15 activity features include: 1) mean, 2) standard deviation (std), 3) variance (var), 4) information entropy (ent), 5) amplitude area (area), 6) root mean square (rms), 7) mean absolute error (MAE), 8) maximum (max), 9) minimum (min), 10) median (med), 11) mode, 12) range, 13) intermediate range (mid), 14) interquartile range (IQR), and 15) energy (EN).^{[ 34–37 ]} In addition, the maximum–minimum normalization method is used to normalize the feature value to eliminate the influence of unit and scale differences between features [0,1].

Aiming at the 120D features extracted earlier, the feature selection method of random forest and feature correlation analysis is used to reduce the feature dimensionality. First, the Gini coefficient of the random forest is used to evaluate the importance of features as defined by Formula (7).[Image Omitted. See PDF]where K represents the number of categories of feature samples, and p _k represents the proportion of category k to the number of all nodes.

The larger the Gini coefficient of the feature, the higher the contribution of the feature to activity recognition. Taking the feature selection of activities in group A as an example, the top 56 features account for 80% of the total contribution of all features after random forest sorting (Figure  8 ). The last 64 features only consume computing resources and cannot significantly improve recognition accuracy and are removed from the feature set.

Figure 8. Flowchart of feature selection.

The Pearson correlation coefficient between features is then calculated to quantify the linear correlation degree between them (Formula 8). Two features with a correlation coefficient >0.9 are considered highly correlated, redundant features, and only the feature with a higher contribution rate will be retained.[Image Omitted. See PDF]where $\text{Cov}(X,Y)$ represents the covariance of vectors X and $Y,{\sigma }_x,{\sigma }_y$ represent the standard deviations of vectors X and Y, E represents the expectation, and ${\rho }_{xy}$ represents the Pearson correlation coefficient between vectors X and Y.

The overall flowchart of feature selection is shown in Figure 8. For the feature set $F_i$ containing i elements, feature dimensionality reduction is performed in two steps. In the first step, the first k features with a cumulative contribution of 80% are retained, and the last i-k features with lesser importance are dropped. In the second step, the correlation between the two features is sequentially compared. If the correlation coefficient is >0.9, only the feature with greater importance will be retained; otherwise, both of them will be retained. After feature selection, the feature dimension is greatly reduced, which reduces computational complexity and running time. The results of the feature selection are shown in Table  5 .

Table 5 Changes in feature number after feature selection

Group	Original	After random forest	After correlation analysis
A	120	56	26
B	120	50	27
C	120	51	17

Results and Discussion Result of First-Level Activity Recognition

In the first level, 20 types of activities are preliminarily divided into 3 activity groups. In our experiment, a total of 2398 samples were tested to extract the feature values. As discussed earlier, three features ($A_{\mathrm{h}}\text{\_area}$, $G_{\mathrm{h}}\text{\_area}$, and $A_x\_\text{FV}$) are extracted to describe the differences in amplitude and periodicity between the three activity groups in the first-level activity recognition process. To find a suitable classifier for this level of decision, five classifiers, including linear discriminant analysis (LDA), naive Bayes (NB), classification and regression trees (CART), support vector machine (SVM), and K nearest neighbors (KNN), are adopted for activity group classification. The final results are shown in Figure  9 .

Figure 9. Comparison of recognition results in the first level of hierarchical decision.

The KNN classifier shows the highest recognition accuracy among the five classifiers (Figure 9). Moreover, it can be also observed to spend less time in the recognition process. Therefore, KNN was adopted as the classifier for the first level of activity group recognition. After fivefold cross-validation, the average recognition accuracy of this level is 99.2%. The detail of the final recognition results is shown in Figure  10 .

Figure 10. Final recognition results of different groups in the first-level decision.

Result of Second-Level Activity Recognition

In the second level, the three best feature subsets (containing 26, 27, and 17 features) obtained after feature selection are input into 5 classifiers to identify activities within groups A, B, and C. To select a better classifier for second-level recognition, the recognition process was conducted in each group with different classifiers. The recognition results are shown in Figure  11 .

Figure 11. Comparison of the recognition results of different classifiers in the second level of hierarchical decision.

SVM has the highest accuracy in recognizing the activities in each activity group (Figure 11). Therefore, the SVM algorithm is selected as the classifier for the second level of activity group recognition.

In addition, the response speed of the system is hoped to improve by reducing the number of input features under the premise of ensuring high algorithm accuracy. The effect of the number of features on classification accuracy and running speed was analyzed (Figure  12 ).

Figure 12. Relationship between recognition accuracy and running speed with the number of features.

Thus, as the number of features increases, the accuracy of the algorithm increases, while the running speed decreases. Moreover, when the number of features of groups A, B, and C is at 12, 15, and 6, respectively, they can ensure a high running speed while guaranteeing high-precision recognition.

Relying on the above trade-off between running speed and recognition accuracy, 12, 15, and 6 features were finally selected to identify different activities in groups A, B, and C, respectively. After fivefold cross-validation, the recognition accuracies of groups A, B, and C are 98.3%, 98.4%, and 99.7%, respectively. The confusion matrix of second-level activity recognition results is shown in Figure  13 .

Figure 13. Recognition results of different groups in the second-level decision.

Final Recognition Results of Hierarchical Decision Algorithm

Based on the aforementioned study, three features for the first-level recognition were finally selected, and 12, 15, and six features for groups A, B, and C, respectively. Two different classifiers, KNN and SVM, are specially adapted for first- and second-level decisions. These two-level recognition methods were then combined, and 20 activities with different classifiers for comparison were recognized. The results are shown in Figure  14 .

Figure 14. Recognition results from a hierarchical decision algorithm.

The algorithm of the current study shows the highest recognition accuracy among these algorithms, which reaches 98.1% (Figure 14). In addition, it also has a fast-running speed when compared with some other algorithms (e.g., CART, LDA, NB, and SVM). Although it is slightly lower than KNN, the recognition accuracy is higher by 5.4%, which shows an advantage.

Conclusion

Previous HAR solutions often fail to meet the requirements of portability, high precision, and multiactivity recognition simultaneously, which greatly limits their applications in daily life. In this work, a high-precision HAR system based on a single-node MEMS (Micro Electro Mechanical Systems) sensor and a hierarchical decision algorithm is proposed to identify 20 common human daily and exercise activities. The MEMS sensor is embedded in a ring carrier and worn on the subject's right index finger to collect activity data. The ultra-small size and mass of the smart ring make it portable and easy to wear. The hierarchical decision algorithm is used to decompose the multiclassification problem of complex activities, and the feature selection algorithm integrated into the framework was used to reduce the feature dimensionality. These measures effectively shorten system running time and improve recognition accuracy. The experimental results show that the current HAR scheme has an average accuracy of 98.10% in identifying 20 kinds of daily and exercise activities and is both practical and accurate. The portable, reliable, and high-precision HAR solution proposed in this article is of great significance to lifestyle supervision, time planning, and healthcare management.

Acknowledgements

This work was supported by the National Natural Science Foundation of China (Grant no.61873307), the Hebei Natural Science Foundation (Grant no. F2020501040, F2021203070, F2022501031), the Fundamental Research Funds for the Central Universities under Grant N2123004, the Administration of Central Funds Guiding the Local Science and Technology Development (Grant no. 206Z1702G), and in part by the Science, Technology, and Innovation Commission of Shenzhen Municipality Project (SGDX2019081623121725).

Conflict of Interest

The authors declare no conflict of interest.

Data Availability Statement

Research data are not shared.