A Yoga Pose Difficulty Level Estimation Method Using OpenPose for Self-Practice System to Yoga Beginners

Abstract

Yoga is an exercise preferable for various users at different ages to enhance physical and mental health. To help beginner yoga self-practitioners avoid getting injured by selecting difficult yoga poses, the information of the difficulty level of yoga poses is very important to provide an objective metric to assist yoga self-practitioners in selecting appropriate exercises on the basis of their skill level by using the yoga self-practice system. To enhance the developed yoga self-practice system, the yoga difficulty level estimation function will enable users to clearly understand whether the selected yoga poses are suitable for them. In this paper, the newest difficulty level estimation method of yoga poses is proposed by using and analyzing OpenPose two-dimensional (2D) human body keypoints. The proposed method effectively uses the selected six keypoints areas of the upper and lower body, body support types, center of gravity calculations, and body tilt angles and slopes to produce estimations. Firstly, the method calculates the weighted centers of the upper and lower human body for each pose by using keypoints. Secondly, it refers the slope of the centroid line between the two centers and infers the body’s balance state. Lastly, the system estimates the difficulty level by additionally considering the keypoints of the body to contact the ground. For evaluations of the proposal, more than one hundred yoga poses are collected from the Internet and applied to classify them into five difficulty levels. Through comparisons with subjective levels from one instructor and 10 users, the validity of the estimation results is confirmed, a comparison is performed with existing designs, and it is implemented in embedded systems.

1. Introduction

In recent years, demand for personalized self-practice systems for physical exercise has grown considerably. Yoga, with its diverse range of poses, stands out for its accessibility and benefits for physical, mental, and spiritual health. However, yoga practitioners often struggle to gauge the difficulty of various poses, leading to incorrect practice and injury. Thus, an objective method for estimating the difficulty of yoga poses is essential. Previously, we developed a yoga self-practice system based on OpenPose [1,2] to guide users through poses by comparing their movements with those of an instructor. Although this system provides real-time feedback, it lacks a reliable method for estimating the difficulty of poses—a feature crucial for tailoring practice to an individual’s skill level.

As we know, yoga has many challenging poses, and the level of difficulty varies depending on an individual’s flexibility, strength, and experience. At the training School of Yoga [3,4], the yoga teachers help you perform these challenging yoga poses [5,6]. For example, Figure 1 shows two challenging yoga poses that require advanced strength, flexibility, and balance. Figure 1 depicts the “Handstand” yoga pose. This yoga pose is an inversion pose that requires strong strength, balance, and focus in the upper body, core, and shoulders. Before yoga self-practitioners practice this challenging pose, to avoid injuring by directly selecting this difficult pose, yoga practitioners must be informed of the yoga pose’s difficulty level, and the yoga self-practice system will suggest for them to practice the forearm plank, dolphin pose, and handstand against a wall in advance to build body strength and confidence [3]. Similarly, Figure 1 depicts the “Scorpion” yoga pose. This is also a challenging backbend pose that poses a high difficulty level regarding flexibility in the spine, shoulders, and hips. To train for this very difficult pose, yoga self-practitioners can be suggested to practice the camel pose, wheel pose, and forearm stand beforehand to raise their flexibility and force in the spine and shoulders [3]. To prevent self-practice users from getting injured by attempting difficult poses that are not suitable for them, the objective-based yoga pose difficulty level estimation is very important to provide information for at-home yoga self-practitioners.

Figure 1. Left: “Handstand” pose. Right: “Scorpion” pose.

In this paper, we propose a difficulty level estimation method for yoga poses that integrates body area, body tilt, and body support considerations to deliver accurate estimations. Our method incorporates the OpenPose two-dimensional (2D) BODY25 model, which provides comprehensive keypoint detection for the human body. By calculating the weighted centers of the upper and lower body, analyzing the slope of the centroid line between these centers, and considering ground contact points, the proposed system can estimate the difficulty level of each pose. This method aims to provide an objective metric to assist yoga self-practitioners in selecting appropriate exercises on the basis of their skill level. By considering numerous yoga pose combinations, our goal is to improve the user-friendliness of the yoga self-practice system, and the proposed approach aims for a win–win scenario, promoting both health and technological advancement. Evaluation of the method is conducted by using the method to classify more than 100 yoga poses into five difficulty levels, and the results are validated through comparisons with subjective difficulty ratings provided by a professional instructor and 10 users. By comparing the yoga pose difficulty levels between the proposed objective-based design and the subjective opinions of users and professional yoga instructors, the results prove the reliability of this study.

For the proposed yoga pose difficulty level estimation, only the human body keypoints of yoga instructors and yoga teachers are selected for reference and evaluation. The proposed method does not refer to or use any human body keypoint of the yoga beginner and yoga self-practitioner. Therefore, the proposed method does not require to compare the body keypoints between the yoga teacher and student, as they may have different body proportions.

On the other hand, for the proposed yoga self-practice guide system, the yoga student must follow the correct yoga poses from the guidance videos of yoga teachers. To guide yoga students and beginners, the proposed image-based yoga self-practice system will compare in real-time the angles of yoga poses from the detected keypoints between the yoga teacher and student. Since the pose comparison is angle based, the proposed image-based yoga self-practice system can overcome the problem of having different body proportions.

The contributions of this study are listed as follows: (1) Based on the authors’ best knowledge and understanding in this research field, the proposed method is the newest state-of-the-art study to use the information of human body keypoints to research the objective-based yoga pose difficulty level estimation. (2) By using the 2D OpenPose model, the newest proposed method effectively uses the six selected keypoints areas of the upper and lower body, body support types, center of gravity calculations, and body tilt angles and slopes to produce estimations. To evaluate the yoga difficulty level, the three parameters, i.e., “base difficulty”, “angle difficulty”, and “area difficulty”, are combined for the utilization. (3) To avoid getting injured by selecting difficult yoga poses, the proposed method provides an objective metric to assist yoga self-practitioners in selecting appropriate exercises on the basis of their skill level by using the yoga self-practice system.

The remainder of this paper is organized as follows. Section 2 reviews the relevant literature. Section 3 describes the yoga self-practice system proposed in this paper. Section 4 presents the difficulty level estimation method. Section 5 reports the experimental results of the proposed method for evaluations. Finally, Section 6 concludes the paper and provides suggestions for future works.

2. Related Works

Skeleton detection projects are increasingly implemented in pose estimation research, and they are widely used in various fields, such as interactive games, intrusion detection, behavior analysis in retail, sports training assistance, fall detection, and public safety on production lines. The present study focuses on related research in sports education and interactive games.

In the field of yoga pose recognition, Saluja et al. [7] studied common convolutional neural network models to train a yoga pose classifier. They meticulously compared the accuracy differences between various models, providing a comprehensive understanding of which models perform more favorably in different scenarios. They also discussed the medicinal value of yoga, particularly its effectiveness in improving joint pain and enhancing overall physical well-being. The present study laid a foundation for integrating artificial intelligence in health care, demonstrating that yoga could be a valuable tool in medical treatment and rehabilitation when correctly practiced and monitored through artificial intelligence systems.

Gupta et al. [8] contributed substantially to the field of yoga practice by highlighting the importance of correct pose execution. They indicated that although yoga is an excellent form of exercise, incorrect poses can have negative health consequences. To address this, they proposed a system based on OpenPose 18-point skeleton detection to guide users in performing yoga poses correctly. Their system provides real-time feedback, ensuring that users can adjust their poses immediately to prevent injuries and maximize the benefits of their yoga practice. Their study underscored the key role of artificial intelligence in enhancing the accuracy and safety of yoga practice.

Reyes et al. [9] took another approach by using a Mediapipe detection framework to design a human pose estimation method. Their method objectively measured the performance of users in terms of yoga pose execution by performing angle calculations to ensure the consistency of experimental parameters. This objective approach is crucial in providing reliable feedback and improving the overall performance of users. By emphasizing the importance of objective metrics, their study contributed to the development of accurate and reliable pose estimation systems.

Singh et al. [10] studied the yoga pose recognition using AI and ML and aimed to develop automatic structures that can find and classify yoga poses based solely on visual input. This era can help yoga professionals, teachers, and fans by introducing ongoing critiques, personalized guidance, and overall performance reviews. The website’s artificial intelligence model can accurately identify different postures thanks to a rich dataset of yoga poses.

Talaat et al. [11] proposed the yoga pose detection by an artificial intelligence technology for providing standardized and appropriate yoga poses. Incorrect yoga postures lead to serious injuries and long-term complications. By analyzing human posture, the system detected and corrected abnormal postures to improve user’s health at home. The pose estimator extracts yoga asana attributes from well-presented images. The purpose of this study is to explore different ways to classify yoga poses by the LGDeep model, which combines a new residual convolutional neural network and three deep learning approaches: Xception, VGGNet, and SqueezeNet. Experimental results revealed that the LGDeep classifier outperformed other approaches and achieved the highest classification accuracy ratio.

Vinutha et al. [12] studied to detect and correct yoga poses by machine learning technology. This project focused on developing a network architecture for a yoga pose recognition system by transfer learning with Convolutional Neural Network VGG19. The performance of the proposed model was evaluated by its ability to accurately classify different yoga poses under different conditions including changing posture backgrounds, lighting, and environment. The developed yoga pose recognition system aims to help yoga practitioners and teachers achieve correct posture alignment and provide instant feedback.

In Kishore et al. [13], the recent pandemic led to an increase in the number of yoga practitioners, many of whom practice without proper guidance. This research was proposed with the aim of facilitating the work of these practitioners by implementing a deep learning-based method that can estimate the correct poses to be performed. In this study, the authors implemented this approach by using four different deep learning architectures: EpipolarPose, OpenPose, PoseNet, and MediaPipe. This database contained images of five commonly practiced yoga poses: tree pose, triangle pose, half-moon pose, mountain pose, and warrior pose. By the authentic database for training the model, the deep learning-based system was used in real-time applications.

Verma et al. [14] surveyed that the merging of yoga and artificial intelligence opened up new opportunities to improve yoga practices and made yoga more accessible. Yoga algorithms driven by artificial intelligence provided the tailored yoga practices to help yoga practitioners achieve their goals and overcome limitations. Virtual yoga teachers provided personalized guidance, and their presence helped beginners feel more confident in their practice. In addition, artificial intelligence was used to analyze and understand the benefits of different yoga poses and exercises, and the combination of these two fields had the potential to revolutionize the way yoga was practiced and taught.

Archana et al. [15] described the importance of automatic yoga pose detection in providing instant feedback and improved self-correction. In this review, different methods were considered to classify yoga poses by focusing on PoseNet and KNN classifiers. It leveraged deep learning to identify yoga poses by using effective algorithms such as OpenPose, human pose estimation, and computer vision technology. By combining CNN and LSTM on real-time surveillance videos to detect yoga poses, the effective algorithm created body skeletons of yoga poses and extracted joint angles as feature for learning the deep learning models. In addition, a deep learning technology was developed to detect incorrect yoga postures, which enabled the system to detect abnormal angles between the actual and desired posture, and it prompted users to improve their posture.

According to the work of Rajendran et al. [16], yoga is a great form of physical activity, and posture-based yoga improves flexibility, energy, and overall brain activity, and reduces stress, blood pressure, and back pain. When users perform yoga asanas, the posture adjustment is a very important aspect. Many yoga poses require forward or backward bending, which, if performed incorrectly, puts strain on the users’ joints, ligaments, and spine, and it leads to possible hip problems. Therefore, it is important to pay attention to proper yoga postures while practicing. By using the advances in computer vision algorithms and sensors, yoga pose prediction and automatic motion analysis are possible. This study explored an in-depth analysis of a yoga pose recognition system by computer vision, machine learning, and deep learning techniques.

All the aforementioned studies [7,8,9,10,11,12,13,14,15,16] focused on human pose detection and classification. Although Mediapipe [17] is frequently used for hand and body pose detection, numerous studies have used Microsoft Windows as the experimental platform instead of embedded platforms. The choice of platform considerably influences the performance and applicability of a pose detection system. Embedded platforms, such as NVIDIA Jetson Nano [18], provides advantages in portability and low power consumption, making them suitable for real-time applications in various environments. The OpenPose 18-point model can be used to identify yoga pose joint angles. This model is typically sufficient for most applications. However, for more complex activities, such as detailed yoga poses, a richer set of keypoints is necessary. The two-dimensional (2D) BODY25 model, which includes 25 keypoints, provides more detailed information on poses and is more suitable than the OpenPose 18-point model for intricate tasks such as yoga pose estimation.

Most studies have focused on detecting whether yoga movements are performed correctly by the user or instructor. However, few studies have examined the selection and estimation of the difficulty of yoga poses. Applying the concept of difficulty estimation proposed by our prior works [19,20], the present study explored difficulty estimation in greater depth, proposing a novel method and validating its effectiveness. In our prior work [19], this study proposed a preliminary yoga pose difficulty estimation method by considering the keypoints, movement duration time, and stability of the yoga poses. The estimated pose levels were provided as a reference for yoga students. The difficulty estimation applied to the yoga instructor videos in the system and the estimated difficulty results were relative. In our prior work ref-20, by using OpenPose, the two different difficulty assessment methods were developed. Firstly, the angular velocity calculations were used to perform the dynamic and static regions of each yoga instructor’s motion for the evaluation of difficulty based on angular and temporal analyses. Secondly, only the body area enclosed by four selected keypoints of the upper and lower body was calculated for the yoga difficulty analysis by the area difference comparison.

In this study, by using OpenPose, the newest difficulty level estimation method of yoga poses is proposed. The newest proposed method effectively uses the six keypoint areas of the upper and lower body, body support types, center of gravity calculations, and body tilt angles and slopes to produce estimations. To evaluate the yoga difficulty level, with the parameters calculated from the body area, the definition of pose difficulty is established. To construct the difficulty evaluation formula, the three parameters including “base difficulty”, “angle difficulty”, and “area difficulty” are combined for the application. The present study provides more comprehensive findings relative to the aforementioned research for joint richness, detection model comparisons, yoga pose identification, and yoga pose difficulty estimation. Through the integration of BODY25, the proposed method can achieve high accuracy in detecting and analyzing yoga poses. Additionally, the use of embedded platforms enables the development of portable and user-friendly applications, which are crucial for real-world use.

3. Review of Yoga Self-Practice System

In this section, the previous studies of the yoga self-practice system are reviewed in several parts. The first part provides an overview of yoga self-practice, including the overall system design concept, software and hardware architecture, and system presentation. The second part introduces the human skeleton estimation model used in this research area and the selection of keypoints. The third part discusses the main algorithms, including the angle calculations and vectors obtained using keypoints. The fourth part compares the poses of users and instructors, discussing the definitions of dynamic and static regions. The fifth part describes the scoring system designed using fuzzy theory. The final part covers image output and web interactions, which are designed using Flask.

3.1. The System Hardware Architecture

The hardware architecture of the system in our prior work [21] uses NVIDIA Jetson Nano to implement yoga training for users. Figure 2 depicts the hardware computation flow of the system, and Figure 3 presents the hardware devices used.

Figure 2. System implementation on NVIDIA Jetson Nano.

Figure 3. Yoga self-practice system hardware environment in our prior work [20].

3.2. Software Algorithm Architecture

OpenPose is used as a detection model. OpenPose extracts human keypoint coordinates to perform angle difference calculations. The main algorithms are implemented in Python, including preprocessing and system processing (Figure 4). Finally, Flask is used to build a web interface for user demonstrations.

Figure 4. Architecture of yoga self-practice system in our prior work [21].

3.3. Human Body Keypoints Extraction

For human keypoint extractions, we use the 2D BODY25 model proposed by Cao et al. [1,2] as the prediction model. We input images of the instructor or user and extract joint keypoint information. Subsequent calculations are performed using coordinate positions. Figure 5 displays the keypoints extracted from the instructor’s image by using OpenPose.

Figure 5. Architecture of yoga self-practice system, where the “Red”, “Green”, “Yellow”, “Dark blue”, and “Cyan” colors indicate the detected parts of the torso, left arm, right arm, left leg, and right leg, respectively.

Figure 6 depicts the BODY25 model, which enables feature extraction for 25 keypoints of the whole body. Compared with other models, this model provides richer and more accurate joint keypoints. Overall, this advantage is the main reason for selecting OpenPose. However, not all keypoints are included in the yoga self-practice system calculations to reduce the computational load, allowing the method to run in real-time on Jetson Nano. Table 1 presents the keypoint information used in the yoga self-practice system to ensure both system performance and full operability.

Figure 6. Illustration of the MPII, COCO, and BODY25 models model in OpenPose.

Table 1. Five yoga pose difficulty levels.

Yoga Pose	Selected Keypoints
Mountain	3rd, 4th, 6th, 7th
Side_bend	3rd, 4th, 6th, 7th
Seat_1	3rd, 4th, 6th, 7th
Seat_2	3rd, 4th, 6th, 7th
Warrior	3rd, 4th, 6th, 7th, 10th, 11th, 13th, 14th

In Figure 6, the different colors indicate the different body parts in the three model. By the MPII model, “Red”, “Dark blue”, “Green”, “Yellow”, “Cyan”, and “Teal” colors indicate the detected parts of the head, torso, left arm, right arm, left leg, and right leg, respectively. By the COCO model, “Red”, “Green”, “Yellow”, “Dark blue”, and “Cyan” colors indicate the detected parts of the neck, left arm, right arm, left leg, and right leg, respectively. By the BODY25 model, “Red”, “Green”, “Yellow”, “Dark blue”, and “Cyan” colors indicate the detected parts of the torso, left arm, right arm, left leg, and right leg, respectively.

3.4. Process of System Calculation

After extracting the keypoints, we then calculate the angle differences and input the results into a fuzzy system, assigning weights to each keypoint to obtain the overall Shape Score of a pose. Figure 7 depicts how this process enhances the yoga self-practice system, further refining it.

Figure 7. Human body Shape Score calculation on Jetson Nano.

3.5. Previous Difficulty Estimation

Regarding the yoga self-practice system, our prior work [19] proposed a method that uses angle differences and total distance to define the difficulty parameters for the upper and lower body (Figure 8). This method calculates the difficulty values based on the keypoints of the upper and lower body, and then compares these values with user feedback. However, this method has several limitations. For example, calculations for the entire body are performed only for the Warrior Pose, whereas only the upper body is analyzed for all other poses.

Figure 8. Estimation of pose difficulty in related design, where the red boxes define the upper and lower body areas for difficulty estimation.

Data normalization is required for this method because the upper and lower body parameters are calculated differently. Figure 9 depicts how the normalized values are used to determine the difficulty level of a pose.

Figure 9. Estimating difficulty values normalization in our prior work [19].

The proposed method identifies the Warrior Pose as the most difficult yoga posture (Figure 10). However, given the vast number of yoga poses, other more challenging poses must exist.

Figure 10. Estimating difficulty values normalization in our prior work [19].

4. Proposal of Difficulty Level Estimation Method

This section discusses the proposed method, which is an innovative computational model specifically designed for the yoga self-practice system. Yoga poses require maintaining balance and holding positions for specific durations, which are fundamental aspects of a training regimen. Traditionally, the assessment of the difficulty of yoga poses is typically based on subjective opinions or difficulty indicators suggested by professional yoga instructors. The primary motivation of this study is to objectively estimate the difficulty of yoga poses, enabling users to accurately select suitable poses before training, thereby benefiting all users and enhancing the usability of the yoga self-practice system. To estimate the yoga pose difficulty level by preprocessing the yoga pose images from yoga experts, only the body keypoints of yoga teachers can be selected to be used for the applications, and the proposed design does not use the body keypoints of students to perform the yoga pose difficulty level estimation.

4.1. Overview of Pose Difficulty Estimation

Maintaining yoga poses will be difficult, especially for difficult yoga poses. This section presents an overview of the estimation method applied in the present study. This method uses estimation parameters such as body support points, areas of the upper and lower body, and the body tilt angle. By summing these parameters and inputting them into the difficulty level calculation system, we can determine the objective difficulty level of a yoga pose. Figure 11 presents an overview of the system. This estimation method is designed to optimize the initial pose selection for users of the yoga self-practice system, providing them with enhanced safety and suitability.

Figure 11. Processing flow of yoga self-practice system.

4.2. Extraction of Yoga Pose Keypoints by Using OpenPose

To select a detection model, the OpenPose 2D human body keypoints BODY25 model is employed. The detected human keypoint information of this model is presented in Figure 12. The BODY25 model is selected over other models (e.g., MPII or COCO18) because it can provide more detailed human keypoint information. The BODY25 model provides advantages in terms of accuracy and keypoint information, making it highly suitable for analyzing and estimating the difficulty of yoga poses.

Figure 12. OpenPose BODY25 model, where “Red”, “Green”, “ Yellow”, “Dark blue”, “Cyan” colors indicate the parts of torso, left arm, right arm, left leg, right leg, respectively.

After extracting body keypoints, we apply a calculation method that uses six keypoints from both the upper and lower body to estimate overall body balance without increasing the computational load. Given that yoga movements predominantly involve the hands and feet, this method allows us to effectively assess balance outcomes. By focusing on these keypoints, we avoid unnecessary computations that could otherwise strain the system’s capacity. Table 2 lists the selected body keypoints used in the present study.

Table 2. Selected human body keypoints.

Human Body	Keypoint Numbers	Corresponding Joint Names
Upper-Body	Keypoint 2	Right Shoulder
	Keypoint 3	Right Elbow
	Keypoint 4	Right Wrist
	Keypoint 5	Left Shoulder
	Keypoint 6	Left Elbow
	Keypoint 7	Left Wrist
Lower-Body	Keypoint 9	Right Hip
	Keypoint 10	Right Knee
	Keypoint 11	Right Ankle
	Keypoint 12	Left Hip
	Keypoint 13	Left Knee
	Keypoint 14	Left Ankle

Because this system is designed for beginners to practice and learn yoga poses, the selected yoga pose dataset comprises the following five poses in our prior work [21]: Mountain, Warrior, Side Bend, Seat 1, and Seat 2. To increase the credibility of our research, we also extracted nearly 100 yoga poses [22] and included yoga poses that we recorded for application. Figure 13 depicts the 5 yoga poses used in the previous system and the nearly 100 yoga poses used in the current design for difficulty estimation.

Figure 13. Datasets for evaluation.

4.3. Estimation of Difficulty Parameters

The present study proposes an objective method to estimate the difficulty of yoga poses by calculating the area of six keypoints on both the upper and lower body. This method also incorporates other parameters, such as body weighted centers, the slope of the body tilt curve, angle differences, and various types of body support points. In this section, we comprehensively discuss the design principles, validation, and implementation of the estimation method. This method differs considerably from previous methods for estimating yoga pose difficulty. In contrast to the method for estimating pose difficulty proposed by our prior work [19], the present study’s proposed method incorporates a consistent evaluation model that is applicable to a wide range of poses, thereby greatly enhancing the credibility of the present study.

The present study also compares the proposed design with previous related designs (Table 3). Compared with our prior work [19], the proposed method uses more body keypoints, providing more joint information. A previous design used angle differences to calculate upper body estimations and the total distance to calculate lower body estimations. By contrast, the proposed design uses the area of the upper and lower body and other parameters to ensure consistency for full-body estimations. Additionally, a previous design did not perform full-body calculations for all five yoga poses, sometimes evaluating only the upper body. In contrast to previous designs, the proposed method is applicable to all yoga poses and allows for comprehensive full-body yoga pose analysis and estimation.

Table 3. Comparisons of yoga pose difficulty analysis between two different designs.

Comparison Metrics	Our Prior Work [19]	Proposed System
Total number of keypoints	Upper body: 4 points	Upper body: 6 points
	Lower body: 4 points	Lower body: 6 points
	8 points used in total	12 points used in total
Estimation method	Upper body uses angle difference, and lower body uses the total distance	Both upper and lower bodies use area calculation and body slope, body weight
Whole-body analysis	Some poses only use the upper body	Each Yoga pose is applicable

4.3.1. Calculation of Upper and Lower Body Areas and Weighted Centers

The position of the center of mass is crucial because yoga requires body balance. The estimation method in the present study involves the selection of six keypoints from the upper and lower body (extracted by OpenPose) to calculate the body area and the position of the center of mass (Figure 14). The goal is to analyze parameters related to these regions.

Figure 14. Proposed estimation scheme for upper and lower body areas [20].

The areas of the upper and lower body are calculated separately using the shoelace formula. Equations (1) and (2) represent the calculations for the upper and lower body, respectively:

$$A_{\mathrm{upper}}={\displaystyle \frac{1}{2}}\left|\sum _{i=0}^5(x_i{y}_{i+1}-x_{i+1}{y}_i)\right|$$

(1)

$$A_{\mathrm{lower}}={\displaystyle \frac{1}{2}}\left|\sum _{i=6}^{11}(x_i{y}_{i+1}-x_{i+1}{y}_i)\right|$$

(2)

where “A_upper” and “A_lower” represent the areas of the upper and lower body, respectively; x and y represent the x-axis and y-axis coordinates of the selected body keypoint; and “i” represents the total number of joint keypoints for both the upper and lower body, summing up to 12 keypoints. The weighted center for the upper and lower body is calculated by using the areas obtained through the centroid formula. Equations (3) and (4) represent the calculation for the upper body:

$$X_{\mathrm{centroid}\mathrm{upper}}={\displaystyle \frac{1}{A_{\mathrm{upper}}}}\sum _{i=0}^5(x_i+x_{i+1})(x_i{y}_{i+1}-x_{i+1}{y}_i)$$

(3)

$$Y_{\mathrm{centroid}\mathrm{upper}}={\displaystyle \frac{1}{A_{\mathrm{upper}}}}\sum _{i=0}^5(y_i+y_{i+1})(x_i{y}_{i+1}-x_{i+1}{y}_i)$$

(4)

For the lower body, the same centroid formula is used as expressed in Equations (5) and (6). These equations are also employed to calculate the body areas and parameters on the basis of the six body keypoints:

$$X_{\mathrm{centroid}\mathrm{lower}}={\displaystyle \frac{1}{A_{\mathrm{lower}}}}\sum _{i=6}^{11}(x_i+x_{i+1})(x_i{y}_{i+1}-x_{i+1}{y}_i)$$

(5)

$$Y_{\mathrm{centroid}\mathrm{lower}}={\displaystyle \frac{1}{A_{\mathrm{lower}}}}\sum _{i=6}^{11}(y_i+y_{i+1})(x_i{y}_{i+1}-x_{i+1}{y}_i)$$

(6)

Equations (3)–(6) are used to compute the x and y coordinates of the weighted center for the lower body by using the centroid equation. In Figure 15, the computed body areas of the six body keypoints and the weighted center are represented visually alongside the yoga image [22].

Figure 15. Yoga pose area and weighted centers result by the proposed method.

4.3.2. Calculation of Area Difficulty

Area difficulty estimation is performed by considering the difference between the upper and lower body areas. First, the average area for each pose is calculated using Equation (7):

$$\mathrm{Average}\mathrm{Area}\mathrm{Difference}\left(\mathrm{AAD}\right)=={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|Upperbodyare{a}_i-Lowerbodyare{a}_i\right|$$

(7)

Subsequently, the set estimation method is used to determine whether the upper or lower body area is larger. If the upper body area is larger than the lower body area, it indicates that more effort is required to balance the body for upper body poses, making them more difficult. Conversely, if the lower body area is larger, it indicates a reliance on the legs for balance, which generally implies a simpler pose. The decision criteria are expressed in Equation (8):

$$\mathrm{Area}\mathrm{Difficulty}=\left\{\begin{array}{cc}0\hfill & \mathrm{i}f\mathrm{Area}\mathrm{Difference}\le \mathrm{AAD}\hfill \\ 50\hfill & \mathrm{i}f\mathrm{Area}\mathrm{Difference}>\mathrm{AAD}\mathrm{and}\mathrm{Lower}<\mathrm{Upper}\hfill \\ 100\hfill & \mathrm{i}f\mathrm{Area}\mathrm{Difference}>\mathrm{AAD}\mathrm{and}\mathrm{Upper}<\mathrm{Lower}\hfill \end{array}\right.$$

(8)

4.3.3. Calculation of Angle Difficulty

After the weighted centers of the upper and lower body are calculated, we compute the body balance line by using the coordinates of these two points. By connecting these two points and calculating the slope as demonstrated in Equation (9), we can determine the center of gravity balance for the body in a given yoga pose:

$$\mathrm{Slope}={\displaystyle \frac{y_{\mathrm{lower}\mathrm{centroid}}-{\mathrm{y}}_{\mathrm{upper}\mathrm{centroid}}}{{\mathrm{x}}_{\mathrm{lower}\mathrm{centroid}}-{\mathrm{x}}_{\mathrm{upper}\mathrm{centroid}}}}$$

(9)

Subsequently, we apply the slope-to-angle conversion formula to convert units into angles and design the angle difficulty estimation formula as expressed as Equation (10). This formula indicates that if the body is in a more vertical position during a yoga pose, indicating a stable posture, the difficulty level is lower. Conversely, if the body is more horizontal, more effort is required to maintain balance, thereby indicating a higher difficulty level:

$$\mathrm{Angle}\mathrm{Difficulty}=\left|\mathrm{Angle}\mathrm{degree}-90\right|.$$

(10)

4.3.4. Calculation of Body Support Point Difficulty

To perform yoga difficulty estimations, we set thresholds to detect keypoints in contact with the ground. Because pose difficulty is based on which body parts are used to support the entire body, we introduce a parameter called “base difficulty”. This parameter reflects the pattern of the body support points in contact with the ground. For instance, supporting the body solely with the hands is regarded as the most difficult because the hands are bearing the entire weight of the body in this pose. By contrast, supporting the body with the feet represents the lowest difficulty. The definition of base difficulty is provided in Equation (11):

$$\mathrm{Base}\mathrm{Difficulty}=\left\{\begin{array}{cc}10\hfill & \mathrm{if}\mathrm{pose}\mathrm{contact}\mathrm{type}\mathrm{is}“\mathrm{two}\_\mathrm{hand}\mathrm{and}\mathrm{two}\_\mathrm{leg}”\hfill \\ 25\hfill & \mathrm{if}\mathrm{pose}\mathrm{contact}\mathrm{type}\mathrm{is}“\mathrm{two}\_\mathrm{leg}”\hfill \\ 35\hfill & \mathrm{if}\mathrm{pose}\mathrm{contact}\mathrm{type}\mathrm{is}“\mathrm{single}\_\mathrm{hand}\mathrm{and}\mathrm{two}\_\mathrm{leg}”\hfill \\ 50\hfill & \mathrm{if}\mathrm{pose}\mathrm{contact}\mathrm{type}\mathrm{is}“\mathrm{single}\_\mathrm{leg}\mathrm{and}\mathrm{single}\_\mathrm{hand}”\hfill \\ 75\hfill & \mathrm{if}\mathrm{pose}\mathrm{contact}\mathrm{type}\mathrm{is}“\mathrm{single}\_\mathrm{leg}”\hfill \\ 100\hfill & \mathrm{if}\mathrm{pose}\mathrm{contact}\mathrm{type}\mathrm{is}“\mathrm{two}\_\mathrm{hand}”\hfill \\ 0\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(11)

4.4. Evaluation of Yoga Difficulty Level

With the parameters calculated from the body area, the definition of pose difficulty is established. To construct the difficulty evaluation formula, parameters such as base difficulty, angle difficulty, and area difficulty are combined in Equation (12) as follows:

$$\mathrm{Difficulty}=\mathrm{Base}\mathrm{Difficulty}+\mathrm{Angle}\mathrm{Difficulty}+\mathrm{Area}\mathrm{Difficulty}$$

(12)

Subsequently, we define the five difficulty levels as follows: beginner, novice, intermediate, advanced, and master. These levels are assigned on the basis of the numerical values within their respective ranges. Table 4 lists the five yoga pose difficulty levels as determined in accordance with the proposed method for reference. Table 5 provides examples of yoga pose difficulty assessments conducted using the method outlined in Table 4.

Table 4. Five yoga pose difficulty levels.

Difficulty Evaluation	Difficulty Levels
Difficulty ≤ 30	(Beginner, Lv.1)
30 ≤ Difficulty ≤ 60	(Novice, Lv.2)
60 ≤ Difficulty ≤ 90	(Intermediate, Lv.3)
90 ≤ Difficulty ≤ 100	(Advanced, Lv.4)
100 ≤ Difficulty	(Master, Lv.5)

Table 5. Examples of yoga pose difficulty levels in Table 4.

Input Yoga Image	Output	Difficulty Levels
Scale	101.4470	Master (Lv.5)
Tree	79.2968	Intermediate (Lv.3)
Half-Moon	93.9552	Advanced (Lv.4)
Lord of the Dance	95.4079	Advanced (Lv.4)
Garland	25.5112	Beginner (Lv.1)

5. Evaluation Results

During the evaluation phase, the collection and analysis results for yoga poses are obtained first. Next, by compared the difficulty estimations with those from the reference data, this section highlights and discusses the cases involving substantial discrepancies. Upon obtaining the objective yoga pose difficulty estimations, the assigned difficulty classifications are implemented in the user interface of the yoga self-practice system, marking the start of the realization phase of the present study.

5.1. Collection and Analysis of Yoga Pose Images

In the experiment, we first analyze five yoga poses in the yoga self-practice system (Figure 16). Each yoga posture is evaluated for multiple difficulty parameters.

Figure 16. Calculation results of five yoga pose images in [20], where the “Red”, “Green”, “Yellow”, “Dark blue”, and “Cyan” colors indicate the detected parts of the torso, left arm, right arm, left leg, and right leg, respectively.

The images from [22] are used to present the results of yoga pose difficulty estimations (Figure 17). In each yoga pose, 12 body keypoints are used to calculate the areas, centroids, and the body balance lines for the upper and lower body. On the basis of these analyses, active assessments of the poses are conducted. Yoga pose images in the dataset [22] where body keypoints cannot be detected are excluded to avoid estimation errors.

Figure 17. Calculation results of 16 yoga pose images in [22], where the “Red”, “Green”, “Yellow”, “Dark blue”, and “Cyan” colors indicate the detected parts of the torso, left arm, right arm, left leg, and right leg, respectively.

5.2. Results of Difficulty Level Estimations and Comparisons

After the various pose parameters are obtained, the proposed system assigns difficulty levels to the yoga poses. In the present study, the objective experimental results obtained through the proposed system are compared with the reference for yoga pose difficulty levels [22]. As revealed in Figure 18, the reference categorizes yoga poses by difficulty [22] into three levels, whereas the proposed system uses a five-point scale with endpoints ranging from 1 (beginner) to 5 (master). However, significant differences in the difficulty ratings for specific poses are identified. These are discussed in subsequent subsections using several examples.

Figure 18. Comparison of difficulty levels for yoga poses in [22].

Figure 19 compares the evaluation results from our prior work [19] with those obtained using the proposed design (Figure 19), revealing that the proposed method produces more objective results. These results also indicate that the five yoga poses evaluated in our prior work [19] are not difficult to perform and are therefore highly suitable for self-practice training by yoga beginners.

Figure 19. Comparison of difficulty levels for five yoga poses in our prior work [19].

The validation steps for this method also involve input from 10 users and a professional yoga instructor, who provide subjective opinions on these poses. The significant discrepancies in the results (Figure 20) are likely due to specific keypoints being obscured, leading to large estimation errors. However, the results closely align with the opinions of the professional yoga instructor (Figure 21), demonstrating the method’s reliability based on objective criteria.

Figure 20. Comparison of difficulty levels with 10 users.

Figure 21. Comparison of difficulty levels with professional instructor.

Figure 22 reveals the comparisons of difficulty estimation by the evaluations from reference [22], professional yoga instructors, users, and the proposed method. In this study, the error calculation is included to demonstrate the accuracy of the proposed method as shown in Equation (13). In (13), the calculation is performed by using the average subjective difficulty of the professional yoga instructors as the ground truth (GT), and the other three ones, including the reference subjective difficulty in [22], the average subjective difficulty of 10 users, and the proposed method, are used as the inputs of the “comparative difficulty”:

$$\mathrm{Percentage}\mathrm{Error}=|{\displaystyle \frac{\mathrm{Comparative}\mathrm{Difficulty}-\mathrm{Ground}\mathrm{Truth}\left(\mathrm{GT}\right)}{\mathrm{Ground}\mathrm{Truth}\left(\mathrm{GT}\right)}}|$$

(13)

Figure 22. Comparisons of difficulty estimation by the evaluations from reference [22], yoga instructors, users, and the proposed method.

By using the average subjective difficulty of the professional yoga instructors as the ground truth, the mean percentage error (MPE) is calculated using Equation (13), and the MPE values by the reference subjective difficulty in [22], the average subjective difficulty of 10 users, and the proposed method are 0.373, 0.307, and 0.438, respectively.

By analyzing the corresponding MPE results, although the MPE values between the proposed method and the subjective ground truth from professional yoga instructors are slightly higher than other reference benchmarks, the estimated difficulty levels by the proposed design are still within the acceptable range, and there is no significant deviation. The result reveals that the proposed method has stability and reference value, and this analysis further supports the performance consistency of the proposed model under different benchmarks and verifies the rationality and practicality of its design.

5.3. Implementation in User Interface of Yoga Self-Practice System

The present study implements the difficulty level estimation function for yoga poses in the yoga self-practice system. The corresponding difficulty levels are displayed in the user interface of the system (Figure 23). From left to right, the difficulty levels for the yoga poses Mountain, Side Bend, Warrior, Seat 1, and Seat 2 are categorized as beginner, beginner, novice, beginner, and beginner, respectively.

Figure 23. UI diagram of yoga self-practice system.

5.4. Significant Differences in Estimations

In the experimental results, several yoga poses (i.e., Cobra, Dolphin, Lord of Dance, Half Frog, and Warrior III) with significant differences in difficulty level ratings are identified (Figure 24). Because the evaluation of yoga pose difficulty in the present study considers additional factors, such as body angles and keypoints in contact with the ground, safety during yoga self-practice training is also a crucial indicator. To prevent yoga beginners from injuring themselves by executing incorrect or difficult poses during yoga exercises, specific yoga poses originally classified as suitable for beginners [22] are reclassified into higher difficulty levels (Figure 25).

Figure 24. Yoga poses for significant grade differences in [22].

Figure 25. Yoga poses for beginner levels in [22].

In Figure 25, the “Red” color indicates torso, the “Green” color indicates left arm, the “Yellow” color indicates right arm, the “Dark blue” color indicates left leg, and the “Cyan” color indicates right leg. The green boxes highlight the poses that, due to the requirements for support points or flexibility, have been reclassified into higher difficulty levels compared to their original categorization as simple poses. Compared with existing difficulty estimation techniques, the experimental results indicate that the proposed method provides more objective and consistent difficulty level estimations. In contrast to our prior work [19], which employed different approaches for the upper and lower body, the proposed method achieves a comprehensive and consistent evaluation by considering the entire body. To enhance the yoga self-practice system, the proposed yoga difficulty level estimation function enables users to clearly understand whether the selected yoga poses are suitable for them.

In this study, the proposed difficulty level estimation method of yoga poses can evaluate proper results by using the front body view of yoga poses from yoga teachers. If the OpenPose model detects the important selected body keypoints accurately by the side views of yoga poses, the proposed method can also evaluate the yoga pose difficulty level. However, when the yoga teacher turns his/her body to the side, there is a greater possibility that the selected body keypoints will not be detected due to the overlap of the limbs. Therefore, there are some possible solutions to solve this issue.

For the proposed yoga pose difficulty level estimation, only the human body keypoints of yoga instructors and yoga teachers are selected for the evaluation. When the videos or images of yoga poses from yoga teachers are collected, according to the posture changes of yoga teacher, the yoga teacher is automatically reminded to turn or adjust his/her yoga pose so that all important selected keypoints can be captured to detect by OpenPose correctly. For example, when the pose difficulty level estimation function detects that the yoga teacher turns their body to the side, the system can prompt the yoga teachers to adjust their body views, or the system can also change the capture angle and view of the camera. It means if the OpenPose model cannot detect the body keypoints of the side view accurately, the yoga poses of the front view or other angle side views from yoga teacher will be selected to be used for the yoga difficulty level evaluation.

When the yoga pose is just the crossing limbs but all selected keypoints can be detected by OpenPose, the proposed method can evaluate the yoga difficulty level properly. For example, in Figure 26, although the yoga pose has the condition of crossing limbs, the required keypoints are detected correctly, and the difficulty level of the yoga pose can be calculated. If some of the selected body keypoints are missing when the yoga teacher has the crossing limbs, it may be possible to infer the missing parts based on the existing visible body keypoints and the previous yoga actions, but this will involve a more complex inference estimation methodology. In addition, the OpenPose model can be considered to be re-trained by the image datasets of the failure-detected pose conditions, especially the yoga poses with crossing limbs.

Figure 26. Yoga pose with the crossing limbs, where the “Red”, “Green”, “Yellow”, “Dark blue”, and “Cyan” colors indicate the detected parts of the torso, left arm, right arm, left leg, and right leg, respectively.

6. Conclusions and Future Works

In the present study, a novel yoga pose difficulty level estimation method based on OpenPose is presented to define five difficulty levels for yoga poses. The proposed method effectively uses six keypoint areas of the upper and lower body, body support types, center of gravity calculations, and body tilt angles and slopes to produce estimations, thereby contributing considerably to the yoga self-practice system in aspects such as computation, data collection, method validation, comparison with existing designs, and implementation in embedded systems.

Given the benefits of the yoga pose difficulty level estimation method for the yoga self-practice system, future studies should enhance the formula definitions by incorporating body flexibility as a parameter, which is crucial for yoga poses and varies among individuals and age groups. The present study used nearly 100 datasets for evaluation and comparison. However, increasing the dataset size in future research could enhance the reliability of the assessment. By considering numerous yoga pose combinations, our goal is to improve the user-friendliness of the yoga self-practice system. The integration of artificial intelligence in daily life ensures that all age groups can easily perform and practice yoga poses. The proposed approach aims for a win–win scenario, promoting both health and technological advancement.