Aerobics movement quality assessment based on adaptive multi-scale ST-GCN model

Abstract

With the widespread application of deep learning technology in image analysis, visual-based data processing and action recognition have become current research hotspots. To improve the accuracy and real-time evaluation of the quality of aerobics actions, a new lightweight OpenPose network is developed to extract key points of athletes’ skeletons. At the same time, the adaptive multi-scale optimization is carried out on the spatiotemporal graph convolutional network to grasp the temporal characteristics of actions. Finally, a new aerobics action quality evaluation model is designed by integrating Siamese network. The accuracy of the proposed model reached 96.8%, which was 5.3% higher than existing methods. The highest action evaluation matching degree was 89.7%, the highest action improvement rate was 21.34%, the highest action similarity evaluation was 92.17%, and the shortest evaluation time was 0.71 seconds. Its efficiency and effectiveness in the detection and evaluation process are significantly higher than other advanced models. From this, the proposed method for evaluating the quality of aerobics actions has strong application potential while ensuring efficiency and accuracy, and can provide certain technical support for aerobics competitions and actions training.

Keywords

aerobics action evaluation OpenPose ST-GCN model skeleton nodes

Introduction

With the deep popularization of the concept of national fitness, aerobic exercise, as a fitness method that combines music rhythm and body movement, has been widely promoted and applied worldwide. Especially in the fields of sports competition, fitness exercise, and entertainment performance, aerobic exercise not only helps to improve the participants’ cardiorespiratory function, coordination, and strength endurance but also becomes an important criterion for assessing the level of motor skills in programs such as competitive aerobics.¹ In competitive sports, the execution of high-quality aerobic movements directly affects the score and performance of the competition, while in the field of public fitness and recreation, standardized and regulated movement instructions can effectively improve the effect of exercise and reduce the risk of sports injuries. Therefore, accurate assessment of aerobic movement quality not only has important competitive value but also has a wide range of social application needs. However, at this stage, aerobic movement quality assessment still mainly relies on the subjective judgment of professional referees or coaches, which has the problems of inconsistent judgment standards, strong influence of subjective factors, and limited real-time feedback capability. With the development of intelligent sports and sports health field, how to realize the objective, efficient, and real-time assessment of aerobic movement with the help of advanced technology has become an important topic to be solved.² Especially in the fast and changeable scenarios with rich movement details such as competitive aerobics, the traditional methods have significant deficiencies in the recognition of movement details, dynamic coherence capture, and complex posture adaptability, which seriously constrain the standardization and automation process of movement quality assessment. To address this problem, many researchers at home and abroad have successively explored it. Liu Q believes that the traditional aerobic exercise recognition judgment model has the challenges of limited cost and insufficient recognition accuracy, for this reason, the scholar proposes a novel recognition judgment model that integrates Convolutional Neural Networks (CNN) and long and short-term memory networks. The experimental results show that the model recognizes fewer training parameters and has higher subsequent judging accuracy, and performs well in static or short-time action recognition, but there are still some limitations in dealing with complex dynamic action sequences and scenarios with high real-time requirements.³ Lv D et al. In order to improve the quality assessment effect of aerobics exercise, the researchers combined the infrared target imaging system and monocular detection algorithm to propose a novel assessment model. The experimental results show that the model can effectively track the athlete’s body posture and accurately assess the quality of movement with good tracking accuracy, but it is sensitive to the environmental lighting conditions and weakly adapted in multi-character interaction scenarios.⁴ Cao T attempted to enhance the effectiveness of wireless sensor networks in aerobics training posture assessment in the scholar combined the deep learning method and multiple inertial measurement units to The scholar combined the deep learning method and multiple inertial measurement units to construct an aerobics training posture evaluation system. The experimental results show that the system has lower classification error and higher real-time performance than the existing methods, and is suitable for single training in controlled environment, but it is more equipment-dependent, which restricts its generalization to large-scale free-motion scenarios.⁵ Zuo N et al. In order to overcome the problem of low success rate of tracking and decomposition of the traditional aerobics movement, the researchers combined the binary classifiers and convolution filters, and proposed a new action evaluation model. A novel movement evaluation model. The experimental results show that the method has a significant improvement in tracking and decomposition success rate compared with the traditional method, which is suitable for fast processing of action segments, but there is still room for improvement in the modeling of action continuity and the capture of detailed features.⁶ Zhang Y proposed a frame bounding box smoothing non-maximum suppression strategy to strengthen the stability of aerobic quality assessment, which integrates multiple prediction frames by integrating the information to obtain the results. Experimental results show that the method effectively improves the stability and tracking quality of motion quality assessment localization, and performs well in scenes with high action coherence, but there are still some limitations on the adaptability to drastic posture changes and abnormal movements.⁷ Through skeletal joint point extraction, spatio-temporal feature modeling, and action similarity analysis, deep learning methods are able to more accurately capture the fine changes in locomotor actions and achieve automated, real-time action evaluation. However, there are still many limitations in the existing deep learning-based assessment methods, such as the high consumption of computational resources in the skeleton feature extraction phase, insufficient temporal modeling of complex dynamic actions, and the lack of efficient matching mechanism for action quality determination, which leads to the difficulty of balancing the assessment accuracy and real-time performance, and the robustness of the assessment in the diversified action scenarios needs to be improved.

With the rapid development of deep learning and Graph Convolutional Network (GCN), action recognition methods on the basis of GCN have become one of the mainstream technologies for processing complex spatiotemporal data. Among them, the Spatial Temporal Graph Convolutional Network (ST-GCN) combines the characteristics of GCN and recurrent neural networks, and can effectively handle the spatial structure of human joints and the temporal dependence of action sequences.⁸ Hou Z et al. in order to explore the application of dynamic analysis algorithms in sports training injury risk assessment, a physical training injury risk assessment model was proposed based on ST-GCN with the introduction of adaptive graph convolution and residual channel attention. The experimental results showed that the model improved the accuracy by 1.94% and the F1 value by 4.3% compared with the original ST-GCN model, which demonstrated a good assessment effect in specific training scenarios, but there was still a certain response latency when dealing with high-speed dynamically changing data.⁹ Xu Y et al. in order to solve the limitations encountered by individual visual sensors in capturing motion data, the researchers combined the ST-GCN after proposing a novel multi-view skeleton data detection method. Experimental results show that the method significantly improves the smoothness and accuracy of skeleton data, improves pose recognition, and is suitable for multi-angle acquisition environments, but still suffers from high consumption of computational resources in large-scale real-time applications.¹⁰ Chen H et al. found that most of the existing methods use the 3D coordinates of the joints to implicitly demonstrate the dynamic motion context, which lacks the key information of the pose sequence. For this reason, the research team proposed a dual-stream augmented human motion pose detection model in combination with ST-GCN. Experimental results show that the model achieves advanced performance in both qualitative and quantitative evaluations, especially in fine-grained action recognition, but the network structure is complex, and the training time and hardware requirements are high.¹¹ Enkhbat A et al. proposed an ST-GCN-based approach for simultaneous human behavior recognition and note recognition in music performances. Experimental results show that the model achieves an accuracy of 81.4% in a specific task, which is better than the traditional ST-GCN model, and demonstrates good simultaneous detection of actions and musical notes, but the scope of application is relatively limited, and the generalization ability still needs to be strengthened.¹² Janbi N F et al. came up with a unique ST-GCN violence detection model based on an integrated spatio-temporal skeleton, which automatically uses spatial and temporal data to detect violence in surveillance videos. The experimental results show that the model has a maximum accuracy of 93%, which is more than 10% higher than the traditional model, and is effective in violence detection scenarios, but further validation of its generalizability in other action category generalization applications is still needed.¹³ A comparison of the above literature review is shown in Table 1.

Table 1.

Comparison of different literature reviews.

Researcher	Method	Achievements	Limitations	References
Liu Q	Aerobics recognition and evaluation model integrating CNN and LSTM	Reduced training parameters and improved evaluation accuracy for static or short-term movements	Limited ability to handle complex dynamic sequences; insufficient real-time performance	[3]
Lv D et al.	Action evaluation model based on infrared target imaging and monocular detection	Improved posture tracking accuracy; suitable for specific lighting environments	Sensitive to lighting changes; poor adaptability in multi-person interactive scenes	[4]
Cao T	Aerobics posture evaluation system based on wireless sensors and deep learning	Enhanced classification accuracy and real-time performance; suitable for single-person controlled training	Strong dependence on devices; difficult to extend to large-scale free-motion scenarios	[5]
Zuo N et al.	Action tracking and decomposition model based on binary classifier and convolution filtering	Significantly improved tracking and decomposition success rate; fast processing of action segments	Insufficient modeling of motion continuity; limited detail capture	[6]
Zhang Y	Frame bounding box smoothing non-maximum suppression strategy	Enhanced stability of motion localization and improved tracking quality	Limited adaptability to violent posture changes	[7]
Hou Z et al.	Injury risk evaluation model based on ST-GCN with adaptive graph convolution and residual attention	Accuracy improved by 1.94%, F1 score increased by 4.3%	Delayed response to high-speed dynamic changes; real-time performance still insufficient	[9]
Xu Y et al.	Multi-view skeleton data detection method based on ST-GCN	Improved skeleton smoothness and posture recognition accuracy	High computational cost for large-scale real-time applications	[10]
Chen H et al.	Dual-stream enhanced ST-GCN human motion posture detection model	Improved performance in both qualitative and quantitative evaluation; excellent fine-grained motion recognition	Complex network structure; high training time and hardware resource consumption	[11]
Enkhbat A et al.	ST-GCN-based human behavior and musical note synchronization recognition method	Achieved 81.4% accuracy in specific tasks; good synchronization detection performance	Limited generalization ability; narrow application scope	[12]
Janbi N F et al.	Ensemble spatiotemporal skeleton-based ST-GCN violence detection model	Violence detection accuracy improved to 93%, over 10% higher than conventional methods	Generalization ability to other action categories needs further verification	[13]

In summary, despite the improvement in the accuracy of aerobics quality assessment, these methods still have certain limitations, including insufficient adaptability to complex dynamic movements and challenges in real-time. In order to solve these problems, an efficient, accurate, and widely adaptable aerobics movement quality assessment model is constructed. The study proposes an innovative movement quality assessment method based on deep learning techniques. Specifically, compared with the existing techniques, the study significantly reduces the number of model parameters and inference latency and improves the real-time processing capability by introducing a lightweight OpenPose network while maintaining the accuracy of skeleton key point extraction; on this basis, an adaptive multi-scale spatio-temporal graph convolution network (AMS-ST-GCN) is proposed to dynamically adjust the spatial relationship modeling between skeleton joints, and fully captures different levels of movement change features through multi-scale temporal convolution; in addition, the fusion twin network (Siamese Network, SN), which realizes efficient matching and quality quantification between standard and actual movements in the movement feature space, fills the gap of traditional movement recognition methods that make it difficult to accurately assess the movement accuracy and fluency. The study aims to provide an efficient and accurate solution for quality assessment in aerobics, and also provides a reference for action recognition and assessment in other fields.

Methods and materials

To improve the quality evaluation accuracy of aerobics actions, an aerobics action analysis framework based on lightweight OpenPose and adaptive multi-scale ST-GCN is built. First, the lightweight OpenPose network is applied to extract the skeletal features of athletes, capturing changes in joint positions and skeletal structures. Subsequently, an adaptive multi-scale ST-GCN model is adopted to enhance the spatiotemporal feature extraction capability, to comprehensively reflect the complexity of actions and athlete performance. Finally, combined with SN, the current action is compared with the standard action library to evaluate the similarity of actions, and the action quality score is optimized through a spatiotemporal error minimization comparison loss function.

Skeleton feature extraction of aerobics athletes based on OpenPose

The quality assessment of aerobics actions relies on the accurate capture and analysis of athletes’ actions. Skeleton feature extraction can provide position information of action joints and changes in skeleton structure during dynamic and complex action processes, helping to analyze athlete posture, action accuracy, and coordination.^14,15 OpenPose, as an efficient multi-body pose estimation algorithm, can extract real-time and accurate joint position information of athletes from images or videos, and generate corresponding skeleton structure maps. OpenPose uses CNN to process images, locate key points, and generate skeleton maps through the connections between key points. The processing results of each frame of the image include the 2D or 3D coordinates of various joints of the athlete, such as shoulders, knees, ankles, etc.^16–18 However, in practical applications, due to the dynamic characteristics and rapid changes of aerobics actions, the skeletal information of a single frame image is not sufficient to fully characterize the athlete’s performance throughout the entire action process. In addition, due to the rapid changes in athletes’ posture, factors such as joint occlusion, posture changes, and interactions between athletes may affect the accuracy of skeleton extraction.^19–21 Therefore, a lightweight OpenPose network model is proposed, and its structure is shown in Figure 1.

Figure 1.

Lightweight OpenPose network model structure.

In Figure 1, the entire lightweight OpenPose receives video action information from the MircoNet network, and then extracts the action information into inference feature 1 and inference feature 2 through 5 rounds of 7×7 convolution operations. During this period, the Loss function is used to predict the confidence level of skeletal joints, partial affinity domain fields, and image features of the Visual Geometry Group (VGG19) in the output image. After completion, deeper convolutional feature extraction is performed in the pose refinement stage. Similarly, inference features 3 and 4 are extracted. Compared to the traditional OpenPose network model, this lightweight OpenPose uses a more lightweight MircoNet network to front-end VGG19, reducing running parameters, distinguishing channels and regions more finely during runtime, and reducing computational resources. Second, the pose refinement stage is used to replace the dual branch multi-stage CNN repeated convolution extraction, reducing redundant calculations. The network unit structure of MircoNet is displayed in Figure 2.^22–24

Figure 2.

Network unit structure of MircoNet.

As shown in Figure 2, MircoNet consists of three basic units, namely Mirco-A, Mirco-B, and Mirco-C. Among them, Miro-A consists of two layers of Maximum Dynamic Displacement Activation Function (MDDAF), one layer of Micro-decomposition Point Convolution (MPC), and one layer of Micro-decomposition Deep Convolution (MDC). Miro-B consists of three layers of MDDAF, two layers of MPC, and one layer of MDC. The structure of Mirco-C is similar to that of Mirco-B, but the decomposition between layers is done through conventional connections. Compared to Mirco-A and Mirco-B, conventional connections can reduce computational complexity in simpler or resource limited situations. Assuming $I$ is the input image, the output of the $k$ -th layer after 5 convolution operations is $F_{k}$ , and the calculation process is shown in equation (1)

F_{k} = χ (F_{k - 1}, W_{k}) + b_{k}

(1)

In equation (1),

F_{k}

represents the feature map processed by the

k

-th layer convolution.

χ (_)

represents convolution operation.

F_{k - 1}

signifies the input feature map.

W_{k}

signifies the

k

-th convolutional kernel.

b_{k}

signifies the bias term of the

k

-th layer. After each convolution operation, the dimensionality of the feature map gradually decreases, but the spatial information and joint features contained therein are gradually extracted. Next, by further processing the features of the skeleton joints, the network predicts the confidence map for each joint. Assuming the joint confidence map output by the network is

C_{j}

, the calculation process is shown in equation (2)

C_{j} = σ (W_{C} \cdot F_{k} + b_{C})

(2)

In equation (2),

C_{j}

signifies the confidence level of the

j

-th joint, that is, the reliability of predicting the position of that joint.

σ

signifies the Sigmoid activation function.

W_{C}

signifies the weight matrix.

b_{C}

signifies the bias term. In addition, the partial affinity domain field represents the spatial relationship between joints, which can be predicted by CNN, as shown in equation (3)^25,26

P A F_{i j} = χ (F_{k}, W_{P A F})

(3)

In equation (3),

P A F_{i j}

represents the affinity field between joint

i

and joint

j

W_{P A F}

represents the convolution weight of the affinity domain field. By calculating the confidence and affinity fields, the model obtains the position and spatial relationship of the joints. At this point, global optimization and pose refinement are used to maximize the combination of joint position confidence and affinity field, resulting in the final pose prediction result, as shown in equation (4)^27,28

{\begin{cases} P_{f i n a l} = \arg \max_{P_{j}} \sum_{i, j} C_{i} \cdot P A F_{i j} \\ P_{j}^{'} = P_{j} + Δ P_{j} \end{cases}

(4)

In equation (4),

P_{f i n a l}

represents the final optimized skeleton joint position.

\arg \max_{P_{j}}

represents the optimal joint position obtained by combining the maximum joint confidence with the affinity domain field.

P_{j}

and

P_{j}^{'}

represent the initial and adjusted joint positions, respectively.

Δ P_{j}

represents the adjustment amount calculated during the pose refinement stage. In the feature extraction, although the MircoNet reduces the parameters of the model, it also reduces the retrieval performance. Especially, face with a large and complex amount of image data, its nonlinear properties are more exposed.²⁹ For this purpose, the MPC is introduced in the study. This method preserves more representative features through micro-decomposition of convolutional layers, while reducing the interference of nonlinear effects and improving the inference ability of the network in complex scenes.^30,31 The principle is displayed in Figure 3.

Figure 3.

Schematic diagram of micro-decomposition point convolution.

As shown in Figure 3, MPC decomposes high-dimensional calculations into multiple low dimensional operations through data permutation matrices, greatly reducing the model’s demand for computational resources. MDC is decomposed into multiple convolution kernels, which reduces the complexity. The MPC and MDC are connected. By combining micro-decomposition methods, the number of channels in the spatial filter of the convolution operation is first increased, thereby expanding the coverage range of the receptive field. Assuming the size of the input convolution kernel is $k \times k$ , the input feature map is $X \in R^{H \times W \times C_{i n}}$ , and the output feature map is $Y \in R^{H \times W \times C_{o u t}}$ , the combined point convolution at this time is shown in equation (5)

Y_{i, j} = \sum_{c = 1}^{C_{i n}} \sum_{m = 1}^{k} \sum_{n = 1}^{k} X_{i + m, j + n, c} \cdot K_{m, n, c}

(5)

In equation (5),

Y_{i, j}

signifies the

i, j

-th pixel value of the output feature map.

X_{i + m, j + n, c}

represents the pixel values of the input feature map.

i, j

represents the index of a specific location.

m

and

n

represent the local region positions of the convolutional kernel.

c

represents the input channel.

X_{i + m, j + n, c}

represents the convolution kernel at the

m, n

-th element of the

c

-th channel.

C_{i n}

represents the number of channels in the input feature map. The decomposed output feature map is shown in equation (6)

Y_{s p l i t} = \sum_{i = 1}^{k} \sum_{j = 1}^{k} \sum_{c = 1}^{C_{i n}} X_{i, j, c} \cdot W_{i, j, c}

(6)

In equation (6),

Y_{s p l i t}

represents the output feature map obtained through MPC.

W_{i, j, c}

represents the split convolution kernel.

Construction of aerobics action extraction and quality evaluation model based on adaptive multi-scale ST-GCN

After constructing the lightweight OpenPose network model, the lightweight OpenPose network can successfully extract skeletal features of aerobics athletes, and obtain joint position information and changes in skeletal structure. However, to achieve precise extraction and quality evaluation of aerobics actions, relying solely on static skeletal features is not sufficient to fully reflect the complexity of the actions and the performance of the athletes.^32,33 Therefore, the study further introduces the ST-GCN model. The spatiotemporal feature maps of athlete skeletons in the ST-GCN model and conventional Kinect are displayed in Figure 4.³⁴

Figure 4.

Skeleton spatiotemporal feature maps of ST-GCN and conventional Kinect. (a) Kinect (b) ST-GCN.

Figure 4(a) shows the spatiotemporal feature map of the human action skeleton in the conventional Kinect. Figure 4(b) shows the spatiotemporal feature map of the human action skeleton in the ST-GCN model. In Figure 4, ST-GCN can effectively capture the spatiotemporal features of athlete skeletons through graph convolution and temporal convolution operations. Unlike ST-GCN, the skeleton extraction method adopted by Kinect mainly relies on depth sensors to obtain the three-dimensional position coordinate information of athletes. Although it can accurately obtain joint positions in space, it is relatively simple in temporal processing. To this end, an Adaptive Multi-Scale Spatial Temporal GCN (AMS-ST-GCN) model is proposed. The structure of AMS-ST-GCN is shown in Figure 5.

Figure 5.

Structure of AMS-ST-GCN.

In Figure 5, the AMS-ST-GCN mainly includes Adaptive Semantic Graph Convolutional Networks (AS-GCN), Multi-scale Time Graph Convolution (M-TCN), and Adaptive Multi-channel Graph Convolutional Networks (AM-CGN). In the process, first, the input human skeletal data is modeled as a spatio-temporal graph, where nodes represent joint point locations and edges connect spatially adjacent joints to the same joint trajectory in time. The model uses a Spatial-Temporal Graph Convolution (STGC) structure for feature extraction in the spatial and temporal dimensions, respectively. Spatial convolution is used to capture the local relationships between joint points, while temporal convolution is used to model the dynamic properties of action evolution over time. To further enhance the feature representation, the study introduces an adaptive neighbor matrix mechanism based on the traditional ST-GCN. Compared with the fixed-skeleton topology, the improved model is able to dynamically adjust the connection strength between nodes according to the distribution characteristics of the input data, thus enhancing the adaptability to different action patterns and individual differences. In addition, the model is designed with multi-scale convolutional branches, and different branches perform convolutional operations on the first-order, second-order, and farther-neighborhoods, which can effectively fuse the local fine-grained action features with the global action trend information. Through adaptive neighbor optimization and multi-scale feature fusion, the overall network shows better results in detail discrimination and action dynamic modeling. Finally, after feature fusion, the output layer completes the action category evaluation and score prediction based on the extracted multiscale spatio-temporal features, which provides reliable support for downstream performance analysis. For a given skeleton graph, the graph convolution operation of AS-GCN is displayed in equation (7)

H^{(k + 1)} = σ (\hat{A} H^{(k)} W^{(k)})

(7)

In equation (7),

H^{(k)} \in R^{N \times D_{k}}

represents the node feature matrix of the

k

-th layer.

N

represents the number of joints.

D_{k}

represents the feature dimension.

W^{(k)}

signifies the weight matrix of the

k

-th layer. Next, the M-TCN module processes temporal information and combines multi-scale convolution kernels to capture temporal features. The temporal convolution operation is shown in equation (8)

Z_{t} = \sum_{k = 1}^{K} X_{t - k} * W_{k}

(8)

In equation (8),

Z_{t}

represents the output feature at time step

t

X_{t - k}

represents the input feature at time step

t - k

. Finally, the AM-CGN module introduces the multi-channel graph convolution operation and performs information fusion through an adaptive weighting mechanism. Assuming that in multi-channel graph convolution, the convolution output corresponding to each channel is

H^{(c)}

, as shown in equation (9)

H_{o u t} = \sum_{c = 1}^{C} α_{c} \cdot H^{(c)}

(9)

In equation (9),

H^{(c)}

represents the graph convolution output of the

c

-th channel.

α_{c}

represents the adaptive weighting coefficient.

H_{o u t}

represents the final output feature after fusion. Accurate assessment of actions is a key task in the analysis of aerobics actions. It not only requires identifying the basic types of actions but also evaluating the accuracy, coordination, and fluency of their execution. The study introduces an action similarity evaluation model based on SN. This model compares the action characteristics of aerobics athletes with a standard action library by constructing a SN structure, in order to evaluate the similarity between the current action and the standard action. The advantage of SN lies in its ability to handle high-dimensional feature matching problems, especially suitable for handling complex and ever-changing action data.^35,36 The structure is presented in Figure 6.

Figure 6.

Structure of action similarity evaluation model based on SN.

As shown in Figure 6, the model is mainly composed of two sub-networks that share weights. These two sub-networks, respectively, receive input action features, process them through the same network structure, and generate corresponding action embedding vectors. Then, by calculating the similarity between the embedding vectors of these two actions, the model can evaluate the similarity between the current action and the standard action. Specifically, the input action features first pass through a feature extraction layer, which encodes the spatiotemporal features extracted by adaptive multi-scale ST-GCN, mapping the spatial and temporal features of the action to a low dimensional space. Subsequently, the two sub-networks share the same convolutional and fully connected layers, ensuring that the comparison of action features during processing can be performed in the same parameter space, improving the consistency and generalization ability. Specifically, an adaptive weighting mechanism is used to generate embedding vectors for each action. These embedded vectors capture the spatiotemporal features of actions. In order to measure the similarity between two action features, a weighted similarity measurement formula is proposed, as shown in equation (10)

S (v_{1}, v_{2}) = \sum_{o = 1}^{D} ω_{i} \cdot \frac{{(v_{1, i} - v_{2, i})}^{2}}{{‖ v_{1} ‖}_{2} {‖ v_{2} ‖}_{2}}

(10)

In equation (10),

S (v_{1}, v_{2})

represents the weighted similarity between actions

v_{1}

and

v_{2}

v_{1}

and

v_{2}

represent action embedding vectors generated through SN.

D

represents the feature dimension.

v_{1, i}

and

v_{2, i}

signify the elements of

v_{1}

and

v_{2}

in the

i

-th dimension, respectively.

ω_{i}

represents adaptive weights. The similarity measure of temporal consistency is shown in equation (11)

T (v_{1}, v_{2}) = \sum_{t = 1}^{T} [\sum_{k = 1}^{K} α_{k} \cdot c o n_{k} (v_{1} (t), v_{2} (t))] \cdot \frac{{‖ v_{1} (t) - v_{2} (t) ‖}_{2}}{{‖ v_{1} (t) ‖}_{2} + {‖ v_{2} (t) ‖}_{2}}

(11)

In equation (11),

T (v_{1}, v_{2})

represents the temporal similarity between actions and

v_{1}

and

v_{2}

, respectively.

v_{1} (t)

and

v_{2} (t)

represent the action embedding vectors of actions

v_{1}

and

v_{2}

at t

t

c o n_{k}

signifies the convolution operation of multi-scale convolution kernels at

t

α_{k}

signifies the adaptive weight of each convolution kernel. Finally, in order to train the network to effectively distinguish between similar and dissimilar actions, a contrastive loss function that minimizes spatiotemporal errors is proposed. This loss function not only considers the global similarity of actions but also introduces spatiotemporal error terms to ensure comprehensive optimization of temporal and spatial features. The calculation is shown in equation (12)

L_{c o n t r a s t i v e} = \frac{1}{2} [\begin{array}{l} y \cdot d^{2} (v_{1}, v_{2}) + (1 - y) \cdot \max {(0, m - d (v_{1} - v_{2}))}^{2} + \\ λ \cdot \sum_{t = 1}^{T} ({‖ v_{1} (t) - v_{2} (t) ‖}_{2}^{2}) \end{array}]

(12)

In equation (12),

L_{c o n t r a s t i v e}

represents the contrastive loss function.

y

represents the label. If

y = 1

, it means that the two actions are similar. If

y = 0

, it means that the two actions are not similar.

d (v_{1} - v_{2})

represents the Euclidean distance between two action embedding vectors.

m

represents a constant threshold used to distinguish between similar and dissimilar pairs of actions.

λ

represents the regularization parameter of the spatiotemporal error term, used to balance the error terms in both spatial and temporal dimensions. In summary, a new model for evaluating the quality of aerobics actions is proposed, and the process of this model is shown in Figure 7.

Figure 7.

New aerobics action quality evaluation model flow.

As shown in Figure 7, firstly, the lightweight OpenPose network extracts the skeletal joint features of athletes and obtains the spatiotemporal coordinate information of each joint during the action process. Next, the adaptive multi-scale ST-GCN captures the spatiotemporal dynamic features of the skeleton, enhancing the recognition ability of action details and coherence. Then, SN is used to compare the athlete’s current action with the standard action library. The similarity of actions is evaluated by calculating the similarity between embedded vectors. Finally, by introducing a spatiotemporal error minimization contrastive loss function, the model is optimized, taking into account the accuracy, coordination, and fluency of the actions. The final output is the athlete’s action quality score, which guides the athlete to improve their action execution.

Results

The study first constructs an experimental environment and conducts optimization tests on key hyper-parameters, with action accuracy and evaluation time as the core indicators for optimization. Multiple performance tests are conducted using two types of publicly available datasets, including ablation testing, comparative testing, and action feature extraction evaluation. At the same time, multiple advanced models are introduced for comparative testing, with action improvement rate, action similarity, and action quality as indicators, to verify the superiority and reliability in existing methods.

Performance testing of the quality evaluation model for aerobics actions

To verify the performance of the new aerobics action quality evaluation model, two publicly available datasets, Human3.6 M dataset and Carnegie Mellon University Motion Capture Dataset (CMU MoCap), are used as test data sources. Human3.6 M and CMU MoCap datasets, as standard datasets widely used in human gesture recognition and motion capture research, are characterized by high frame rate, high annotation accuracy, and rich coverage of motion types. Among them, Human3.6 M covers basic whole-body movements such as walking, sitting, clapping, avoiding, etc., while CMU MoCap has a large number of continuous dynamic sequences, which is closer to the natural movement state. Thus both provide skeletal motion information for aerobic exercise assessment that can be used for generic modeling. It should be noted that the movement samples in these datasets are mostly recorded in controlled environments and lack the diversity and complex background interference in real fitness scenarios, which may limit the generalization ability of the models in real aerobic training assessment. The detailed testing data environment and parameters are presented in Table 2.

Table 2.

Experimental environment and parameter configuration.

Experimental equipment	Value
CPU	Intel Core i9-10900K
GPU	NVIDIA RTX 3090
Memory	32 GB
Graphics memory	24 GB GDDR6X
Development environment	Windows 10, Python 3.9
Programming tools	PyTorch 1.6.0, OpenCV 4.5.2
Initialise learning rate	0.001
Learning rate batch size	32
Momentum parameters	0.7
Training period	150 epochs

The study first conducts value selection tests on two types of hyper-parameters that have a significant impact on the final model performance, namely the number of channels $C_{i n}$ in the input feature map and the action adaptive weight $ω_{i}$ , in order to maintain optimal model performance and facilitate effective testing in the future, as displayed in Figure 8.

Figure 8.

Hyper-parameter selection test. (a) Test the number of channels in the input feature map (b) Action adaptive weight test.

Figure 8(a) shows the channel count test of the input feature map. Figure 8(b) displays the action adaptive weight test. In Figure 8(a), the Loss value of the model shows an overall decreasing trend as the number of input feature map channels is gradually increased from 32 to 128. Specifically, when the number of channels is set to 32, the Loss value is initially about 3.5, and then decreases to about 2.7 with training; when the number of channels is increased to 64, the Loss value further decreases to 2.0; and when the number of channels reaches 128, the Loss value decreases to about 0.5, which shows a more significant convergence effect. When continuing to increase the number of channels, the decrease in the Loss value tends to slow down, while the model training time increases significantly and the computational overhead is aggravated. From Figure 8(b), it can be seen that there are differences in the Loss decline rate of the model under different action adaptive weights. The larger the weight value, the faster the Loss value decreases and the final convergence value is lower. Among them, when the weight is set to 0.8, the final Loss value of the model is the lowest, which is about 0.6; and when the weight is 0.2, the final Loss value is only reduced to about 1.4, which is not ideal for both convergence speed and effect. From a comprehensive point of view, a reasonable increase in the number of channels of the feature map can effectively improve the model’s ability to perceive the details of the action, but the computational burden brought by too many channels should not be ignored; at the same time, a moderate adjustment of the action adaptive weights can help to optimize the balance of the features between the actions during the training process, which can further reduce the Loss value and improve the overall training effect. The ablation test is conducted on the proposed model, as displayed in Figure 9.

Figure 9.

Ablation test results. (a) Human3.6M (b) CMU MoCap.

Figure 9(a) displays the ablation test results in the Human3.6 M dataset, and Figure 9(b) displays the ablation test results in the CMU MoCap dataset. As shown in Figure 9(a), the OpenPose module alone exhibited low iteration accuracy, only 85.3%. After implementing lightweight improvements, the iterative efficiency of the model significantly was increased by about 4%. After adaptive multi-scale improvements to ST-GCN, the detection accuracy of Light Weight OpenPose-AMS-ST-GCN reached 91.4%. After integrating the SN module, it increased again to 93.7%. According to Figure 9(b), its data performance was similar to Figure 9(a). After sequentially improving OpenPose and ST-GCN, the detection accuracy of the model reached about 95.1%. After integrating the SN module, the final performance of the model remained stable at 96.8%. Although the number of iterations slightly increased, the model performance was effectively improved. This indicates that the introduction and improvement of each module in the study are effective and significant. Advanced action detection and evaluation models are compared, such as Pose Residual Network (PoseResNet), Self-Supervised Pose Network (SPIN), and High-Resolution Network (HRNet), as presented in Table 3.

Table 3.

The results of multi-index testing of the model.

Environment	Model	Precision/%	Recall/%	F1/%	Average processing time/s
Low light	PoseResNet	81.57	78.42	79.97	0.84
	SPIN	83.69	80.31	81.96	0.76
	HRNet	85.41	82.27	83.81	0.72
	Our model	91.82	89.64	90.72	0.61
Normal light	PoseResNet	84.26	80.85	82.52	0.83
	SPIN	85.93	83.17	84.53	0.74
	HRNet	87.48	85.24	86.35	0.69
	Our model	93.14	91.76	92.44	0.58
High light	PoseResNet	80.37	77.51	78.91	0.85
	SPIN	82.26	79.03	80.61	0.79
	HRNet	84.12	81.47	82.77	0.74
	Our model	90.85	88.29	89.56	0.59

As can be seen in Table 3, under different lighting environments, the proposed model significantly outperforms the comparison methods in the three metrics of Precision, Recall and F1 value, and keeps the lowest average processing time. Specifically, in low-light environment, Our model’s Precision reaches 91.82%, Recall is 89.64%, and F1 value is 90.72%, which are higher than that of PoseResNet, SPIN, and HRNet, and the average processing time is only 0.61 seconds, which reflects good real-time performance and robustness. Under normal lighting conditions, the indexes of our model are further improved, in which the F1 value reaches 92.44%, which is more than 6% higher than that of HRNet, and the average processing time is shortened to 0.58 seconds, which verifies the stability and high efficiency of the model in the standard environment. And in the strong light environment, despite the overall recognition difficulty, Our model still maintains 90.85% Precision and 89.56% F1 value, the performance is ahead of other baseline methods, and the inference time is controlled within 0.59 seconds. In summary, the research method shows high accuracy and excellent processing efficiency in action recognition under different light intensities, and has good scene adaptability and practical promotion potential.

Simulation testing of the quality evaluation model for aerobics actions

To verify the effectiveness in action recognition detection and quality evaluation, 8 types of basic aerobics actions are randomly selected from the Human3.6 M dataset for testing. The selected actions are displayed in Figure 10.

Figure 10.

8 types of basic actions of aerobics. (a) Propelling forward (b) Upthrow (c) Forward downward (d) Forward upward lift (e) Lateral raise (f) Side down (g) Side upward (h) Backward downward lift.

Figure 10(a) shows the propelling forward action, Figure 10(b) shows the upthrow action, Figure 10(c) shows the forward downward action, Figure 10(d) shows the forward upward lift action, Figure 10(e) shows the lateral raise action, Figure 10(f) shows the side ward action, Figure 10(g) shows the side upward action, and Figure 10(h) shows the backward downward lift action. Combining the above 8 types of aerobics actions, taking the Percentage of Correct Keypoints (PCK) as the detection indicator, as displayed in Figure 11.

Figure 11.

PCK detection results of 8 types of actions by different models. (a) PoseResNet (b) SPIN (c) HRNet (d) Our model.

Figure 11(a) shows the PCK detection results of 8 types of actions in the PsoeResNet model, SPIN model, PsoeResNet model, and the proposed model. From Figure 11, the proposed model performed well in PCK detection results, especially in actions such as propelling forward, upthrow, and lateral raise, with significantly higher PCK values than other models. For example, in the “propelling forward” action, the PCK of the proposed model reached 95.2%. In contrast, the PoseResNet model accounted for only 85.2%, the SPIN model for 88.5%, and the HRNet model for 91.3%. This gap was also reflected in other actions, and the proposed model maintained high accuracy in most actions, especially in the complex actions of “forward upward lift” and “side upward,” with PCK of 92.5% and 93.3%, respectively. The proposed model has stronger robustness and accuracy in handling details and complex actions. Compared to traditional models such as PoseResNet and SPIN, it can better capture key point information and adapt to the action characteristics of different athletes. The study randomly selects four types of actions and tests the action evaluation matching degree of the four evaluation models in the process of action feature extraction and matching. The results are shown in Figure 12.

Figure 12.

The matching test results of action evaluation of different models. (a) Propelling forward (b) Upthrow (c) Lateral raise (d) Backward downward lift.

Figure 12(a) shows the evaluation matching results of different models in the propelling forward action, Figure 12(b) shows the evaluation matching results of different models in the upthrow action, Figure 12(c) shows the evaluation matching results of different models in the lateral raise action, and Figure 12(d) shows the evaluation matching results of different models in the backward downward lift action. As shown in Figure 12, the proposed model performed outstandingly in the evaluation matching test of four types of aerobics actions. In the “propelling forward” action, the matching degree of the model reached 89.7%, significantly higher than PoseResNet’s 84.2%, SPIN’s 76.6%, and HRNet’s 80.2%. Similarly, in the upthrow action, the evaluation matching degree of the proposed model was 80.7%, which had significant advantages over HRNet, PoseResNet, and SPIN. In the complex “lateral raise” and “backward downward lift” actions, the evaluation matching degree of the proposed model was 78.6% and 83.7%, respectively, which was 3–5% higher than other models. Especially in the “lateral raise” action, the accuracy of the proposed model accurately matched the key points, overcoming the bias that traditional models were prone to in complex poses. These results indicate that the proposed model has a far superior ability in action feature extraction and matching compared to other models, which can better capture action details and improve the accuracy and stability of action evaluation. Finally, tests are conducted using indicators such as action improvement rate, action similarity, action quality, and evaluation time, as displayed in Table 4.

Table 4.

Action evaluation results of different models.

Athlete body type	Model	Movement improvement rate/%	Action similarity/%	Action quality	Evaluation time/s
Thin	PoseResNet	72.35	75.41	Medium	0.74
	SPIN	74.92	77.58	Medium	0.68
	HRNet	78.23	80.64	High	0.61
	Our model	86.79	88.45	Very high	0.49
Normal	PoseResNet	75.48	77.35	Medium	0.73
	SPIN	77.62	79.48	Medium	0.65
	HRNet	80.91	82.77	High	0.58
	Our model	89.16	90.23	Very high	0.45
Fat	PoseResNet	70.29	72.61	Low	0.78
	SPIN	73.84	75.93	Medium	0.69
	HRNet	76.53	78.81	High	0.63
	Our model	85.02	86.74	Very high	0.52

As can be seen from Table 4, the proposed model shows significant advantages in terms of action improvement rate, action similarity and action quality rating in the action assessment task for athletes of different body types, and the shortest assessment time. Specifically, in the group of lean athletes, our model achieved a movement improvement rate of 86.79% and a movement similarity of 88.45%, with a movement quality rating of Very High, and an evaluation time of only 0.49 seconds, which is superior to baseline methods such as PoseResNet, SPIN and HRNet. In the test with normal-sized athletes, our model also achieved 89.16% improvement rate and 90.23% similarity of movements, with the quality of movements maintained at Very High level, and the average evaluation time shortened to 0.45 seconds, which demonstrated high efficiency and accuracy. Even in the group of obese athletes, our model still maintains 85.02% of movement improvement rate and 86.74% of movement similarity, the movement quality level is much higher than that of other methods, and the evaluation time is 0.52 seconds, which is significantly better than that of other models. In summary, the proposed model shows higher action optimization effect and action recognition accuracy under different athletes’ body type conditions, and has good generalization ability and practical application potential.

Discussion

Aiming at the problems of computational complexity, inadequate dynamic modeling and inefficient action matching in aerobics action quality assessment, the study proposes an efficient assessment model that combines lightweight OpenPose, adaptive multi-scale ST-GCN and twin network. The experimental results show that the model achieves excellent performance on both Human3.6 M and CMU MoCap datasets, with a detection accuracy of 96.8%, an action evaluation matching of 89.7%, and an average evaluation time of only 0.71 seconds, which is significantly better than the mainstream methods such as PoseResNet, SPIN, and HRNet. The computational overhead is significantly reduced by the lightweight design of MircoNet, AMS-ST-GCN enhances the extraction of complex spatio-temporal features of actions, and the twin network structure improves the accuracy and consistency of action similarity assessment. Compared to the CNN-LSTM fusion model proposed by Liu Q, this study further strengthens the ability to recognize and model complex dynamic actions³; compared to the infrared imaging-based method of Lv D et al, the proposed method does not rely on specific lighting conditions and is more adaptable⁴; compared to the sensor-based scheme of Cao T, this study achieves wider applicability without the need for external hardware with real-time feedback capability.⁵ In addition, compared with the traditional tracking model proposed by Zuo N et al. and Zhang Y’s NMS strategy approach, this study is more advantageous in terms of detail capture and temporal coherence.^6,7

The model has a broad application potential in the fields of sports training, health monitoring, and dance recognition. The evaluation results can provide technical feedback for competitive aerobics and gymnastics programs, as well as automated movement specification monitoring in home fitness and rehabilitation training, supporting public health management and intelligent physical education. The combination of multimodal data, 3D visual inputs and incremental learning mechanisms is expected to further improve the dynamic adaptability and cross-individual generalization ability of the model. Nonetheless, the research still has shortcomings. The current training samples are mostly derived from standardized datasets, which have not been extensively validated in complex environments with multiple background disturbances. Meanwhile, the model still suffers from performance fluctuations when facing pose occlusion or non-standardized movements. The robustness and generalization ability of the model in real applications can be improved by introducing more real scene samples and fusion of multi-source information in the future.

Conclusion

Aiming at the problem of insufficient accuracy and real-time performance in aerobics movement quality assessment, the study proposes an assessment method based on lightweight OpenPose and ST-GCN model. Experiments show that the model Loss is reduced to 0.5 when the number of input feature map channels is 128 and the action adaptive weight is 0.8. By improving the OpenPose, ST-GCN and SN modules, the detection accuracy reaches 96.8%, and each module effectively improves the model performance. Under different lighting environments, the proposed model achieves the highest Precision 91.82%, 93.14%, and 90.85%; Recall 89.64%, 91.76%, and 88.29%; and F1 values 90.72%, 92.44%, and 89.56%, and the average detection time is shortened to less than 0.61 seconds, which is overall better than the PoseResNet, SPIN, and HRNet. In the movement assessment of athletes of different body types, the model achieved the highest movement improvement rates of 86.79%, 89.16%, and 85.02%, and the highest movement similarity of 88.45%, 90.23%, and 86.74%, and the movement quality was rated as Very High, and the assessment time was controlled at 0.49–0.52 seconds, showing excellent adaptability and robustness. Nevertheless, the study still has limitations. The current training data is mainly based on standard action samples, which lacks extensive validation in diverse and complex environments, which may affect the generalization ability of the model; there is still room for improvement in feature extraction and recognition stability in extreme occlusion or background interference scenarios; in addition, the model relies on a single skeleton input and lacks multimodal information fusion. Future research can consider: (1) optimizing the AMS-ST-GCN structure and introducing dynamic adjacency and attention mechanisms to improve the ability of complex action modeling; (2) fusing multi-source data such as images and IMUs to enhance the comprehensiveness and robustness of the evaluation; (3) applying on-line incremental learning to improve the adaptability to individual differences; and (4) carrying out large-scale extensive testing in real environments to further validate and expand the value of the model application. Model application value.

Footnotes

ORCID iD

Hua Pan

Funding

The author received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statements

The data will be made available on the request.*

References

Yang

Wang

, et al. ST-GCN human action recognition based on new partition strategy. Comput Integr Manuf Syst 2023; 29(12): 4040–4059. DOI: 10.13196/j.cims.2021.0453.

Sun

Gao

, et al. Evaluating the application of using biological pulse sensor in aerobics. Wirel Netw 2023; 29(8): 3567–3582. DOI: 10.1007/s11276-023-03428-z.

Liu

. Aerobics posture recognition based on neural network and sensors. Neural Comput Appl 2022; 34(5): 3337–3348. DOI: 10.1007/s00521-020-05632-w.

Yan

. Retracted article: application of visual tracking based on infrared target imaging in aerobics action evaluation system. Opt Quant Electron 2024; 56(3): 329–331. DOI: 10.1007/s11082-023-06016-9.

Cao

. Analysis of aerobic training posture using machine vision for body area networks. Wirel Netw 2023; 29(4): 1611–1620. DOI: 10.1007/s11276-022-03123-5.

Zuo

Liu

. Research on tracking and decomposing method of aerobics movement based on machine learning. Int J Inf Commun Technol 2023; 22(4): 362–376. DOI: 10.1504/IJICT.

Zhang

. A novel non-maximum suppression strategy via frame bounding box smooth for video aerobics target location. Sens Imaging 2022; 23(1): 29–32. DOI: 10.1007/s11220-022-00399-x.

Tian

Xia

. Feature extraction of gymnastics images based on multi-scale feature fusion algorithm. Scalable Comput Pract Exp 2024; 25(5): 3394–3407. DOI: 10.12694/scpe.v25i5.3190.

Hou

Xue

. Sports training injury risk assessment combined with dynamic analysis algorithm. Mol Cell BioMech 2024; 21(3): 484–493. DOI: 10.62617/mcb484.

10.

Wei

Yin

. Optimization of human posture recognition based on multi-view skeleton data fusion. Inf Technol Control 2024; 53(2): 542–553. DOI: 10.5755/j01.itc.53.2.36044.

11.

Chen

Zhang

, et al. Spatiotemporal consistency learning from momentum cues for human motion prediction. IEEE Trans Circuits Syst Video Technol 2023; 33(9): 4577–4587. DOI: 10.1109/TCSVT.2023.3284013.

12.

Enkhbat

Shih

Cheewaprakobkit

. Human action recognition and note recognition: a deep learning approach using STA-GCN. Sensors 2024; 24(8): 2519–2524. DOI: 10.3390/s24082519.

13.

Janbi

Ghaseb

Almazroi

. ESTS‐GCN: an ensemble spatial–temporal skeleton‐based graph convolutional networks for violence detection. Int J Intell Syst 2024; 2024(1): 3337–3346. DOI: 10.1155/2024/2323337.

14.

Liyuan

. Retracted article: application of light imaging detection based on data information visualization in aerobics training management competition simulation system. Opt Quant Electron 2024; 56(2): 148–153. DOI: 10.1007/s11082-023-05771-z.

15.

Tsai

Huang

. Enhancing accuracy of human action recognition system using skeleton point correction method. Multimed Tools Appl 2022; 81(5): 7439–7459. DOI: 10.1007/s11042-022-12000-4.

16.

Kim

. Comparative analysis of OpenPose, PoseNet, and MoveNet models for pose estimation in mobile devices. Trait. Signal 2022; 39(1): 119–124. DOI: 10.18280/ts.390111.

17.

Van Hooren

Pecasse

Meijer

, et al. The accuracy of markerless motion capture combined with computer vision techniques for measuring running kinematics. Scand J Med Sci Sports 2023; 33(6): 966–978. DOI: 10.1111/sms.14319.

18.

Zhang

Huang

, et al. Partitioning leakage detection in water distribution systems: a specialized deep learning framework enhanced by spatial-temporal graph convolutional networks. ACS EST Water 2024; 4(8): 3453–3463. DOI: 10.1021/acsestwater.4c00285.

19.

Ergeneci

Carter

Kosmas

. sEMG onset detection via bidirectional recurrent neural networks with applications to sports science. IEEE Sensors J 2022; 22(19): 18751–18761. DOI: 10.1109/JSEN.2022.3198882.

20.

Chowdhury

Shivakumara

Raghavendra

, et al. An episodic learning network for text detection on human bodies in sports images. IEEE Trans Circuits Syst Video Technol 2021; 32(4): 2279–2289. DOI:10.1109/TCSVT.2021.3092713.

21.

D’Antonio

Taborri

Mileti

, et al. Validation of a 3D markerless system for gait analysis based on OpenPose and two RGB webcams. IEEE Sens J 2021; 21(15): 17064–17075. DOI: 10.1109/JSEN.2021.3081188.

22.

Bini

Serrancoli

Santiago

, et al. Validity of neural networks to determine body position on the bicycle. Res Q Exerc Sport 2023; 94(4): 905–912. DOI: 10.1080/02701367.2022.2070103.

23.

Byun

. Development of aerobic exercise equipment using universal design: treadmill and arm ergometer. Healthcare, MDPI 2022; 10(11): 2278–2279. DOI: 10.3390/healthcare10112278.

24.

. Research on intelligent recognition technology of gymnastics posture based on KNN fusion DTW algorithm based on sensor technology. Int J Wireless Mobile Comput 2023; 25(1): 58–67. DOI: 10.1504/IJWMC.2023.132423.

25.

Oyama

Murayama

Ohta

. Relationship between the recognition error of dynamic postural control ability and the extent of exercise in middle-aged older women. Sport Sci Health 2022; 18(4): 1293–1299. DOI: 10.1007/s11332-022-00900-9.

26.

Shiguang

. Retracted article: simulation of sports injury prevention and rehabilitation monitoring based on fiber optic sensors and machine learning algorithms. Opt Quant Electron 2024; 56(4): 616–621. DOI: 10.1007/s11082-024-06321-x.

27.

Alsawadi

Rio

. Skeleton split strategies for spatial temporal graph convolution networks. Comp Mater Cont 2021; 71(3): 4643–4658. DOI: 10.32604/cmc.2022.022783.

28.

Tang

Xiong

, et al. ROpenPose: a rapider OpenPose model for astronaut operation attitude detection. IEEE Trans Ind Electron 2021; 69(1): 1043–1052. DOI: 10.1109/TIE.2020.3048285.

29.

Mokayed

Quan

Alkhaled

, et al. Real-time human detection and counting system using deep learning computer vision techniques. Artif Intell Appl 2023; 1(4): 221–229. DOI: 10.47852/bonviewAIA2202391.

30.

Zhang

Zhou

Deng

. Cascaded hierarchical CNN for 2D hand pose estimation from a single color image. Multimed Tools Appl 2022; 81(18): 25745–25763. DOI: 10.1007/s11042-022-12780-9.

31.

Bhosle

Musande

. Evaluation of deep learning CNN model for recognition of devanagari digit. Artif Intell Appl 2023; 1(2): 114–118. DOI: 10.47852/bonviewAIA3202441.

32.

Lovanshi

Tiwari

. Human skeleton pose and spatio-temporal feature-based activity recognition using ST-GCN. Multimed Tools Appl 2024; 83(5): 12705–12730. DOI: 10.1007/s11042-023-16001-9.

33.

Wang

Deng

. TFC-GCN: lightweight temporal feature cross-extraction graph convolutional network for skeleton-based action recognition. Sensors 2023; 23(12): 5593–5596. DOI: 10.3390/s23125593.

34.

Zheng

Huang

Wang

, et al. Spatio-temporal fusion for human action recognition via joint trajectory graph. Proc AAAI Conf Artif Intell 2024; 38(7): 7579–7587. DOI: 10.1609/aaai.v38i7.28590.

35.

Xing

Zhu

, et al. An improved spatial temporal graph convolutional network for robust skeleton-based action recognition. Appl Intell 2023; 53(4): 4592–4608. DOI: 10.1007/s10489-022-03589-y.

36.

Zhuang

Qin

Ding

, et al. Temporal refinement graph convolutional network for skeleton-based action recognition. IEEE Trans Artif Intell 2023; 5(4): 1586–1598. DOI: 10.1109/TAI.2023.3329799.