The real-time posture estimation system, based on a deep neural network, has been researched extensively. For example, the system used convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to determine a person’s posture from a single RGB camera. The system exhibited a high accuracy in estimating an individual human’s three-dimensional stance and utilizing a publicly accessible dataset [8,9,16–19]. Li et al. [20] proposed a posture estimation system using a similar methodology based on an RGB-D camera. The convolutional neural network estimated the human joint positions using a depth map and the pose. The system could calculate the pose accurately and in real-time on a publicly available dataset. A novel posture estimation system was created, employing convolutional neural networks and recurrent neural network techniques. The technology was designed to precisely forecast the human body’s position in real time with a sliding window mechanism. The suggested method successfully executed real-time posture estimation and achieved remarkable accuracy on a publicly available dataset [8,16,20–22]. The study uses media pipes and convolutional neural networks to measure the human body’s position in three dimensions. This multi-step method accurately assesses human posture in publicly available datasets [16,20,22–24]. GCN can be used to effectively model human joints and improve performance in pose estimation and action recognition tasks. Using GCN with spatiotemporal information enhances accuracy in complex scenes [25]. Action recognition improvement is also achieved by combining behavioral dependencies and contextual cues. This method allows the model to distinguish between similar poses [26].

Researchers have analyzed human actions using a spatiotemporal feature analysis method based on deep learning. In action recognition, temporal sequence recordings are essential. Shahroudy et al. [27] established the NTU RGB+D dataset, which has now become one of the largest and most diversified datasets for 3D action recognition. People extensively use it to evaluate the effectiveness of spatiotemporal models. Liu et al. [28] used global context-aware attention LSTM networks on 3D frame data, which helped the model focus on important joints and time segments. Graph Convolutional Networks (GCNs) have demonstrated efficacy as a powerful instrument for spatiotemporal modeling. Spatial-Temporal Graph Convolutional Networks (ST-GCN) serve as a technique for skeleton-based action recognition. The spatial interaction between joints and the temporal dynamics along the sequence are wrapped in this network [29]. The adaptive two-stream graph convolutional network (2s-AGCN) was proposed by Shi et al. [30] to support this idea of two-stream adaptive graph convolutional networks (2s-AGCN). These networks enhance accuracy by adaptively learning the graph topology from the supplied input. Moreover, Temporal Convolutional Networks (TCNs) exhibit considerable potential in this domain. Long Short-Term Memory (LSTM) models aren’t as good at sequence modeling as Temporal Convolutional Networks (TCN) models because they can’t capture long-range relationships as well as TCN models [31].

Besides deep learning methods, various methodologies have contributed to spatiotemporal analysis. Action posture is encoded using a bag of 3D points from depth data, and action graph dynamics are modeled for each action. Koniusz et al. [32] portrayed that space-time action sequences can efficiently be captured in a tensor space. Such methods have already been used to recognize badminton actions in RGB videos [33,34]. Feature extraction methods that have been applied are bounding box and histogram of oriented gradient (HOG). However, these methods produce misclassification for some classes involving the same pose during stroke [33]. The sliding window and Haar-like methods have also been applied. However, the recognition rate for different badminton players is not ideal [34]. In addition, detection of shot badminton players converts video data into sequential frames, applying sliding windows for feature extraction to detect shot timing and position with an accuracy of 95.9%. Similarly, image recognition of badminton swing motion based on a single inertial sensor transforms motion data into sequences, using sliding and action windows for accurate segmentation. This enables the Deep Residual LSTM model to recognize six swing types with over 90% accuracy, automating the extraction and recognition process for performance analysis in badminton [35,36]. Atanassov’s intuitive 3D fuzzy Histon roughness index method has been discovered to detect moving objects with colored data. (IA-IA3DFHRI) [14]. The proposed method demonstrates resilience to dynamic backgrounds.

Ensemble learning has been applied to several other fields, as pointed out by prior investigations. For example, Karim et al. [37] used ensemble learning combining logistic regression, decision trees, and SVM in healthcare. Regarding handwritten recognition, Karray et al. [38] proved improved accuracy by ensemble deep learning, RF, and SVM models. The accuracy of HAR can be enhanced by integrating SVM and HMM [39]. Combining SVM, KNN, and LR with a soft voting classifier can improve accuracy to 92.78% [40]. Das et al. [41] combined ensemble models (DT, RF, SVM, and KNN) with soft voting techniques and weighted optimization using FOX optimization. Research in human action recognition has also integrated deep learning models. The suggested ensemble learning algorithm includes DNN and CNN stacked at the gated recurrent unit (GRU) to recognize human actions [42]. Kaur et al. [43] combined ResNet50 and a custom CNN in human action recognition.

Badminton action recognition methods are categorized into two groups: machine learning-based classification and deep learning-based classification. Ting et al. [44] applied SVM for badminton recognition, classifying badminton strokes into ten distinct classes. Later, researchers further proved SVM’s efficacy in badminton action recognition. Anik et al. [45] identified badminton stroke techniques. In this study, the accuracy of SVM reached 88%. Ghazali et al. [46] also confirmed that SVM provides superior classification results compared to decision trees, KNN, and SVM, with SVM achieving an accuracy of 83.4%. He concluded that future research should enhance recognition accuracy by expanding and refining feature extraction techniques. Another conventional machine learning method applied to BAR is AdaBoost. The recognition rates of AdaBoost are higher than those of HMM [34].

Deep learning has been applied in human action recognition. Wang et al. [47] applied the Adaptive Feature Extraction Block (AFEB) AlexNet. Rahmad et al. [4] have proven that GoogleNet has the highest classification performance compared to AlexNet, VGGNet-16, and VGGNet-19. GoogleNet provided the best accuracy for differentiating between smash and non-smash techniques [48,49]. Steels et al. [50] confirmed the accuracy of CNN in classifying nine activities using accelerometer data.

Analyzing badminton movements using skeleton key point data is a mostly unexplored area of research. Using the AlphaPose framework, Liang et al. [1] analyzed key point data from certain badminton films. Comparative investigation shows that the Long Short-Term Memory (LSTM) model surpasses the CNN model. Another model implemented is PDDRNet for 3D human pose estimation and XGBoost for classification [2]. Liu et al. [3] applied GCNN with skeletal data for badminton action recognition.

This research contributes to the spatiotemporal analysis of skeleton data and weighted ensemble learning for badminton action recognition. The use of video data processed into 3D skeleton coordinates. The data skeleton was chosen because, according to Sun et al. [51], the data skeleton has the advantage of providing 3D pose information of an object. It’s simple yet informative and not sensitive to viewpoints, background, or light intensity changes. On the other hand, if the data is processed in RGB format, it will be susceptible to changes in viewpoint, light intensity, and changes in the background. The spatiotemporal feature captures the changes in the position of objects or body parts in space and time (temporal). This feature allows the model to understand the movement sequence, which is critical for recognizing actions that involve dynamic changes. In previous research, DTW was applied in human action recognition as a classifier by seeking the similarity between two sequences. In this paper, fast DTW is applied as a method for temporal feature extraction. One of the advantages of using fast DTW as a feature extraction method is its ability to reduce the number of features to a smaller amount. For example, the temporal feature in the right hip distance with dimensions 15 × 33, when processed with FDTW, will become a feature with dimensions 1 × 33.

Another contribution is the weighted ensemble of SVM, LR, RF, and AdaBoost. Machine learning classics were more efficient when compared with deep learning models. With appropriate feature extraction, machine learning models can achieve high accuracy. Ensemble learning methods that combine several classifier models have been proven to enhance accuracy.

Data was collected using a 15MP DSLR camera with the installation as presented in Fig. 2. The video data resolution is 30 fps. An indoor badminton court served as the location for data collection. We used the existing lighting in the field without any additions. The video records a single badminton stroke technique. The execution of the badminton stroke begins from the athlete’s ready position at the center of the court and returns to the center after performing the badminton stroke. The subjects included badminton athletes registered with PBSI, aged 15 to 22 years. Badminton techniques are presented in Fig. 3. The dataset consists of 1333 videos for training and 433 videos for validation.

The data preprocessing stage segmented the T-frame video data into l-frame sequences. We perform the frame segmentation length of l = 15 by taking frames from 1 to T, using the frame interval as presented in Eq. (1). To extract the 3D coordinates for 33 pose landmarks, we apply media pipe to each frame. Fig. 4 presents the 33 pose landmarks. Fig. 5 presents the 15-frame pose landmarks.

(1)interval=round(Tl)

We applied SVM, LR, RF, and AdaBoost in this research. Deep learning models (CNN and LSTM) are applied as a comparison in the machine learning model analysis. Each model was trained and evaluated individually. Five-fold cross-validation was used in validation.

The ensemble learning models in this study are arranged based on several variations of machine learning methods, as shown in Table 1, such as Ensemble-1 (E1) is configured from SVM, LR, RF, and AdaBoost, and configuration individual models for Ensemble-2 (E2), Ensemble-3 (E3), and Ensemble-4 (E4). The weighted soft voting classifier method (E1, E2, E3, and E4) can contribute to the final decision. The voting process will choose the more excellent score for more robust models.

We performed optimization by systematically testing several hyperparameter configurations for each model to determine the most effective combination that produces the maximum degree of performance. This method uses five-fold cross-validation to analyze the hyperparameters thoroughly. The training model that delivers the best performance metrics, such as accuracy, precision, recall, and F1 score, is chosen as the determinant of the best hyperparameters. The best hyperparameter model is shown in Table 2.

We conduct measurements of complexity, training time, and memory usage during model training. Table 3 presents the results. Eq. (5) to Eq. (10) allow for the calculation of the training time complexity (Big O). Compared to CNN and LSTM, classical machine learning (SVM, LR, RF, and AdaBoost) has lower complexity, training time, and memory usage during training.

(5)BigOSVM=O(n2.d)untilO(n3.d) (6)BigOLR=O(n.d) (7)BigORF=O(m.n.dlog⁡n) (8)BigOAdaBoost=O(m.n.d) (9)BigOCNN=O(n.d2.k2.f) (10)BigOLSTM=(n.t.h.d)

With n: number of samples, d: number of features, k: number of kernels, m: number of trees, t: sequence length.

We performed the validation test using 433 distinct data points that differed from the training data. Table 4 shows that the performance of each deep learning method is greater than that of machine learning methods. For example, the accuracy of deep learning (CNN, 93.06%; LSTM, 94.06%). The precision of deep learning (CNN, 94.66%; LSTM, 96.31%), The recall of deep learning (CNN, 94.61; LSTM, 95.26), et cetera., the accuracy of deep learning (CNN, 93.06%; LSTM, 94.06%). The precision of deep learning (CNN, 94.66%; LSTM, 96.31%). The recall of deep learning (CNN, 94.61; LSTM, 95.26), etc.

A soft voting classifier method in an ensemble model enhances accuracy and generalization in badminton action recognition. The weight value of each model SVM, LR, RF, and AdaBoost is assigned based on accuracy. We determine the weights by normalizing each model’s accuracy values. Table 5 shows the comparison evaluation matrix of ensemble models. E2 achieved the highest performance with an accuracy of 95.38%, precision of 96.01%, recall of 94.89%, and F1 score of 95, 25%.

According to the validation test results, ensemble model E2, which combines SVM, LR, and AB, outperforms ensemble model E1, which combines SVM, LR, RF, and AB. The superior performance of E2 indicates that adding RF to E1 does not contribute to performance improvement and may even introduce noise that leads to a decrease in model efficiency. It also emphasizes that the selection of the right individual models is necessary to obtain an optimal ensemble model. E2 successfully leverages the strengths of SVM, LR, and AB more synergistically without the additional complexity that could weaken the overall generalization power of the model, as seen in E1. This finding shows that the number of models in an ensemble does not always correlate with the final performance; instead, the quality and suitability of the combined models play a crucial role in determining the outcome.

Refer to Fig. 7 for the evaluation results in the confusion matrix. An analysis of the confusion matrix from the top two ensemble models, E1 and E2, reveals a notable disparity in performance, especially in the categories of drive backhand and overhead backhand. Group E2 had the highest accuracy. Model E2 has superior sensitivity in detecting drive backhand patterns. Moreover, in the overhead backhand class, model E2 again demonstrates its superiority by achieving greater accuracy in differentiating this movement from other classes. That indicates a superior capability to identify patterns of overhead backhand action.

The classification levels for each class are presented in Fig. 8. Based on that figure, model E2 has an accuracy level of 100% for the drive backhand class, and four classes have an accuracy above 95%. In comparison, there are two classes with an accuracy below 95%, namely the overhead backhand class at 90.2% and the underhand forehand class at 87.2%. The accuracy value for the underhand backhand suggests that the model struggles to differentiate this movement from other movements with similar characteristics in this classification study.

To evaluate the results of the proposed ensemble learning model, we created a comparison table of the results of recognizing badminton stroke techniques using skeleton data as presented in Table 6. Each model in this table uses different data for its recognition process. Ensemble learning with 3D spatiotemporal skeleton features has higher accuracy compared to other models from the state of the art.