Deep Learning Sequence Network for Identifying and Analyzing Archery Shooting Patterns

This study developed a method to evaluate archery shooting patterns and provide feedback by using a deep learning sequence network on landmark joint coordinates tracked via pose estimation. The data used in this study were collected over three months, from February to May 2024, by selecting four national-level male archers in Korea and recording their training sessions. The dataset encompasses full-body joint point coordinates extracted from full HD resolution (1920×1080 pixels), 60 fps videos, where the athlete's entire body is fully captured within the frame. [Each dataset includes the full sequence from the athlete's ready posture through drawing, holding, aiming, and releasing, and was labeled by athlete. Figure 1 illustrates an example of the dataset extraction process used in this study; only shooting samples under 900 frames per shot (less than 15 seconds) were included. The original shooting dataset contained 6,231 samples; however, after filtering, 5,984 samples were used in the final dataset. The original data and data availability used in this study are provided at the URL. (https://github.com/analysispark/deep-archery-sequence-data.git). All participants were fully informed of the study's purpose and procedures prior to data collection, and each provided written consent to participate.

Figure 1.

Environment Experimental Setup and Recording Environment Used for Video Analysis

This study employs the real-time multi-person one-stage (RTMO) framework to analyze the shooting motions of archers. RTMO is a state-of-the-art framework for real-time multi-person pose estimation. It is based on the you only look once (YOLO) architecture and can estimate poses directly without the need for a separate person detector. It achieves an average precision (AP) 1.1% higher than existing one-stage pose estimators on the common objects in context (COCO) dataset. It reaches 74.8% AP on COCO va12017, demonstrating excellent real-time performance with a processing speed of 141 FPS on a single V100 GPU (Chen et al., 2019; Lu et al., 2024).

The joint points extracted from the RTMO framework follow the COCO dataset's keypoint definitions and provide 17 joint points. The list and body parts corresponding to these joint points are shown below. (Table 1, Figure 2)

Figure 2.

Example of Joint Extraction in Archery Shooting

Figure 3.

Shooting Process from Actual Video Clips Used

To analyze only actual shooting motions, this study trimmed the original footage to exclude unnecessary segments before and after the shooting motion, using only clips from the ready position to release. Joint points were extracted frame by frame in JSON format. Each JSON file for a shooting instance contains metadata, including athlete information, shooting date and time, and an array of joint point data, with the structure [n][17][3]. Here, n denotes the number of frames. Moreover, the 3D array of 17 joint points includes x-coordinate, y-coordinate, and confidence score.

To suit the sequence model, the keypoint coordinates from each frame were normalized using MinMaxScaler from the scikit-learn library. This normalization compensates for differences in camera angles across recording dates and variations in distance between the athlete and the camera during each shoot, and helps prevent performance degradation in machine learning models due to scale differences in the data (Deepa & Ramesh, 2022). Additionally, for data with low confidence scores, interpolation using coordinates from −5 and +5 frames was applied to compute average values as a preprocessing method. Furthermore, 11 keypoints (nose, left and right eyes, ears, hips, knees, ankles) were excluded to reduce training cost and enable real-time processing, as these points remain fixed during the shooting motion (Ji et al., 2024; Lee et al., 2024; Zhang & Lin, 2024).

To analyze shooting motions and estimate individual patterns of archers, this study designed deep learning sequence models. Specifically, unidirectional sequence (RNN, LSTM, GRU) and bidirectional models (Bi-LSTM, Bi-GRU) were designed and compared. At this stage, the model was designed to determine the distinctive shooting motion patterns for each archer, forming the basis for subsequent analysis of motion consistency and variability. All models included a Global Average Pooling 1D layer to reduce the time dimension and extract average features from the entire sequence. Hyperparameter tuning of the neural networks was performed using the Optuna library (Akiba et al., 2019) to identify optimal parameters; Table 2 presents the specific model structures.

Model training used RMSE as the loss function and Adam as the optimization algorithm. The learning rate was set to 0.001, and the loss function used was categorical crossentropy. Considering the video features at 60 fps, the batch size was set to 64. Dropout layers were applied, randomly deactivating 30% of neurons to prevent overfitting. Additionally, early stopping was implemented to terminate training early if the validation loss did not improve for 10 consecutive epochs.

Deep learning sequence models were used to learn the shooting motion patterns of archers. The training results for each model are as follows.

The RNN model showed a training accuracy of 89.7%. The GRU algorithm achieved a training accuracy of 98.5%, while the bidirectional GRU (Bi-GRU) achieved 98.7%, showing the best performance. The LSTM model recorded the lowest performance at 55.9%, while the bidirectional LSTM (Bi-LSTM) achieved 96.6%. (Figure 4)

Figure 4.

Sequence Model Training Accuracy

Table 3 presents performance evaluation metrics for each model. Bi-GRU had the highest scores in precision, recall, and F1-score. Regarding LSTM, precision was higher than accuracy, but F1-score was lower. In contrast to the single LSTM, the Bi-LSTM model demonstrated excellent performance in the evaluation metric (accuracy).

The Bi-GRU model, which showed the best performance metrics, achieved a training accuracy of 99.9% and a training loss of 0.0013 over 64 epochs. The accuracy on the test dataset, which was not used during training, was 98.7%, with a loss of 0.0635.

This section visually analyzes the predictions made by the pose estimation-based deep learning model to compare the shooting motion patterns of athletes. The results of visualizing athletes' shooting patterns are as follows.

Figure 5 visualizes joint point movements of each athlete across all shooting data, starting from the top-left. For example, the upper graph for athlete A shows changes in x-coordinates of joints during the shooting motion, while the lower graph shows changes in y-coordinates. The x-axis in each athlete's graph represents frames (time), and the y-axis shows normalized coordinate values. A narrower variance range indicates that the athlete performs consistent motions, and the figure clearly shows distinct differences in shooting patterns among the four athletes.

Figure 5.

Visualization of Athletes' Shooting Patterns.

Note: Normalized x- and y-coordinate trajectories for four athletes across complete shooting sequences. Narrow variance indicates higher motion consistency.

Machine learning enables automated, large-scale, and quantitative analysis of motion consistency and variability, which would be challenging to achieve solely through manual pose estimation methods. Recent advancements in AI technologies have brought revolutionary changes to sports science, enabling quantitative analysis of motion and the development of personalized training programs. In sports like archery, which demand high precision, deep learning-based sequence models have shown excellent capability in capturing kinematic chain reactions. Archery involves shooting arrows at a target from a fixed distance, with scores assigned based on arrow accuracy. The shooting motion in archery can be divided into six stages: (1) grip, (2) draw, (3) hold, (4) aim, (5) release, and (6) full draw (World Archery, 2020). Each stage is closely linked to the previous one, and the quality of the preceding motion affects the next, making it essential to maintain proper form at each step (Vendrame et al., 2024).

Owing to these characteristics, archery shooting motions can be considered sequential data over time, indicating the methodological validity of using sequence models. Sequence models can effectively model temporal dependencies and quantitatively analyze how earlier motions affect later stages. Particularly, RNN-based models like GRU perform well in capturing long-term dependencies, making them suitable for analyzing continuous and complex motion patterns such as archery (Gavilanes, 2023; Purnama et al., 2024; Sunal et al., 2021).

Experimental results showed that the Bi-GRU model achieved the highest performance with 98.7% accuracy. Figure 6 shows the confusion matrix for classifying athletes' shooting patterns, confirming the model's robustness for real-world application. In contrast, the low performance of LSTM (55.9%) is presumed to stem from overfitting or insufficient hyperparameter tuning, while the high accuracy of Bi-LSTM (96.6%) suggests that the bidirectional structure partially overcame these limitations (Yang et al., 2020).

Figure 6.

Confusion Matrix of the Bi-GRU Model.

Note: X-axis = predicted classes (A D), Y-axis = actual classes. Overall accuracy reached 98%.

The proposed system enables coaches to provide more concrete and practical feedback based on motion analysis results, significantly contributing to athletes' technical improvement. By visualizing the variance in each athlete's shooting trajectory, the system quantifies subtle patterns that are not visually detectable, identifying points of motion inconsistency. This allows coaches to guide athletes toward consistent motion, which is critical in archery. Figure 5 visualizes joint point movements of each athlete, clearly showing distinct differences in shooting patterns. Figure 7 further illustrates three shooting instances of Athlete A, enabling precise feedback for consistent motion. When visualized on actual shooting videos, the error rates of joint points can be presented as shown in Figure 8, allowing athletes to analyze their motion more effectively while watching their practice footage.

Figure 7.

Joint Variation Graphs for Three Shooting Instances of Athlete A.

Note: Left panel shows x-coordinate trajectories; right panel shows y-coordinate trajectories for key joints across three shooting instances.

Figure 8.

Field Application Example of Feedback for Consistent Shooting Motion

Note: Error rates for each joint are overlaid on actual shooting video frames. Joints exceeding a standard error threshold of 0.4 are highlighted for corrective training.

Through such applications, the deep learning sequence network-based archery shooting motion analysis system developed in this study is expected to significantly contribute to technical improvement and performance enhancement among archers through consistent motion training.