2.1.3. Object Detection Model and Generation of Datasets for Model Training

An object detection model is required to locate the position of the players and shuttlecock in real-time. Therefore, it is important that the object detection model used in this work is fast enough for the system to work with high temporal resolution. As such, the YOLOv5 object detection model is used due to the high inference speed and high accuracy [31,32]. Using this object detection model, we can obtain information about the position and displacement of the shuttlecock, which can then be used to infer the hitting instant of a badminton service.

To facilitate good detection for the object detection model, a high-quality image dataset is required. The camera was placed at the side of the court, where the service judge is seated at a height of around 1.150 m. A total of six venues were used for the data acquisition to ensure a good diversity of hall environments for a better generalization of data. Collectively, 19 badminton players, including 14 university team players, were involved in the data collection for object detection model training. Among them, 16 were male players. A summary of the dataset was shown in Table 1.

Information about dataset for object detection model training.

To ensure a good mix of different service styles, the service conditions of the men’s singles, men’s doubles, and women’s doubles were recorded. This ensured that the common service conditions (Figure 4), which comprised backhand low serves, backhand flick serves, forehand low serves, and forehand high serves, were included in the dataset.

Various types of badminton service captured at different venues for the datasets.

Capturing footage of player serving has a downside, which is that the label counts for the shuttlecock are much lower compared to the other labels. The imbalance dataset is not ideal as it would cause the object detection model to have less of a chance to learn to detect the shuttlecocks, causing poor accuracy at recognizing the shuttlecock. Since the key feature of the system is to detect whether the shuttlecock is struck above the 1.150 m height limit, the ability of the objection detection model to recognize the shuttlecock is of utmost importance. To balance the dataset, 1900 images that contain only shuttlecocks were taken to increase the number of instances of the shuttlecocks (Figure 5). Overall, the training and validation dataset has a class distribution, as shown in Figure 6.

Shuttlecock-only images were added to increase the label counts of shuttlecocks to balance the dataset.

Distribution of the classes in the dataset.