Research for Detecting Soccer Players using Panoramic Video

In recent years, quantitative tactical analysis based on player positional data has attracted attention in field sports such as soccer. IoT devices can acquire such data, but their use is limited by the difficulty of attaching them to opposing players. In camera-based approaches, a single camera covering the entire field makes players appear too small to detect accurately, whereas multiple cameras increase the burden of installation and operation. This research developed a workflow combining deep learning-based player detection from panoramic video generated by two fixed cameras with manual correction to detect all players on one team in every frame without missed or false detections. The annotation system design, training data size, data augmentation, training epochs, confidence threshold, and frame interpolation were systematically examined to identify the scenario with the shortest total required time. Across three matches, the required time was 28-41 hours per match, approximately 1/21-1/15 of full manual annotation. The positional RMSE was 0.504-0.986 m, indicating potential applicability to group tactical analysis.