Retrieving similar partial trajectories in soccer games considering surrounding player movements

Advancements in soccer tracking technology have enabled the large-scale collection of detailed movement data of players and the ball, crucial for match analysis and strategy development. However, manual analysis of such a vast amount of data is impractical, necessitating the development of auto-mated systems to efficiently retrieve similar scenes from player movement databases. Traditional trajectory search methods using Siamese Neural Networks focus on the movements of all players simultaneously, enabling high-speed searches. However, they struggle with precision when focusing on specific players due to noise from the movements of others. This study proposes a more flexible trajectory search method that considers the surrounding context and enables searches involving any number of players. By using trajectory pairs as the basic unit and incorporating the movements of opposing players, the proposed method enhances search accuracy. Evaluations using J. League data demonstrated the effectiveness of the proposed method, particularly in scenarios without time differences.