Deep Learning for Scene Understanding in Sports: Enhancing Game Analysis through Cultural Heritage Insights

Abstract Scene understanding in sports analytics has witnessed significant advancements with the integration of deep learningtechniques. As sports continue to be a crucial element of cultural heritage, analyzing game dynamics through computationalmethods provides valuable insights into team strategies, player interactions, and tactical formations. Traditional sports analyticsapproaches heavily rely on handcrafted features and predefined heuristics, which often fail to generalize across diversegame scenarios and cultural variations in play styles. In this work, we propose a novel deep learning framework that fusesvisual scene understanding with cultural heritage metadata, offering a more contextualized interpretation of sports footage.Our approach leverages multimodal learning and self-supervised techniques to reduce dependency on extensive labeleddatasets, enabling scalability across modern and historical sports domains. Furthermore, we introduce a knowledge-guidedlearning mechanism that aligns deep learning models with domain-specific sports history, facilitating better recognition oftactical evolutions, historical play patterns, and culturally embedded strategies. Experimental evaluations on benchmarksports datasets demonstrate that our model outperforms existing state-of-the-art methods in both accuracy and contextualunderstanding. Beyond performance gains, our framework provides a holistic perspective that bridges modern analytics withthe preservation of sports heritage, enabling more comprehensive studies of the evolution of sports tactics, cultural influences,and strategic developments across eras. Ultimately, this work contributes to both advancing artificial intelligence applications insports and enriching historical sports research.