Semantic event analysis of sports match videos using domain knowledge and deep features

Semantic event analysis in sports videos faces challenges such as complex actions and high annotation costs. To address these issues, this paper proposes a novel framework that integrates domain knowledge with deep features. The approach first translates sports rules into computable spatio-temporal constraints, then designs a knowledge-injection network to guide deep models toward semantically critical regions. Finally, a knowledge-conditioned attention mechanism is introduced to fuse domain knowledge with visual features effectively. Experimental results on the SoccerNet dataset demonstrate that the proposed method achieves a mean average precision of 71.5%, outperforming strong baselines such as inflated 3D ConvNet and soccer background matting network by 13.3% and 3.6%, respectively. The framework shows significant improvements in detecting complex and sparse events, offering enhanced accuracy, robustness and generalisation capability with reduced reliance on large-scale annotated data.