Football highlights detection based on twin comparison algorithm and multimodal feature fusion

To enhance the accuracy and robustness of football highlights detection (FHD), this study proposes an innovative method integrating Twin Comparison and multimodal features. It employs Twin Comparison for dual-threshold shot segmentation, classifying video scenes into four types: goals, corner kicks, penalty area attacks, and fouls. The approach uses a 3D convolutional neural network to extract spatiotemporal visual features and applies Mel-frequency cepstral coefficients for audio analysis. Multimodal information is integrated via post-fusion and classified using a Softmax classifier. Experiments demonstrated a shot segmentation precision of 92.4% and recall of 88.7%. The goal detection F1-score was 91.9%, while audio cheer detection reached 94.2%. The multimodal model achieved an average accuracy of 84.8%, an F1-score of 88.9%, and real-time processing at 33.5 FPS. The method significantly improves detection accuracy, effectively complements audiovisual features, and offers practical value for intelligent football video analysis, viewing experience, and content dissemination.