Generating Deeply-Engineered Technical Features for Basketball Video Understanding

Abstract

Investigating video-guided basketball movement understanding is essential for enhancing sports coaching. Integrating basketball videos with human-computer interaction (HCI) algorithms significantly improves training efficiency. In this paper, we propose a novel method for basketball player motion recognition and prediction. We engineer the technical features of gameplay through video analysis and introduce a behavioral analysis method using a multi-layer learning architecture. Our main contributions include: 1) an LSTM-based deep learning architecture for player action recognition and prediction; 2) a clustering-based algorithm for basketball court and line detection; and 3) a keyframe selection technique for basketball videos based on spatial-temporal scoring. Experimental validation on a comprehensive basketball video dataset demonstrates the effectiveness of our method in accurately identifying player movements and analyzing behaviors.

Keywords

• Basketball
• motion tracking
• human-computer interaction
• video analysis
• CNNs
• RNNs