College Basketball Teaching based on Action Recognition using Graph Convolutional Network and Spatiotemporal with Long Short-Term Memory

In recent years, basketball action recognition has developed player training and coaching strategies. Traditional basketball teaching methods are struggled to robustly capture joint moments in complex backgrounds, ineffective modeling of long-term dependencies. To overcome this issue, a Graph Convolutional Network and Spatiotemporal with Long Short-Term Memory (GCN-ST-LSTM) approach is proposed. Initially, input data is collected from the NTURGB+D120 dataset which contains skeleton-based motion data is processed 2D-to-3D conversion using VideoPoSe3D to improve movement representation. After processing, GCN-ST is utilized to extract spatial and temporal features from player movements to capture joint dependencies and motion sequences. These extracted features are then transferred into an LSTM network, which processes long-term motion dependencies to classify basketball actions accurately. Lastly, the classified actions are analyzed for skill evaluation and real-time feedback. The proposed GCN-ST-LSTM approach achieved better accuracy (98%), precision (97%), recall (96%), and F1-score (97%) compared to existing skeleton-based action recognition models, making it an effective tool for Artificial Intelligence (AI)-driven basketball coaching and training.