Video classification of Indonesian traditional dance using a hybrid CNN-LSTM model with pose estimation

The preservation and recognition of traditional Indonesian dances face challenges due to limited digital documentation and declining intergenerational transmission. Manual annotation of dance videos is time-consuming and prone to subjectivity, creating urgency for automated solutions. This study proposes a deep learning-based approach combining convolutional neural networks (CNN) for spatial feature extraction and long short-term memory (LSTM) for temporal modeling to recognize traditional dance movements from video sequences. The system leverages OpenPose for keypoint detection and gesture estimation, enabling frame-wise pose representation prior to classification. A hyperparameter tuning process was applied to optimize the CNN-LSTM architecture using 80% of the dataset for training and 20% for testing. Experimental results show the proposed model achieved a macro accuracy of 98.4%, with perfect precision, recall, and F1-score. This research contributes to cultural heritage digitization and intelligent video analysis by enabling accurate, real-time classification of traditional dances, providing a foundation for future systems in education, archiving, and motion-driven applications.