Deep Learning-Based Classification Of Harmandali Dance Figures

Abstract

This study aims to recognize figures from the Turkish folk dance Harmandali using deep learning-based methods. Although AI-based studies on our country's folk dances are quite rare, this study is one of the first examples integrating robotic systems. Accordingly, a special video dataset consisting of Harmandali figures was created; the images extracted from the videos were processed using the Google MediaPipe Pose library, and skeleton keypoints representing the dancer's joint positions were extracted from each frame. The resulting time series data were classified using 1D-CNN, LSTM, and GRU architectures, and their performances were compared. Experimental results show that GRU-based models achieve the highest success with 89.7% top-1, 97.58% top-3 accuracy rates and 0.8953 F1 score. This study demonstrates that deep learning approaches based on skeleton representations are effective for the automatic recognition of dance figures. © 2025 IEEE.