A Study of Dance Movement Capture and Posture Recognition Method Based on Vision Sensors

Abstract

With the development of technology, posture recognition methods have been applied in more and more fields. However, there is relatively little research on posture recognition in dance. Therefore, this paper studied the capture and posture recognition of dance movements to understand the usability of the proposed method in dance posture recognition. Firstly, the Kinect V2 visual sensor was used to capture dance movements and obtain human skeletal joint data. Then, a three-dimensional convolutional neural network (3D CNN) model was designed by fusing joint coordinate features with joint velocity features as general features for recognizing different dance postures. Through experiments on NTU60 and self-built dance datasets, it was found that the 3D CNN performed best with a dropout rate of 0.4, a ReLU activation function, and fusion features. Compared to other posture recognition methods, the recognition rates of the 3D CNN on CS and CV in NTU60 were 88.8% and 95.3%, respectively, while the average recognition rate on the dance dataset reached 98.72%, which was higher than others. The experimental results demonstrate the effectiveness of our proposed method for dance posture recognition, providing a new approach for posture recognition research and making contributions to the inheritance of folk dances. Doi: 10.28991/HIJ-2023-04-02-03 Full Text: PDF