Deep learning driven multi-scale spatiotemporal fusion dance spectrum generation network: A method based on human pose fusion

With the integration of dance art and computer technology, automatic dance score generation has become a new research direction in computer vision and machine learning, but generating the corresponding Laban symbols by capturing the skeletal key points of dance movements is a challenging task. In this study, we propose an automatic dance score generation model that utilizes local spatio-temporal features to address the inefficiency and creativity limitations of traditional choreography methods. Specifically, we propose the Multiscale Spatio-Temporal Convolution (MSConv) module to capture local spatio-temporal features in human skeletal motion sequences. In addition, the Compressed Pyramid Attention (CPA) mechanism is used to achieve effective fusion of global and local features. This mechanism facilitates the interaction between global and local spatio-temporal information and automatically generates dance sequences by analyzing motion data from dance videos to extract key features. We validate the proposed method on Laban 16 and Laban 48 dance score datasets, and the generated Laban sequences preserve the original style of the dance sequences with a combined accuracy of 94.2% and 93.7%, respectively.