In this paper, based on computational behavioral science, on the basis of the original OpenPose model, the backbone feature extraction network VGG19 was changed to ResNet18 network, the prediction network in the OpenPose model was optimized by depth-separable convolution, and a linear module was added to the nonlinear network to obtain a classification network for recognizing the basic dance rhythmic behaviors of infants and children. Then the data were collected and processed using qualisys miqus m3 three-dimensional motion capture system and Visual3D biomechanical analysis software, and the statistical analysis was completed by SAS JMP14.2. The results showed that the accuracy of the improved OpenPose network model in recognizing dance rhythmic behaviors of infants and toddlers reached more than 99%. the performance characteristics of dance rhythmic behaviors of infants and toddlers aged 2~3 years old did not differ significantly in basic rhythmic movements (P>0.1), but there were significant differences in body exploration and social interaction movements (P<0.001). Especially in the social interaction category dimension, male toddlers’ completion of rotational imbalance and forward-leaning-backward movements was significantly higher than that of female toddlers. In conclusion, male and female infants and toddlers can synergistically control dance rhythmic behaviors through physiological structures, neuromotor control and behavioral patterns.
