Dance movement recognition method based on convolutional neural network
Zhiqun Lin · 2023
In order to improve the accuracy of dance movement recognition, we propose a dynamic adaptive video segmentation strategy. In this method, motion information statistics are carried out under the optical flow field with median compensation extracted, and video of different frame lengths are obtained by setting the threshold of frame segmentation, which is used as the input of convolutional neural network. The 3DSTPP structure is introduced into the transmission mapping of the network, and the characteristics of the traditional 3D convolutional neural network are extended by using the idea of pyramid pool. Feature description operators of the same dimension are extracted from video input sequences, and feature dimension reduction is carried out by long-term time pooling method, which is solved by two-level optimization function. The test results show that the generalization performance of the improved network is improved, and the classification accuracy of the two test data sets can reach 92.25% and 66.48%, respectively.