Dance Video Analysis System Based on YOLO Pose and CNN+LSTM Deep Learning Models

The development and popularization of artificial intelligence and deep learning have driven the development of human pose recognition. Video analysis of human body posture has become a popular demand and important research direction in sports, fitness, dance, rehabilitation and other fields. However, related application research is still in its early stages of development. In response to this situation, this article applies artificial intelligence to the action classification and analysis of human posture videos. The content includes using the joint point sequence of the dance action dataset as data to train the CNN+LSTM model, achieving the goal of classifying actions. By using YOLO Pose to recognize human pose and joint points in test videos, and converting them into sequences to send to the CNN+LSTM model for testing, the goal of classifying human poses in videos outside the dataset can be achieved. Due to the fact that dance movements usually have musical rhythms, using music rhythms as keyframes for similarity calculation can not only achieve static image similarity comparison, but also observe the similarity change curve of continuous keyframes. The similarity algorithm adopts algorithms such as fastDTW algorithm, cosine similarity, Pearson correlation coefficient, etc., to compare video movements and use them as analysis and evaluation indicators for human posture in videos.