With the development of the times, exploring and protecting folk dances has become a historical inevitability for their inheritance and development. In order to achieve efficient action recognition and tracking in intelligent dance teaching and improve the quality of folk dance teaching, a DeepLabv3+ segmentation model with deep convolutional networks and dilated convolutions is designed to accurately segment human movements from the background in dance videos. Then, it is improved by combining hybrid dilated convolution and non-local attention mechanism. The IoU value was 0.92 on the training set and 0.93 on the test set. The precision, recall, and F1-score reached 93.05%, 95.08%, and 94.80%, respectively. In summary, the intelligent dance teaching folk dance dynamic segmentation algorithm based on HN-DeepLabv3+ has high segmentation accuracy and robustness, providing strong technical support for intelligent dance teaching.