| Plum Blossom Boxing(PBB)is a typical case to represent traditional Chinese sports culture and is selected as the first batch of national intangible cultural heritage.Artificial intelligence(AI)applications driven by data and algorithms bright PBB a new way to promote further development.This paper is dedicated to broadening the spreading path of PBB,bridging the data shortcomings of the AI applied to traditional martial arts,and providing a basis for realizing an intelligent teaching system of PBB.Based on a survey of the current state of research on AI applied in martial arts,a few-shot action recognition method of the PBB based on graph convolutional neural networks is proposed.The proposed method extracts the human skeleton from PBB videos utilizing monocular 3D human pose estimation.Further,it classifies the PBB actions using metric learning to compare the similarity between supported actions and query actions.Inspired by the traditional martial arts concept of "coordinate hands and feet,coordinate elbows and knees,coordinate shoulders and hips",this study proposes an explainable module based on the attention mechanism.The spatial activation between the different joints is reflected by visualizing the heat map.During the model’s inference stage,inconsistencies in the reference frames of input skeleton sequences,arising from shooting angles and practice directions,present challenges in learning spatiotemporal features of actions.To enhance the model’s performance,this research further proposes a few-shot method for recognizing the PBB actions,which is based on reference frame consistency using graph convolutional relational networks.This method adaptively adjusts the reference frame of the skeleton sequences.The test results demonstrate that the proposed methods can effectively identifies the PBB actions.Moreover,the action recognition method based on reference frame consistency improves the model’s performance.Analyzing the attention heatmaps generated by the interpretable module reveals that the attention mechanism enables the model to focus on information between joints that match the action characteristics,regardless of whether there is an actual skeletal connection between the joints.The graph convolutional neural network-based PBB action recognition method can make PBB more digital and intelligent,alleviate the contradiction between deep learning and lack of data,promote the dissemination of martial arts culture,and has the potential value for applying to action evaluation,creating favorable conditions for realizing the intelligent teaching system of PBB. |
