The Research On Temporal Action Localization In Figure Skating

Research background: The characteristics of high complexity,fine granularity and fast movement of figure skating make it difficult to watch or perform this sport in person.By studying the application of temporal action localization in the field of figure skating,this study provides rapid positioning service for the daily training or competition of figure skating,improves the intelligence level of figure skating training and forms a paradigm.Methods: In this paper,a weakly-supervised action localization method for few-shot in figure skating based on human posture temporal alignment was proposed.In this study,84 videos from world figure skating Championships,Winter Olympic Games and International Skating Grand Prix were collected.42 videos were used as support sets,including 298 movements.The query set consists of 42 videos and contains 302 actions.These videos were converted into 2D and 3D skeletal key-points using 2D HRNet and 3D VideoPose3D.After the reconstruction of skeletal key-points,according to the characteristics of figure skating and the data characteristics of 2D and 3D key bone points,Angle instead of coordinates is used for dynamic feature transformation,and two groups of feature matrices are found to represent the skeletal key-points features of figure skating.Finally,localization is carried out according to the proposed model.Results: In this study,a weakly-supervised action localization method for few-shot in figure skating based on human posture temporal alignment was proposed.The localizing results based on 3D skeletal key-points calculation were superior to those based on 2D skeletal key-points calculation.In the experiment based on 3D skeletal key-points,the method using the spatial consistency morphological features combined with the dynamic variation morphological features is superior to the method using only spatial consistency morphological features.Research conclusion:(1)This study uses unedited figure skating video data to analyze the movement of figure skating technique.By using dynamic feature transformation and using angle instead of bone key point coordinates,the influence of athletes’ posture and shooting distance on experimental results is eliminated.(2)In this study,we propose a weakly-supervised action localization method for few-shot in figure skating based on human posture temporal alignment.Through the study of a few samples,figure skating with high speed and complex movements in real competitive sports is localized.(3)With the help of the model,the experimental results show that the model has good performance,and has been applied in more figure skating videos with little artificial operation.