Football action recognition method based on skeleton graph multi feature and ensemble learning

Abstract In response to the shortcomings of existing football action recognition methods in terms of the utilization of skeleton graph features, capturing logical adjacency relationships, and algorithm robustness, this study innovatively proposes a football action recognition method based on skeleton graph multi-feature and ensemble learning using graph convolutional networks and ensemble learning methods. The proposed method achieved a 98.87% action recognition accuracy, with mean square error and mean absolute error of only 0.013 and 0.016, respectively. Meanwhile, in the presence of occlusion and noise, its root mean square error and average absolute percentage error were only 7.53% and 6.56%, respectively, demonstrating its excellent robustness. In addition, this method performed exceptionally well in terms of action recognition time. The average time consumption during the training stage was only 12.16 s, and the average recognition time for a single sample in the reasoning stage was 0.37 s, demonstrating highly efficient action recognition capabilities. In summary, the proposed method has achieved significant results in improving recognition accuracy, enhancing model robustness, and ensuring high efficiency. These advantages make this method have broad application prospects in football action recognition, and can provide some new guidance and suggestions for future intelligent sports analysis and athlete training evaluation.