目的：针对散打技术动作复杂，评分难度高的特点，提出基于人工智能的散打动作智能评分方 法，以提高比赛中动作识别与评分的准确性。方法：收集2015—2024年间发布于抖音、快手等网络平台 上的全国武术散打锦标赛、全国武术散打冠军赛和全运会武术散打比赛视频，构建并标注了散打动作数 据集。在此基础上，结合图卷积网络（Graph Convolutional Network， GCN）PoseSAGE模型，加入残差连 接，构建了改进模型PoseSAGERES，并开展了与 PoseGNN、PoseSAGE模型的对比实验。实验结果表明， PoseSAGERES模型在小规模数据集上实现了 73.76%的分类准确率，显著优于其他模型。一致性分析显 示，该方法与人工评判结果具有良好一致性，体现出在散打动作智能评分中的应用潜力。研究证实了基 于人工智能的散打智能评分方法的有效性，以及残差链接机制在提升复杂动作识别准确率方面的促进 作用，为散打动作的自动化分析与智能评分提供了创新性解决方案。未来的研究将着力于扩展数据集 规模，丰富动作类别，进一步优化模型性能与泛化能力。

This study addresses the challenges posed by the complexity of Sanda techniques and the difficulty of accurate scoring by proposing an AI-based intelligent scoring method for Sanda actions. The goal is to improve the accuracy of action recognition and scoring during competitions. We collected and annotated a comprehensive dataset of Sanda actions from videos of the National Wushu Sanda Championship， the National Wushu Sanda Tournament of Champions， and the Wushu Sanda events of the National Games between 2015 and 2024， published on popular platforms such as Douyin and Kuaishou. Based on this dataset， an improved model named PoseSAGERES was developed by integrating residual connections into the Graph Convolutional Network（GCN）-based PoseSAGE model. Comparative experiments were conducted with the PoseGNN and PoseSAGE models to evaluate performance. Experimental results demonstrated that the PoseSAGERES model achieved a classification accuracy of 73.76% on a small-scale dataset， significantly outperforming other models. Consistency analysis revealed strong agreement between this method and manual evaluations， highlighting its potential for application in intelligent scoring of Sanda actions. The study has demonstrated the effectiveness of AI-based intelligent scoring methods for Sanda， as well as the positive impact of residual connection mechanisms in improving the accuracy of complex action recognition. It provides an innovative solution for the automated analysis and intelligent evaluation of Sanda movements. Future research will focus on expanding the dataset scale， enriching the diversity of action categories， and further optimizing model performance and generalization capability.

TP391.41； G852.4； TP18

2024年度国家社会科学基金后期资助项目（24FTYB015）， 广西壮族自治区教育厅学位与研究生教育改革项目（JGY2023137）， 广西师范大学 2025年广西学位与研究生教育教改课题（XYJG2025 026）。