Under the background of intelligent development in sports field, it is of great significance to accurately capture complex sports action patterns to improve the training quality and the level of event analysis. Aiming at the limitation of traditional methods in action feature extraction, this paper explores the construction of sports action pattern capture algorithm model by using GCN (Graph Convolutional Network). By transforming human skeleton into graph structure, combining spatio-temporal convolution algorithm and graph attention mechanism, a lightweight network structure is designed to realize efficient extraction of spatio-temporal characteristics of action. In the experimental stage, the self-built data set is combined with NTU RGB+D 120 public data to construct a data set of 4500 samples, and the model is evaluated by 50% cross-validation. The results show that the loss value of this model converges to about 0.32 in training. During the test, the average recognition time is 42.3 ms, the parameter quantity is 18.6 M, the memory occupation is 215 MB, and the accuracy of multi-action continuous recognition reaches 91.7%, which is significantly improved compared with the contrast models such as 3D-CNN and ST-GCN. The research shows that the model based on GCN can effectively break through the bottleneck of traditional methods, and perform well in the efficiency, accuracy and lightweight of action pattern capture, which provides a better solution for sports action analysis.