Human Pose Recognition Method in Sports Scene Based on Improved DLSTM-GCN

Xiao Gan Chen
School of Physical Education, Wuxi Taihu University, Wuxi 214000, Jiangsu Province, China

Abstract:

At present, sports scene motion pose recognition has become an important direction for motion learning and motion action disassembly. It realizes the recognition and analysis of different movements through human pose recognition of sports scenes. However, traditional recognition methods in complex scenes are often not able to accurately recognize and analyze human pose. Therefore, to improve the accuracy of human pose recognition in sports scenes, a human pose recognition method based on improved deep long short-term memory graph convolutional networks for sports scenes is proposed. The new method introduces spatial-temporal graph convolutional networks and PoseCResNet-R framework on the traditional long short-term memory graph convolutional networks to improve the pose recognition accuracy. The results of the study revealed that the improved algorithm had the highest recognition accuracy in the action recognition tests in different motion scenes. Compared with the AlphaPose algorithm, its accuracy was improved by 24.2%, 24.6%, and 27.2%, respectively. Meanwhile, the improved algorithm had the lowest Fréchet inception distance (FID) value closer to the real image scene in different scene tests. In the tennis scene test, the mean average precision (mAP) value of the improved algorithm could reach a maximum of 94.80%. Moreover, the Giga floating-point operations per second (GFLOPs/V) metric of the improved algorithm was as low as 24. In the badminton scene test, the GFLOPs/V metric of the improved algorithm was 46 higher than the AlphaPose algorithm. In summary, the recognition accuracy of the improved algorithm is higher than other algorithmic models. This is a good guidance for improving the accuracy of human pose recognition in sports scenes.