Title: A Sanda action recognition using CNN-LSTM network model

Authors: Jingying Ouyang; Jisheng Zhang; Yuxin Zhao; Shenghai Chen

Addresses: College of Physical Education, Hunan Normal University, Changsha 410012, China ' College of Physical Education, Hunan Normal University, Changsha 410012, China ' School of Business, Hunan University, Changsha 410082, China ' College of Computer Science and Electronic Engineering, Hunan University, Changsha 410012, China

Abstract: This study presents an action recognition algorithm based on convolutional long short-term memory (CNN-LSTM) to enhance the movement analysis precision. The model takes the joint position as the recognition node, and combines with the cylindrical coordinate system with random sliding window to effectively capture the angle and position information of the action frame. The algorithm groups joints, selects root nodes, and extracts local features through a multi-stream network, with classification completed in the pooling and fully connected layers. The proposed model achieves an average accuracy of 98.89%, with a recognition time of 0.61s and a minimal deviation of 0.035 in Sanda movements, demonstrating superior performance in action recognition.

Keywords: action recognition; convolution; long short-term memory network; LSTM; joint; Sanda.

DOI: 10.1504/IJWET.2025.149267

International Journal of Web Engineering and Technology, 2025 Vol.20 No.3, pp.297 - 314

Received: 25 Jan 2024
Accepted: 19 Sep 2024
Published online: 21 Oct 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article