Sensor-Based Gymnastics Action Recognition Using Time-Series Images and a Lightweight Feature Fusion Network

With the development of micro-electromechanical systems (MEMSs) and artificial intelligence technology, the application of wearable devices in human motion capture and recognition has gradually become a research hotspot. However, existing action recognition methods based on wearable sensors still face issues such as limited feature extraction capability and insufficient information utilization, leaving significant room for improvement in recognition accuracy. To address these challenges, this article proposes a motion recognition method based on time-series images and a lightweight feature fusion network. First, two time-series-to-image conversion methods, raw sequence image (RI) and raw sequence change image (RCI), are proposed, which fully leverage the advantages of convolutional neural networks (CNNs) in image processing. Second, a dual-channel feature fusion network is designed, enhancing the ability to extract features of gymnastic movements through the selection of backbone networks and the design of feature fusion modules. Finally, the effectiveness of the proposed method in gymnastics action recognition is validated. The experimental results show that the proposed method achieves an accuracy of 99.35%, which is at least 4.77% higher than existing machine learning methods and at least 2.03% higher than advanced deep learning methods. This demonstrates a significant improvement in recognition accuracy, proving the effectiveness and superiority of the proposed method in human action recognition. This method is expected to be extended to more application scenarios and provide technical support for the development of wearable devices.