DKFG: A Tennis Activity Recognition System for Data Scarcity and Noise Interference Based on Dual-Loop Kalman Filter and GAN

The importance of human activity recognition (HAR) technology in sports training is becoming increasingly prominent. In the context of tennis, an accurate system for tennis activity recognition is essential for improving athletes’ training quality. However, the complexity of specific tennis movements often results in scarce, unevenly distributed samples, while the raw sensor data are further affected by noise and interference. When traditional models are directly applied to tennis activity recognition, their performance tends to be unreliable. To address these challenges, an innovative tennis activity recognition system, termed DKFG, is proposed, which integrates a dual-loop Kalman filtering (DKF) framework with a generative adversarial network (GAN), using a deep convolutional neural network (DCNN) as the classifier. First, the DKF technique preprocesses raw data from wearable devices by filtering and smoothing the data stream, detrending, and dynamically segmenting samples to enhance data quality. Subsequently, a GAN designed explicitly for tennis data augmentation is used to expand the dataset, thereby improving classifier performance. Finally, the system employs a DCNN to perform tennis activity recognition. Experimental results demonstrate that the proposed method can efficiently recognize tennis actions. Moreover, DKFG achieves strong recognition performance and generalization across datasets.