Lightweight Gesture Recognition Model Based on CWT and Enhanced CBAM

Abstract

As an interaction method, gesture is widely used in various fields because of its simplicity and intuition. At present, most radar-based gesture recognition methods use short-time Fourier transform (STFT) to process radar echo information, but STFT cannot improve time resolution and frequency resolution simultaneously. To fully utilize effective information, the continuous wavelet transform (CWT) is used to process the radar echo signals. In view of the complexity of gesture recognition networks, a novel network incorporating CWT and an enhanced convolutional block attention module (CBAM) mechanism is proposed. First, features are pre-extracted using a feature extraction network. Then, the CBAM module is improved and integrated. Finally, the classification result is formed. To verify the model’s effectiveness, experiments collected data for nine distinct gestures. The results demonstrate a recognition accuracy of 96.3% via participant-stratified cross validation. Moreover, the model parameters are optimized, facilitating relatively simple implementation. It also exhibits strong performance on unknown datasets, proving its excellent generalization capability.