HWD-DFFormer: Wavelet-Enhanced Dynamic Filtering for Micro-Doppler Human Activity Recognition

Radar-based human activity recognition has attracted increasing attention for contactless health monitoring. However, existing methods rely on conventional downsampling techniques, which irreversibly discard high-frequency micro-Doppler features and limit fine-grained discrimination between similar activities. We propose HWD-DFFormer, a lightweight model that integrates Haar wavelet-based downsampling (HWD) and dynamic filtering to efficiently extract micro-Doppler features in the frequency domain. HWD replaces conventional convolutional downsampling to preserve high-frequency micro-Doppler features, while the dynamic filter performs adaptive frequency-domain weighting to enhance micro-Doppler feature extraction. Evaluated on the Kitchen-HAR dataset across seven activities, HWD-DFFormer achieves 99.97% accuracy with only 3.83M parameters and a 15.3MB model size, outperforming four state-of-the-art methods.