Few-Shot Human Activity Recognition Using Millimeter-Wave Radar with WSCNN

Human activity recognition (HAR) using radar is valued for its robustness in all-weather conditions and its adherence to privacy. However, the limited diversity in radar datasets often results in models that are overly fitted to specific environments, impairing their ability to generalize. Addressing this challenge, we introduce a Wavelet Scattering Convolutional Neural Network (WSCNN). This model leverages a small, targeted dataset to generate time-frequency spectrograms. Experimental findings demonstrate that WSCNN not only sustains superior performance with few-shot training but also dramatically reduces the number of parameters required for training.