A Haar Wavelet Down-sampling Linear Deformable ConvFormer for Radar-based Human Indoor Activity Recognition

In the context of modern smart home and healthcare automation, accurately monitoring and identifying human activities indoors is crucial. In this paper, we developed the Haar wavelet down-sampling linear deformable ConvFormer (HWDLD-ConvFormer), a novel model specifically designed for human activity recognition (HAR). The model integrates Haar wavelet downsampling (HWD) with linear deformable convolution (LDConv) within a ConvFormer architecture, enabling efficient extraction and processing of complex radar data. Through extensive experiments, HWDLD-ConvFormer demonstrated an average accuracy improvement of 6.16% over the base ConvFormer model across various HAR events, achieving significant performance gains in Precision, Recall, and F1-Score. Moreover, when compared to other state-of-the-art algorithms, including VIT, VIT+CNN, and 2D-Transformer, HWDLD-ConvFormer outperformed all, showing an 8.3% improvement in accuracy and an 8.9% enhancement in recall over the best alternative.