With the growth of the aging population, the importance of real-time monitoring technologies for the safety of the elderly has become increasingly prominent. In particular, Human Activity Recognition (HAR) technology is being actively researched using...
With the growth of the aging population, the importance of real-time monitoring technologies for the safety of the elderly has become increasingly prominent. In particular, Human Activity Recognition (HAR) technology is being actively researched using various sensors to rapidly detect hazardous behaviors such as falls. Among these, radar-based HAR utilizing micro-Doppler maps is gaining attention due to its advantages in privacy preservation and robustness against environmental conditions. However, existing high-performance models face challenges in real-time processing within on-device environments.
To address these issues, this paper proposes MSPS-Mixer, a lightweight deep learning model that applies the MLP-Mixer architecture—known for its linear computational complexity and excellent performance—to micro-Doppler-based radar HAR for the first time. The proposed model employs a Multi-Scale Time Patch partitioning strategy that includes the entire frequency axis to preserve the temporal context of micro-Doppler signatures. By processing two patches with different temporal resolutions in parallel, it effectively learns multi-scale temporal information. Furthermore, a Shift operation is introduced to enhance temporal locality between adjacent time steps, thereby mitigating the lack of inductive bias in MLP-Mixers, while a hierarchical structure is formed using Downsampling to progressively reduce the number of tokens. Additionally, Knowledge Distillation and Quantization were applied for on-device deployment. Experimental results on the UoG dataset demonstrate that the proposed model achieves an accuracy of 94.91% (1.58M parameters), outperforming the existing SOTA model, LH-ViT, by 2.81%. The distilled lightweight model achieved an accuracy of 93.22% (358K parameters), proving the validity of the compression, and on-device implementation was successfully verified on two types of edge devices.