GS-Transformer: A Lightweight Transformer Model for Real-Time HAR on Edge Devices

Human Activity Recognition (HAR) plays a crucial role in healthcare, sports monitoring, and human-machine interaction, where real-time analysis is essential for delivering personalized services. However, deploying HAR models on edge devices presents significant challenges due to limited computational resources and memory. Traditional deep learning models are inefficient in such environments. To address this, we introduce the Grouped Sparse Transformer (GST), a lightweight transformer optimized for real-time HAR on edge devices. By integrating Sparse Multi-Head Attention with Grouped Attention, GST reduces computational overhead and inference time while handling complex time-series data. Experiments using millimeter Wave (mmWave) and Inertial Measurement Unit (IMU) sensor data show that GST achieves state-of-the-art accuracy with a parameter size under 1MB. Compared to the baseline transformer’s 0.133 seconds inference time, GST achieves 0.057seconds, reducing inference time by over 57.1%, proving its suitability for real-time deployment in resource-constrained environments.