A Real-time IMU Posture Prediction System Based on Lightweight Transformer Model

Human posture prediction is critical for balance control, rehabilitation, and wearable applications. However, traditional IMU-based methods face challenges in accuracy and real-time performance. This paper proposes a lightweight Transformer-based system incorporating a Single-Projection Gated Linear Attention (SP-GLA) mechanism to efficiently model temporal dependencies with linear computational complexity. Experiments showed that the proposed model achieved 95.71% accuracy with an inference latency under 20 milliseconds. The results demonstrate superior efficiency and accuracy compared to baseline LSTM and standard Transformer models, making it suitable for embedded deployment.