Human Sitting Posture Recognition System Based on Pressure Sensor Layout Optimization and Edge Lightweight AI Inference

Abstract

Human sitting posture recognition is crucial for posture correction, rehabilitation, and health monitoring. However, existing pressure-sensor-based posture recognition systems often face challenges in balancing recognition accuracy and hardware cost. To address these challenges, a human sitting posture recognition system is proposed in this work based on sensor layout optimization and on-device lightweight AI inference. First, a two-stage K-means clustering framework (TKCF) is proposed to identify posture-sensitive regions and optimize the layout and number of pressure sensors. Second, this article designs a lightweight convolutional neural network (LW-CNN) integrated with the squeeze-and-excitation (SE) attention mechanism, which achieves a classification accuracy of 97.96% on the PC platform. After optimizing the layout of sensing points, we train the LW-CNN using the optimized 484-point sensor array and deploy the trained model onto the Microcontroller (MCU). The deployed model occupies just 77.97 kB of flash memory and 18.04 kB of static random access memory (SRAM). In contrast to the original sensor layout, the total quantity of sensing points decreases to 47.27% of the original number, SRAM memory overhead decreases to 47.40% of the baseline value, and the model inference latency decreases to 52.59% of the initial benchmark. These experimental results verify that the proposed method can well balance recognition accuracy, hardware resource cost, and user privacy protection, offering a feasible implementation scheme for lightweight sitting posture recognition and monitoring systems.