Deep Learning-Based Adaptive Sitting Posture Recognition System

Prolonged poor sitting posture has been proven to increase the risk of musculoskeletal disorders and chronic diseases significantly. This study proposes a portable and adaptive sitting posture recognition system. The system integrates infrared wide-angle cameras, a deep learning model, and edge computing technology to achieve real-time monitoring and posture feedback. It employs the YOLOv11 classification model along with a position calibration function, making it compatible with various office chair models and user body types. It accurately classifies 18 sitting posture combinations and effectively addresses recognition challenges caused by image occlusion. The system is implemented on the TB-RK3399ProX edge computing development board, utilizing the RK3399Pro chip's neural network processing unit (NPU) for efficient inference, achieving an inference latency of only 8.73 ms and an overall accuracy of 94.64%. Experimental results demonstrate outstanding stability and reliability across diverse office chair environments and under image occlusion challenges. Compared to traditional wearable devices or pressure sensor-based solutions, this system provides a more efficient and portable non-contact solution, offering an innovative option for enhancing health and efficiency in modern work environments.Clinical Relevance- The proposed system provides an efficient and flexible sitting posture recognition solution tailored to the diverse needs of modern office environments, contributing to improved workplace health and efficiency.