A comparative study on real-time sitting posture monitoring systems using pressure sensors

Accurate sitting posture recognition plays a crucial role in improving improper postures and reducing the risk of associated health issues. The inherent complexity of human behavior, however, poses a great challenge to the development of a practical sitting posture monitoring system with pressure sensors. Towards facilitating the use of features, choice of classification models, and way of evaluating a sitting posture recognizer, in this study a comparative study on pressure-sensor-based sitting posture monitoring is conducted. Specifically, we extract discriminant features from the sensor data based on the distribution of pressure sensors and explore different combinations of these features. Then, five commonly used classification models are evaluated towards building a robust sitting posture recognizer. Finally, extensive comparative experiments concerning four performance metrics are conducted on the collected datasets in subject-dependent, subject-independent, and cross-subject settings. Results show that the joint use of sensors at different positions leads to higher accuracy and that random forest generally outperforms the other four classification models. Surprisingly, compared to the subject-dependent and subject-independent settings, cross-subject setting greatly suffers from degraded accuracy, where we preliminarily present the results of transfer learning techniques to mitigate this issue. In addition, we perform parameter sensitivity and time-cost analysis of random forest, which indicates its applicability to practical use.