Real-time Sitting Posture Recognition of Drivers Based on a Force Sensing Resistor Using Machine Learning

Abstract

Monitoring sitting posture has received increasing attention for predicting daily life issues in healthcare and rehabilitation. Moreover, recent studies have indicated drivers suffer from physical pain by exposure to the risky posture, such as back pain and sciatica. Accordingly, we develop a real-time recognition system of driver’s posture by using an intelligent cushion which includes 16 force sensing resistors (FSR) on the seat and backrest. Our experiment consists of two steps: 1) We created classifiers using 16 subjects’ offline data in the preliminary experiment and 2) performed the test of classification rate in the main experiment to estimate the posture in real time. Various supervised machine learning techniques were applied to the pressure values measured by the developed sitting posture recognition system, and the SVM using the RBF kernel showed the highest classification rate of 96.96% The highest classification accuracy of 96.96% were achieved with the SVM classifier using a radial basis kernel compared to other techniques. In addition, PCA-based feature selection method was used to identify the important sensors for monitoring driver’s sitting postures in the SVM classifier using a radial basis kernel. We found that the sensors on the backrest is more effective for the posture estimation than the sensors on the seat. Although we achieved high performance for classifying driver’s sitting posture in real-time, the performance of classifiers for each subject suggest that the classifier was overfitted to specific personal characteristics. By recruiting more subjects, the classifier will become more suitable and more accurate for the general person. Furthermore, PCA-based feature selection method is linear transformation method which are not adequately detect more complex non-linear feature interactions. Nonlinear feature selection algorithms such as nonlinear kernel-based SVM recursive feature elimination method are expected to improve for choosing more optimal deployment of sensors.