A Wearable Multisensor Posture Detection System

Abstract

People of different backgrounds are experiencing strain in their back and neck or different parts of the body more than ever in today's time. This situation is the consequence of poor body posture and other factors that are part of our daily lifestyle and habits, such as sedentary work stations and reduced physical activity. Continuous and automatic posture detection and correction have been observed to be a reliable solution to this problem. In this work, identification, and classification of multiple postures in sitting as well as standing conditions are considered. Three triaxial accelerometer and gyroscope sensors are found to yield classification accuracies higher than those obtained using just one or two sensors. Principal component analysis of the extracted features shows sufficient distinction between the considered postures. Four different classifiers have been used, namely support vector machine, multi-layer perceptron, decision tree, and random forest (RF) out of which RF produced the highest accuracy. The RF classifier yields an average classification accuracy of 95.68% for 9 different body postures (5 sitting postures, 4 standing postures) performed by 7 subjects.