Detection of Poor Posture using Wearable Sensors and Unsupervised Learning

Most working professionals as well as school and college students require very less physical activity and they are mostly seated during most part of the day. Maintaining a good posture while sitting is a step towards avoiding several lifestyle diseases that arise due to decreased physical activity. In this work, posture detection by placing multiple accelerometers on the back of user is presented. Novel use of one-class classification for detecting poor posture is proposed. One-class support vector machine and isolation forest are employed for posture detection. The classifiers are provided unlabeled data containing mostly samples of good posture. During testing, the poor postures, which we be different postures other than that considered as the good posture, are predicted as outliers. The performance of the models is compared quantitatively. The performance of one-class classifiers is also compared with commonly used unsupervised learning algorithm, k-mean clustering. Results are presented for tuning of the considered models and to study the effect of sensor placement. The isolation forest provides the highest accuracy of 99.3% and F1-score of 0.99.