Real-time sensor- and camera-based logging of sleep postures

This paper presents a process of feature selection, and classification algorithm evaluation for a continuous sleep monitoring system, using a tri-axial accelerometer attached to the subject's chest. Two feature selection algorithms, i.e., Relief-F and support vector machine recursive feature elimination (SVM-RFE), and seven classification algorithms, i.e., Bayesian network, naive Bayesian network, support vector machine, pruned decision tree, instance-based learning with one neighbor, instance-based learning with three neighbors, and multi-layer perceptron, were investigated. By using four features according to the rank obtained from Relief-F, and a multi-layer perceptron classifier, an average accuracy of 85.68 percent has been achieved. Based on the selected model, a real-time logging system of sleeping images triggered by a sleep posture change detected using a wireless sensor node has been developed.