Activity Classification With Empirical RF Propagation Modeling in Body Area Networks

Mobile sensor-based systems are emerging as promising platforms for remote healthcare monitoring. One popular application of these systems is to track the real-time body movements of a patient by analyzing and classifying the physiological signals collected by the video sensors or the body-mounted mechanical sensors. However, the existing motion monitoring infrastructures are inconvenient to be carried with the patient. In this paper, we explore the potential of using the inexpensive offthe- shelf inertial sensors that embedded in the smart phones to identify the body movements. In our proposed system, variance, energy, and frequency domain entropy of linear acceleration and rotating orientation are extracted from the inertial sensors to form the feature vector. To enhance the performance of the system, quantitative metrics of RF propagation characteristics: level crossing rate, Doppler Spread, coherence time, Root Mean Square (RMS) Doppler bandwidth and variation of Path Loss are also investigated to provide new descriptors to the feature space. These features are imported and tested by four most commonly used machine learning algorithms: Backpropagation network (BP), Probabilistic Neural Network (PNN), k-Nearest Neighbor algorithm (k-NN) and Support Vector Machine (SVM) algorithm. Results show that using features from both RF sensor and inertial sensor would greatly improve the classification accuracy.