Posture detection with body area networks

Abstract

Body posture detection is extremely useful in health monitoring and rehabilitation. We develop a method to detect body posture that uses signal strength measurements from sensor nodes forming a Wireless Body Area Network (WBAN). We assume that postures (formations) take values in a discrete set and develop a composite hypothesis testing approach which uses a Generalized Likelihood Test (GLT) decision rule. The GLT rule distinguishes between a set of probability density function (pdf) families constructed using a custom pdf interpolation technique. The GLT is compared with the simple Likelihood Test (LT). We also adapt one prevalent supervised learning approach, Multiple Support Vector Machine (MSVM), to compare with our probabilistic methods. Due to the highly variant measurements from the WBAN, and these methods' different adaptability to multiple observations, our analysis and experimental results suggest that GLT is more accurate and suitable for posture/formation detection. Even for very similar postures in our experiments, GLT demonstrates high detection accuracy (around 97% with multiple observations). Besides the body area networks, the formation detection problem has interesting applications in autonomous robot systems.