Fielded autonomous posture classification systems: Design and realistic evaluation

Few Body Sensor Network (BSN) based posture classification systems have been fielded to date, despite laboratory based research work confirming their theoretical suitability for a range of applications. This paper reports and reflects on two algorithms which i) improve the accuracy of real-time, multi-accelerometer based posture classifiers when dealing with natural movement and transitions and ii) maximize a wearable system's battery life through distributed computation at nodes. The EWV transition filters proposed here increase the classification accuracy by 1% over unfiltered results in realistic scenarios and significantly reduces spurious classifier output in real-time visualizations. A 200 fold transmission reduction from the on-body system to an outside system was achieved in practice by combining the transition filters with an event-based design. Furthermore, a method of reducing transmissions between on-body data gathering nodes based on distributed processing of the classifier rules (but maintaining a one-way flow of communications during system use) is also described. This provides a 3.3 fold reduction in packets and a 13.5 fold reduction in data transmitted from one node to the other in a two-node wearable system.