Design Flow of Wireless Body Sensor Network for Human Activity Classification using Long Short-Term Memory (LSTM) Neural Network

The design aspects of a wireless body sensor network (WBSN) consider a variety of parameters to improve the network and algorithm but the placement choice often remains unjustified by the researcher. A common application of the WBSN is for the motion-capture of a limb or human, but sensor placement differs, making each algorithm very specific to a singular movement. In this paper, we explore the system architecture and design flow of an enhanced human activity classification algorithm using long short-term memory (LSTM) neural network. A WBSN system is designed and implemented to collect information for training the LSTM sequential model. The sensors are placed on waist, ankles and wrists to observe the relevance, pertinence and accuracy improvement on increasing the number of sensor nodes and positioning of the sensors. Result shows that with the presented system and algorithm, we can precisely characterize human positions and behaviors. This WBSN system can be further extended to understand different motions and different sensor positions, and further expanded to include other sensors.