Multiple Human Activities Classification Based on Dynamic On-Body Propagation Characteristics Using Transfer Learning
Yanyang Zhang, Yu Shao, Rui Luo, Lian Xiong, Jie Zhang
Abstract
Human activity recognition and classification have been widely used in various fields. To make full use of wireless body area networks (WBANs), multiple human activities classification based on dynamic on-body propagation characteristics is presented in this article. Four on-body links are established to collect propagation data set for four groups of activities, which include hand activities, arm activities, leg activities and head activities. A total of 30 human activities are considered. The line-of-sight (LOS) and non-LOS (NLOS) on-body propagation characteristics are analyzed in detail by comparing measurement with full-wave simulation. Multiple human activities, including some fine-grained motions, can be accurately classified using the propagation data collected by the on-body antennas. Based on the similar propagation mechanism among different links, an interlink transfer learning framework is proposed by pretraining the deep convolutional neural network (DCNN) on one link before training it on other links. The results show that after pretrained in source domain, the interlink transfer learning can improve classification accuracy and accelerate model convergence in target domain with a small amount of training data, which alleviates the complex and time-consuming data collection. The results of this article are particularly useful for the deployment of WBAN integrated communication and sensing.