Human posture recognition based on wearable sensor

In recent years, with the development of economy and technology, people were increasingly concerned about health, and at the same time, human behavior data based on wearable devices revealed great research value for human health. Currently, most human recognition study focused on video images, however, human behavior recognition using video images showed the following shortcomings: Firstly, it was necessary to install and fix the equipment in advance and the range observed by the equipment was limited; Secondly, it was the privacy controversy involved in human behavior recognition based on video images and the inconvenience caused by the joint parts obscured. Therefore, this study investigated human behavior recognition for wearable sensor data and constructed a body sensor-based behavior recognition system, which used a multilayer CNN-BiLSTM with one-dimensional convolution for extracting features, then combined a bi-directional long and short-term memory network (BiLSTM) that was good at processing time series data, and finally used a fully connected layer for classification to obtain the final The final recognition result was obtained by using a fully connected layer for classification. The system had been experimented on two publicly available datasets (MHEALTH and USCHAD), and achieved good effects, with recognition accuracy of 98.82% and 97.62% on the MHEALTH and HPAT datasets, respectively, showing that the system had good generalization capability.