A Real-time Object Detection for WiFi CSI-based Multiple Human Activity Recognition

Abstract

Listen

Translate article

In the recent past, human activity recognition research has focused on using WiFi channel state information (CSI) as a viable alternative to legacy systems like video and sensor-based activity recognition having limitations such as privacy invasion, obtrusiveness, and the inconvenience of wearing sensory devices. While the performance of CSI-based activity recognition models is impressive, many of the models are built using offline processed data from regulated settings which hinders their application in real-time. However, real-life human activity recognition requires models to be responsive to identifying activities in real-time. To address the shortcoming of CSI-based activity recognition models, we propose a deep learning object detection framework and instance segmentation for multiple human activity recognition using WiFi signals. The real-time CSI data from the signal is captured on a sliding window and converted into time-frequency domain images of the activity stream using continuous wavelet transform (CWT). Since it is impossible to pre-segment activities within a stream in real-time, the power profile from the transformed images is exploited to provide insights for deep learning instance segmentation to identify each unique human activity. The evaluation is carried out using real-time CSI data with single and multiple human activities. The results show that real-time model classification accuracy is 93.80% on average and instance segmentation accuracy of 90.73%.