A Cloud-Edge Collaborative Framework for Cross-environment Human Action Recognition based on Wi-Fi

Device-free human action recognition (HAR) based on Wi-Fi signals is an essential support in the field of the Internet of Things and shows bright application prospects. With the rapid development of deep learning (DL), HAR based on DL models has become mainstream and achieved good performance. However, most of these methods are still far from the practical application, the main challenges include poor cross-environment recognition performance and the high requirements for sensing devices of DL models. Based on this, we propose a cloud-edge collaborative HAR framework (Co-WiSensing), which explores the possibility of cross-environment HAR with low resource consumption. Considering the characteristic of the massive resources of cloud servers and the resource constraints of edge devices, a high-performance multi-branch cloud HAR model is delicately designed and the personalized model compression and offloading strategies are proposed to construct lightweight edge HAR models for different environments, this allows the edge users to realize perception under resource-limitation conditions. Extensive experiments are conducted to validate the effectiveness of the proposed framework. Experimental results show that our framework can provide better HAR accuracy across all environments while using less computation and storage cost than the state-of-the-art lightweight models.