Cross-Scene Human Action Recognition Based On Wi-Fi Signal

Human action recognition based on Wi-Fi Channel State Information is a recently emerging technology,which can realize the perception of human activities at low cost,avoid privacy disclosure,and has the advantages of non-contact,and not affected by illumination changes.However,one of the major difficulties of Wi-Fi CSI-based action recognition technology is cross-scene recognition,which will bring a series of challenges,such as complex model,lack of target scene data and whether the action feature extraction algorithm can perform stably in the target environment.With the large-scale popularization of Wi-Fi,human perception based on Wi-Fi signal has gradually become a research hotspot.Through the research of cross-scene action recognition based on Wi-Fi CSI,the main research contents are as follows:(1)Propose a Wi-Fi human action recognition algorithm based on two-way symmetric dense connection.Wi-Fi human action signal records the change of human posture over a period of time,which has a certain correlation in time domain and spatial domain,if this characteristic is ignored,the final recognition result will lead to unsatisfactory.On the basis of existing research,a two-way symmetric densely connected structure for action recognition is proposed.The core of this algorithm is to mine the general feature information in Wi-Fi action signal effectively by enhancing the receptive field of the network and feature multiplexing.Even when the environment changes,the algorithm can effectively extract general feature information and obtain good recognition accuracy.In order to comprehensively evaluate the performance of the algorithm,standard data sets,irregular data sets and similar data sets are constructed to fully verify the actual recognition ability of the algorithm.Experiments show that the algorithm can achieve more than 95% recognition accuracy in different data sets and different environments.(2)Propose a cross-scene lightweight Wi-Fi human action recognition algorithm.One of the difficulties of action recognition based on Wi-Fi signal is that when the environment changes,the Wi-Fi action characteristics will change,so it usually needs to re-collect a large number of data sets,which is a huge workload and time-consuming.At present,there are some methods that can effectively solve the problem of insufficient data sets when the environment changes.However,these methods usually use complex algorithms,which will lead to a very bloated depth model.It is not conducive to deployment in some scenarios with poor computing equipment.On the basis of existing research,a lightweight action recognition algorithm for Wi-Fi action recognition is proposed based on knowledge distillation.The algorithm adopts a simplified network structure,and only a small amount of data is needed to achieve good cross-scene recognition accuracy when the environment changes.