Based On CSI And Wi-Fi Improved Convolutional Neural Network For Military Training Action Recognition

Action recognition technology has made great progress today.In order to seek more refined and diversified practical application scenarios,the technology itself has gradually pursued into various fields of refinement.Among them,action recognition for military training is an emerging and important branch in the field of action recognition.Dynamic monitoring and accurate identification of cadets’ movements during military training is an important method to test and evaluate the standardization and standardization of their military training movements,and it is also an effective method to ensure that their military training activities are in a state of effective guidance.There are mainly three kinds of early action recognition methods: video-based recognition method,wearable device-based recognition method and wireless radio frequency signal-based recognition method.Among them,the video-based identification method and the wearable device-based identification method both have problems such as large demand for data samples,high calculation and processing costs,poor technical universality,and unfriendly user experience.The identification method based on radio frequency requires personnel to have a high professional knowledge base,and at the same time,it requires extremely high hardware configuration.The emerging Wi-Fi signal-based human motion recognition method does not require heavy professional equipment,has low software and hardware computing costs,and can effectively protect user privacy.It has gradually become a popular research direction for human motion recognition in recent years.Based on the CSI data in Wi-Fi signals,this paper studies common military training action classification and recognition problems.The main work and innovations are as follows:(1)Research related theories based on CSI.First,the basic theoretical knowledge of CSI is introduced,followed by the signal processing theory based on CSI,and finally the theory of convolutional neural network model.(2)Aiming at the problem of action recognition in military training,this paper proposes a data augmentation strategy based on multi-subcarrier mixing(MSM)and a lightweight Sha-CNN(Shallow-Convolutional Neural Network)network model for action recognition.In the data preprocessing stage,the MSM data enhancement strategy is more suitable for the data distribution characteristics of CSI.Based on the interception and rearrangement of the CSI amplitude matrix,information far exceeding the amount of original CSI data can be obtained in less time domain.At the same time,the Sha-CNN network model designs the relevant core parameters and uses the dual-channel strategy for feature fusion classification,so that the model can extract more comprehensive CSI data features with lower model complexity,so as to achieve a balance between computational cost and recognition accuracy.The experimental results show that the CSI action recognition method that combines the MSM strategy and the Sha-CNN model has a high recognition accuracy rate for military training actions and a low computational cost.(3)In view of the problem that the data is easily affected by various factors such as environment and human beings,this paper designs three kinds of variable comparison experiments,including different batches of volunteers,different action poses and different experimental scenarios,to verify that the Sha-CNN network is in the real world.Robustness in the scene;at the same time,for the problem of small number of samples caused by the difficulty of CSI data acquisition,this paper uses the multi-subcarrier mixing strategy(MSM)data enhancement method to expand the data set,so that the convolutional neural network can extract More sufficient CSI feature information;finally,the Sha-CNN model is compared with the SVM and KNN models under the same conditions.The experimental results show that the Sha-CNN network designed in this paper has good robustness in real scenes,and after adopting the MSM data enhancement strategy,the recognition performance of the model is improved,and it is better than the current mainstream recognition methods.