A Violence Recognition Method With Commercial WIFI Devices

School violence and bullying is a serious problem that greatly affects the healthy growth of the youth and the children.Current prevention measures mainly depend on propaganda and self-report.So far,there is still no effective solution that can automatically detect the bullying events.The essence of detecting violent bullying is human activity recognition(HAR).HAR has been widely studied in education,medical treatment,business and military fields at present.HAR has been studied in computer vision,wearable sensors and environmental sensor.However,these technologies are limited by some devices and scenarios,which can not meet the needs of violence recognition comprehensively.In recent years,with the popularity of commercial wireless Wi Fi,more and more research work has been done on HAR using WIFI signal.Human activity recognition with commodity WIFI has many advantages,such as low-cost,non-invasive,and ubiquitous etc..In this paper,we,based on the commercial WIFI infrastructure,build a ubiquitous passive violence detection system.The experimental results under different environment show that the proposed method performs superiorly in both accuracy and robustness.And,the main contributions are as follows:1.By analyzing the relationship between the waveform of different subcarriers and the same action of CSI,the correlated features of CSI are extracted;Dynamic threshold method is leveraged to detect the start and end of actions;CSI amplitude is transformed into pictures and input them into feature fusion algorithm to extract correlation features;Analyzing and comparing the influence of extracting correlated features and time series features on classification model.By comparing the accuracy of different machine learning classification algorithms,this paper studies the use of support vector machine to establish the classification model of violent activities.It is proved by experiments that SVM can improve the prediction accuracy of the system.2.Based on the extraction of time series features and correlated features,the features of the two perspectives are fused;In order to effectively combine the features of these two perspectives,we adopts two fusion strategies,one is PCA algorithm fusion,the other is based on multi-view,KCCA fusion technology,which combines the features of the two perspectives as the final input to the classifier model.In order to confirm the validity of the fused feature,the experiment verifies that the fused feature can improve the accuracy of the model.The effectiveness of PCA fusion method and KCCA fusion method is analyzed and compared,and the effect of each fusion method on the classifier is verified by experiments.Experiments verify the influence of many factors on the recognition system to ensure the robustness of the system.