Indoor Dangerous Behavior Recognition Based On Wi-Fi Signal

As more and more machine learning and deep learning algorithms improve,human activity recognition technology plays a pivotal role in security,smart home and health care and other fields.Traditional human activity recognition technology is mainly through devices such as cameras or wearable sensors and other equipment to collect activity data and recognize data.However,in an indoor environment,cameras have the problem of violating privacy,and wearable sensor devices not have universality.Therefore,human activity recognition based on Wi-Fi signals has become a new solution in this field.WiFi devices are inexpensive and widely available in indoor environments,making them very suitable as a signal source for indoor activity recognition.Initially,Wi-Fi received signal strength indication(RSSI)was used as the base signal for activity recognition,but it has been gradually phased out due to its low recognition accuracy.The channel state information(CSI)can describe the multipath changes caused by human motion,and realize more complex and more accurate activity recognition.After the CSI data of Wi-Fi signals can be obtained through commercial Wi-Fi network cards,the focus of activity recognition based on Wi-Fi signals gradually turns to CSI.Different from the previous work of using CSI data only to realize human activity recognition,this thesis uses the amplitude image generated by CSI data as the recognition data set.Compared with CSI data,which can only use algorithms to classify or predict continuous sequences to complete activity recognition,this scheme can use various image classification algorithms or models to achieve the goal of activity recognition.After the low-pass filtering algorithm and principal component analysis(PCA)were used to preprocess CSI data,amplitude images with more obvious features were obtained,which were used as classified samples to construct a data set.The data set not only subdivided actions,but also divided actions into normal behaviors and dangerous behaviors.The dataset is classified using the improved model Efficient Net with the most superior performance of the current convolutional neural network(CNN),and an adaptive progressive training method based on training accuracy is proposed to optimize the Efficient Net.The classification model generated by 5-fold cross-validation method has an average recognition accuracy of 95.3% for normal behaviors,92% for dangerous behaviors,and94.4% for all actions in the validation set.On the test set,the average recognition accuracy of the classification model generated before optimization by using the adaptive progressive training method based on training accuracy is 91.25%,while the average recognition accuracy of the optimized classification model is 93.13%.The experiment proves that the scheme of using CSI amplitude image to identify indoor dangerous behavior is feasible and effective,and achieves high identification accuracy.