Human Body Posture Recognition Based On Empirical Mode Decomposition And Neural Network

Under the background of new-type human-computer interaction and the widespread popularization of smart devices,human body posture recognition technology is regarded by many researchers in various fields as an indispensable means of human-computer interaction.Human body posture belongs to a natural and intuitive,simple and effective way of interacting in many man-machine interaction modes.It is often used as a supplementary communication mode in human life.With the expansion of wireless network coverage and the emergence of smart home applications,gesture recognition based on WiFi signals is booming.Currently,the recognition based on wireless signal mainly rely on RSSI and CSI according to domestic and international references.However,the RSSI just roughly partition the state of motion,and the information might be lost after processing the fine-grained CSI.Therefore,through the full analysis of the IEEE 802.11a protocol,this article researches on the long preamble of WiFi data frame and consists of many fields such as wireless signal processing,pattern recognition,machine learning etc.The main work can be summarized as follows:(1)We use the USRP to complete the experimental design and data acquisition of human pose recognition based on WiFi signals.(2)We study feature extraction and classifier,and propose an improved algorithm based on Empirical Mode Decomposition(EMD).The method filters the original data according to the conditional formula,and then performs the reorganization of the filtered data.Ajfter feature being extracted,the SVM with high classification speed and high recognition rate is selected for classification and identification.It can be seen from the simulation that this method has higher recognition rate.(3)We analyze the principle and characteristics of neural network,and propose identification method based on CNN and NN.We use the vector extracted by CNN as the input of NN to complete the recognition.The CNN training time is optimized by this method,it means to reduce the complexity.We analyze the applicability of this model under noise interference.Experimental results show that this method can identify low-volume data quickly and accurately.