Research On Human Behavior Recognition Technology Based On WiFi

Behavior recognition plays an important role in environmental monitoring,smart medical treatment,smart furniture,human-computer interaction and other fields,and has always been a research hotspot in the academic community.At present,Wi Fi-based human behavior recognition technology research is a new direction in this field.Compared to the traditional human behavior recognition methods,such as sensor-based,image-based,UWB-based recognition,which overcomes the drawbacks of wearing,sensitivity to light,high recognition cost,and its perception sensitivity and recognition accuracy are high,suitable for human behavior recognition in indoor environments,so it will be an important technological innovation of intelligent furniture control.In this thesis,the CSI data of Wi Fi signals are collected and pre-processed and extracted actions from them,the dataset samples are established,and the Wi Fi behavior recognition technology is proposed in combination with behavior recognition algorithms.The main tasks of this thesis are as follows:1.Based on the Fresnel zone theory,the experiments are designed to obtain stable and reliable CSI data.A series of data preprocessing methods include removing the DC component of the signal,detecting and correcting outlier values,signal filtering,etc,to remove the influence of random noise in the environment.For the pre-processed data,a variable sliding window method based on the movement variance to identify the start and end of an action is proposed in this thesis.Experiments show that this action extraction method can take into account both high accuracy and high efficiency.2.Based on the experimental data extracted by the action,the evaluation index of subcarrier performance and the scheme of subcarrier selection and fusion are proposed,which reduce the data calculation amount by 2/3 and the accuracy of behavior recognition is improved.Comparing the common behavior recognition methods,the behavior recognition based on the improved CNN network is proposed,which increases the average accuracy of 7 kinds of behavior recognition in the non-interference environment to 90.74%.Among these 7 kinds of actions,the best behavior recognition accuracy rate is 99%,and the worst is still 83%.In order to verify the robustness,the experimental scene is replaced by other two rooms,and the average accuracy can still reach 82.43%and 88.29%.3.Based on the behavior recognition network established in the non-interference environment,the behavior recognition in the dynamic interference of a person is explored.By analyzing the time domain characteristics of signal strength,the proposed algorithm can determine whether there is interference in the environment with an accuracy rate of98.68%.The relationship between the identification accuracy and the interference distance is fitted based on the coefficient of 0.9909.Two key interference distances,4.43 m and 1.08 m,are estimated based on the fitting relation,which means the closest interference distance that is equal to non-interference environment and the closest interference distance that the behavior recognition network is effective.