Research On Unobtrusive Human Activity Recognition And Early Warning Algorithm Based On Wireless Channel State Information

Assuming a person will enter the toilet 6 to 8 times in one day,and then each person may spend two to three years in the toilet.Because of the toilet's poor ventilation or the slippery ground,many accidents such as fall,faint and even sudden death may happen in the toilet.The toilet has become a high incidence of accidents.Taking into account the privacy of the toilet itself,many of the accidents in the toilet are not timely treated,which will lead to some serious consequences,especially for the elderly.Therefore,the timely detection of accidents in the toilet is very important,which can adapt to the requirements of privacy and not disturb the user.To achieve the toilet monitoring is very practical significance.In this paper,we propose an unobtrusive toilet activity recognition and early warning method based on the Channel State Information(CSI).We can monitor the activity continuously in the toilet.This study can automatically identify activities and achieve independent interception of activities.We analyze the identified activities to find the risk of accidents and provide an early warning to help them timely.This paper mainly completed the following four aspects of the study.First,this paper proposes analysis of the accidents occurring in the toilet and establishes a model of toilet activities.We break down the toileting into several actions and we aim to identify a few key actions.With the key actions,we can determine whether there is an accident.Then a passive monitoring mode and preprocessing method based on channel state information are proposed.Based on the influence of human activities on CSI signal,a CSI data acquisition scheme for toilet is designed.We design a target detection algorithm which can adapt in the actual environment.After collecting data,the CSI data preprocessing methods are executed for ambient and internal noise.The principal component analysis can combine the characteristics of multiple CSI subcarriers and reduce the noise of CSI.An active interception algorithm based on anomaly detection is constructed which can meet the real-time requirement.For the traditional activity recognition algorithm requires a complex feature extraction method,this paper proposes an active recognition scheme based on convolution neural network.We use characteristics of convolution neural network that can improve extraction of features and reduce noise of channel state information spectrum.In this way,we avoid the complex feature extraction process and can adapt to different activities.The acquisition methods of the spectrum are mainly composed of discrete wavelet transform and short-time Fourier transform.Considering the few category of classified activities,we utilize Le Net-5's network architecture.We take data augmentation and dropout methods to avoid over-fitting.Finally,its result compares with the traditional support vector machine classifier using statistical features.Convolution neural network can achieve higher accuracy and avoid the complex feature extraction stage.This paper carries out a large number of experiments and evaluations,and collected a certain amount of real and reliable data.The results show that the accuracy of the activity recognition based on convolution neural network can reach about 93%,which is better than that of the traditional method.The results verify the effectiveness and reliability of the proposed method,and can adapt to different environments.The scheme can be executed in tolerable delay and meet the real-time requirements,and finally achieve a satisfactory accuracy.