On Human Motion State Detection Technology Based On WiFi

With the rapid development of intelligent technology,people's lives have gradually entered the era of information and intelligence.People's actual life needs have spawned various emerging technologies.Among these,human body motion detection technology has become a new research hotspot.The traditional vision-based and sensor-based human body motion detection methods have the disadvantages of high light requirements,and they are easy to leak privacy and hard to carry.The WiFi-based detection method can overcome the shortcomings of these methods,so it has become a popular research direction.Thus,this thesis focuses on using channel state information(CSI)of WiFi signals to detect the body movements and gestures of human motion state.The main research work in this thesis are:First,the host computer interface program is designed by modifying the source code of the network card driver,and the CSI human motion detection system platform based on Atheros 9590 and Intel 5300 network cards are built under the Linux system.Most of the existing research works based on CSI are built using the Intel 5300 network card with only 30 channels.The Atheros network card platform developed in this thesis can obtain 56 channels of CSI measurements.In addition,the system can automatically realize the data acquisition and classifying CSI data from different people and activities.The system is employed to collect the original data sets of body actions and gesture actions under different environments in this thesis.Secondly,traditional handcrafted feature extraction methods for human body motion detection are time-consuming and labor-intensive.In order to deal with these issues,this thesis propose an ensemble approach,i.e.CNN-SVM-RF,which integrates Convolutional Neural Network(CNN),Support Vector Machine(SVM),Random Forest(RF).First of all,convolutional neural network is employed to automatically extract features from raw CSI measurements,and the extracted features are inputted into CNN,SVM,RF classifiers for detection respectively.Then,a probability-based heterogeneous classifier integration method is used to fuse the probabilistic prediction results of three classifiers to obtain estimated output.The performance of the proposed algorithm is analyzed by collecting data on the human motion detection platform.The experimental results show that,compared with the s algorithms,proposed CNN-SVM-RF model can significantly improve the recognition accuracy.Finally,in order to deal with the low accuracy of existing human gesture detection and recognition methods,a Deep Convolution Decision Forest(DCDF)based recognition model is designed.This model embeds the traditional decision tree classifier into the deep network for end-to-end learning.First,a convolutional network is constructed to improve the feature expression ability of CSI gesture information.Second,this thesis proposes a probabilistic decision tree,and the split nodes of this tree are controlled by the path function with uncertainty.Then,the probabilistic decision forest composed of decision trees is embedded in a standard convolutional neural network for end-to-end training and learning.Finally,the experimental analysis proves that the DCDF model attains significant gains compared to the state-of-art human motion detection methods.