A multi-person identity recognition method based on commercial WiFi signals

Abstract

A multi-person identity recognition method based on commercial WiFi signals, firstly, uses a transceiver device to build a channel state information collection platform, the data is preprocessed and enhanced, and then a tensor similar in size to a visual image is output through a deconvolution network, and input into a deep convolutional neural network to extract discriminative identity features, and finally the features are used to identify the human body. The present invention can identify multiple people at a relatively low cost and can achieve an ideal recognition accuracy rate. The present invention has certain application value in the fields of intelligent access control and security in the case of multiple people.

Description

Technical Field

[0001] The present invention relates to the field of wireless sensing technology, and in particular to a method for applying WiFi signals to human identity recognition through wireless sensing. Background Art

[0002] Human identity recognition is an important aspect of human perception. If the biological characteristics of the human body can be judged and estimated, then each person's identity can be accurately identified through these characteristics.

[0003] With the rapid development of Internet of Things technology, WiFi devices are widely deployed in various locations. WiFi signals can be used in various fields of wireless sensing, including personnel tracking, indoor vital signs monitoring, human-computer interaction and gesture recognition. Among them, WiFi-based human identity recognition technology has attracted the interest of many researchers. Compared with traditional human identity recognition methods, WiFi-based recognition methods do not require close contact with any sensor equipment, are not restricted by light conditions, and do not cause privacy issues.

[0004] Currently, WiFi-based human identity recognition methods have achieved good results, but there are still some shortcomings. For example, most methods require users to walk back and forth along a predetermined path, which has certain limitations on application scenarios. In addition, these methods can only achieve a high recognition accuracy rate when there are no more than 10 people. When the number of people increases, the accuracy rate will drop rapidly. Summary of the invention

[0005] In order to overcome the shortcomings of existing human identity recognition detection methods, such as complex deployment, low recognition accuracy and small number of recognized people, the present invention proposes a multi-person identity recognition method based on commercial WiFi signals to realize human identity recognition, which has simple deployment and can still maintain relatively high recognition accuracy when the number of recognized people gradually increases.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is:

[0007] A method for multi-person identity recognition based on commercial WiFi signals comprises the following steps:

[0008] Step 1: Build a channel state information (CSI) data collection platform;

[0009] Step 2: Build a human feature library through the offline training phase. The human body stands still between the transceiver devices, and different CSI feature data are generated by using the impact of the human body on electromagnetic waves. Each pair of antennas generates 30 subcarrier CSI data, with a total of RxN pairs of data, where R is the number of transmitting antennas and N is the number of receiving antennas.

[0010] Step 3: Take the CSI data on one pair of antennas and pre-process it. The process is as follows:

[0011] Step 3-1: Filter the raw data using a Butterworth low-pass filter;

[0012] Step 3-2: Perform CSI data enhancement;

[0013] Step 3-3: Use PCA principal component analysis to reduce the dimension of the data;

[0014] Step 4: The processed data is input into the generation module for size modification;

[0015] Step 5: Data is input into the feature extraction network to extract identity features;

[0016] Step 6: Send the extracted features into the deep convolutional neural network for identity recognition training;

[0017] Step 7: Get data of other antenna pairs and repeat steps 3 to 6;

[0018] Step 8: The identity features extracted from the above steps are used as the fingerprint information of each person to complete the establishment of the human feature database;

[0019] Step 9: Perform human identity recognition in the online test phase and also collect test data packets;

[0020] Step 10: Process the test data, specifically steps 3 to 7;

[0021] Step 11: Perform convolutional neural network identity recognition and testing on each data sample of the test data;

[0022] Step 12: Output the identification result.

[0023] Furthermore, in step 3-2, the CSI data enhancement method is as follows: the original CSI data sample set is denoted as: T = {t 1 ,t 2 ,…,t n}, where t j is a sample element, j∈{1,n}, and n is the total number of samples; the added enhanced segment CSI is generated using the following formula:

[0024]

[0025] In the formula, every k elements of the entire data set are divided into a group, so there are a total of [N / k] groups, t j is a single element data in each group, and then the arithmetic mean is calculated for the i-th group Finally, the CSI dataset after data enhancement

[0026] Furthermore, in step 4, the CSI data is processed from 30×1×1 through 8 deconvolution layers to generate a tensor of the size of the input image of the convolutional neural network.

[0027] Furthermore, in step 5, feature extraction uses a DenseNet network, and the feature map output by the previous layer of each DenseLayer is used as the input of all subsequent layers. The input of the lth layer is as follows:

[0028] x l =H l ([x 0 ,…,x l-1 ])

[0029] where [x 0 ,…,x l-1 ] refers to the feature map from the previous 0 to l-1 layers, H l It is a l-level cascade.

[0030] Furthermore, in step 6, the feature map after feature extraction by the convolutional neural network is input into the adaptive mean pooling layer, and the one-dimensional tensor after pooling is predicted by the fully connected neural network to predict the possibility of each identity. The identity recognition calculation is as follows:

[0031]

[0032] in is the identification number, K is the number of all users, and argmax() calculates the output identity vector prediction value (p 1 ,p 2 ,…,p K ) is the subscript of the highest predicted value in .

[0033] The beneficial effects of the present invention are:

[0034] 1. Make full use of wireless LAN equipment as the experimental platform, which is simple to deploy, easy to operate and has high recognition accuracy.

[0035] 2. The present invention does not require the human body to move back and forth, is simple and reliable, and has certain application value in the fields of intelligent identification, security, etc.;

[0036] 3. The present invention applies the user's unique biological characteristics (body shape, body fat, muscle, etc.) to human identity recognition, providing a new research idea for human identity recognition;

[0037] 4. Deep convolutional neural networks are mainly used for feature extraction and identity recognition, and they also have good recognition effects when there are a large number of people to be recognized. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is the overall structural diagram of the indoor human identity recognition system;

[0039] Figure 2 It is a specific implementation diagram of the data acquisition and detection system;

[0040] Figure 3 The figure is a comparison of the original CSI and filtered CSI signal amplitudes. (a) is the original background and human CSI amplitude, and (b) is the filtered background and human CSI amplitude.

[0041] Figure 4 It is the structural diagram of feature extraction module;

[0042] Figure 5 It is a network structure diagram for identity recognition and classification of feature graphs;

[0043] Figure 6 It is the recognition accuracy under different number of people. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0045] Reference Figure 1 to Figure 6 , a method for multi-person identity recognition based on commercial WiFi signals, comprising the following steps:

[0046] Step 1: Build a channel state information (CSI) data collection platform;

[0047] Step 2: Build a human feature library through the offline training phase. The human body stands still between the transceiver devices, and different CSI feature data are generated by using the impact of the human body on electromagnetic waves. Each pair of antennas generates 30 subcarrier CSI data, with a total of RxN pairs of data, where R is the number of transmitting antennas and N is the number of receiving antennas.

[0048] Step 3: Take the CSI data on one pair of antennas and pre-process it. The process is as follows:

[0049] Step 3-1: Filter the raw data with a Butterworth low-pass filter to remove high-frequency jitter noise;

[0050] Step 3-2: Perform CSI data enhancement to improve model robustness;

[0051] The CSI data enhancement method is as follows: The original CSI data sample set is denoted as: T = {t 1 ,t 2 ,…,tn}, where t j is a sample element, j∈{1,n}, and n is the total number of samples; the added enhanced segment CSI is generated using the following formula:

[0052]

[0053] In the formula, every k elements of the entire data set are divided into a group, so there are a total of [N / k] groups, t j is a single element data in each group, and then the arithmetic mean is calculated for the i-th group Finally, the CSI dataset after data enhancement

[0054] Step 3-3: Use PCA principal component analysis to reduce the dimension of the data;

[0055] Step 4: The processed data is resized through the input generation module; the CSI data is processed from 30×1×1 through 8 deconvolution layers to generate a tensor of the input image size of the convolutional neural network;

[0056] Step 5: Data is input into the feature extraction network for identity feature extraction; feature extraction uses the DenseNet network, and the feature map output by the previous layer of each DenseLayer is used as the input of all subsequent layers. The input of the lth layer is as follows:

[0057] x l =H l ([x 0 ,…,x l-1 ])

[0058] where [x 0 ,…,x l-1 ] refers to the feature map from the previous 0 to l-1 layers, H l It is a l-level cascade;

[0059] Step 6: The extracted features are sent to the deep convolutional neural network for identity recognition training; the feature map after feature extraction by the convolutional neural network is then input into the adaptive mean pooling layer, and the one-dimensional tensor after pooling is predicted by the fully connected neural network to predict the possibility of each identity. The identity recognition calculation is as follows:

[0060]

[0061] in is the identification number, K is the number of all users, and argmax() calculates the output identity vector prediction value (p 1 ,p 2 ,…,p K) is the subscript of the highest predicted value;

[0062] Step 7: Get data of other antenna pairs and repeat steps 3 to 6;

[0063] Step 8: The identity features extracted from the above steps are used as the fingerprint information of each person to complete the establishment of the human feature database;

[0064] Step 9: Perform human identity recognition in the online test phase and also collect test data packets;

[0065] Step 10: Process the test data, specifically steps 3 to 7;

[0066] Step 11: Perform convolutional neural network identity recognition and testing on each data sample of the test data;

[0067] Step 12: Output the identification result.

[0068] In this embodiment, the experiment was conducted in a 6.1m×4m room with some chairs and tables for daily use. Two laptops equipped with Intel 5300 wireless network cards were used as transmitters and receivers. The distance between the transmitting antenna and the receiving antenna was 1.14m, and the height of the transmitting and receiving antennas was 1.15m. Both laptops were running on the Ubuntu operating system, and the system kernel was modified and adjusted to the wireless network card driver to use the 802.11n csitool. Figure 2 shown.

[0069] During the training phase, each time the biometric features of a person are collected, the person stands Figure 2 The data collection points are used to collect data packets containing channel state information; each collection time is 30 seconds. After the collection is completed, each person obtains about 3000 CSI samples, and then steps 3 to 7 of the present invention are used to extract features from the collected data and train the overall network model;

[0070] During the testing phase, the human body Figure 2 The test data is collected at the test point in the test data set, and the collection time for each person is 30 seconds; after the above processing, the test data is identified using the feature extraction network and identity recognition network of the present invention, and each network framework is as follows: Figure 4-6 As shown, in order to verify the performance of the present invention under different conditions, according to the detailed process of steps 9 to 11 in the content of the invention, the above steps are repeated when the number of training samples is reduced and the number of people is increased; thus, we obtain the performance curve as shown in Figure 6 As shown in the results, the recognition accuracy has been greatly improved, and it still has a high recognition accuracy among a large number of people, showing a significant advantage, and also has a significant advantage in training time.

[0071] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A multi-person identity recognition method based on commercial WiFi signals. It is characterized in that The method comprises the following steps: Step 1: Build a channel state information (CSI) data collection platform; Step 2: Build a human feature library through the offline training phase. The human body stands still between the transceiver devices, and different CSI feature data are generated by using the impact of the human body on electromagnetic waves. Each pair of antennas generates 30 subcarrier CSI data, with a total of RxN pairs of data, where R is the number of transmitting antennas and N is the number of receiving antennas. Step 3: Take the CSI data on one pair of antennas and pre-process it. The process is as follows: Step 3-1: Filter the raw data using a Butterworth low-pass filter; Step 3-2: Perform CSI data enhancement; Step 3-3: Use PCA principal component analysis to reduce the dimension of the data; Step 4: The processed data is input into the generation module for size modification; Step 5: Data is input into the feature extraction network to extract identity features; Step 6: Send the extracted features into the deep convolutional neural network for identity recognition training; Step 7: Get data of other antenna pairs and repeat steps 3 to 6; Step 8: The identity features extracted from the above steps are used as the fingerprint information of each person to complete the establishment of the human feature database; Step 9: Perform human identity recognition in the online test phase and also collect test data packets; Step 10: Process the test data, specifically steps 3 to 7; Step 11: Perform convolutional neural network identity recognition and testing on each data sample of the test data; Step 12: Output the identification result; In step 3-2, the CSI data enhancement method is as follows: The original CSI data sample set is denoted as: T = {t 1 ,t 2 ,…,t n }, where t j is a sample element, j∈{1,n}, and n is the total number of samples; the added enhanced segment CSI is generated using the following formula: In the formula, every k elements of the entire data set are divided into a group, so there are a total of [N / k] groups, t j is a single element data in each group, and then the arithmetic mean is calculated for the i-th group Finally, the CSI dataset after data enhancement In step 6, the feature map after feature extraction by the convolutional neural network is input into the adaptive mean pooling layer, and the one-dimensional tensor after pooling is predicted by the fully connected neural network to predict the possibility of each identity. The identity recognition calculation is as follows: in is the identification number, K is the number of all users, and argmax() calculates the output identity vector prediction value (p 1 ,p 2 ,…,p K ) is the subscript of the highest predicted value in .

2. The method for identifying multiple people based on commercial WiFi signals as claimed in claim 1, It is characterized in that In step 4, the CSI data is processed from 30×1×1 through 8 deconvolution layers to generate a tensor of the size of the input image of the convolutional neural network.

3. The method for identifying multiple people based on commercial WiFi signals as claimed in claim 2, It is characterized in that In step 5, feature extraction uses a DenseNet network, and the feature map output by the previous layer of each DenseLayer is used as the input of all subsequent layers. The input of the lth layer is as follows: x l =H l ([x 0 ,…,x l-1 ]) where [x 0 ,…,x l-1 ] refers to the feature map from the previous 0 to l-1 layers, H l It is a l-level cascade.

Images