Port Personnel Behavior Recognition System Based on WiFi Cross-Band Communication

Abstract

The present invention relates to a port personnel behavior recognition system based on WiFi cross-band communication, belonging to the field of Internet of Things technology. This system utilizes WiFi signals for cross-band communication to obtain multi-channel CSI data, and identifies personnel behaviors through feature extraction and classification algorithms. For the first time in behavior recognition, the present invention adopts the cross-frequency communication method to expand the feature library, and proposes a channel selection algorithm based on feature weights, effectively improving the recognition accuracy. The present invention also proposes a high-rate cross-frequency communication scheme that can be implemented on commercial WiFi devices, ensuring the synchronization of multi-channel data. The present invention can be applied to fields such as port / wharf personnel safety monitoring, behavior analysis, and human-computer interaction, and has broad application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of the Internet of Things, and relates to a port personnel behavior recognition system based on WiFi cross-band communication. Background Art

[0002] With the rapid development of science and technology and the popularization of computers in the 21st century, human-computer interaction has become the focus of attention and research in many countries, and human behavior recognition occupies an important position. Human behavior recognition is of great significance for promoting the development of human-computer interaction technology and improving the quality of human life. Traditional human behavior recognition technologies generally require the use of special devices. For example, human behavior recognition based on computer vision and human behavior recognition based on wearable sensor devices. Although their recognition accuracies are relatively high, the device prices are expensive, and the application scenarios and scopes of action are limited, and they will cause discomfort to people and involve personal privacy issues. Some scholars have proposed DI-Gesture, which is a real-time gesture recognition system with an average recognition rate of up to 97%. However, due to the many limitations of millimeter-wave radar, it is currently difficult to be promoted to actual life scenarios.

[0003] In recent years, with the rapid development of wireless communication technology, human behavior recognition technology based on WiFi perception has been widely studied. Currently, the behavior recognition technologies that can be practically applied to commercial WiFi devices are mainly divided into two categories. The first is based on the Received Signal Strength Indication (RSSI). Since the performance of the RSSI-based system is greatly affected by the link quality. In a complex environment, multipath propagation will cause fluctuations in the link quality, so good performance cannot always be guaranteed. The second is based on Channel State Information (CSI). CSI is a fine-grained value device derived from the WiFi physical layer. It can reflect the frequency response of the channel to capture the phase and amplitude information of different subcarriers of the signal. More information is extracted to capture the subtle changes in the link quality. Therefore, the CSI-based system has better robustness to complex environments.

[0004] Although in a channel state information (CSI) system, high recognition accuracy rates are achieved when identifying various human behaviors with high resolution, their deficiencies can be divided into the following three aspects: (1) There is no evaluation of the recognition accuracy rate for similar behaviors, or when identifying similar behaviors, their recognition accuracy rates are significantly low. (2) When using various classification algorithms based on deep learning, they all select a sufficient number of training sample quantities, but in some special scenarios, it is usually impossible to obtain a large number of behavior samples in a short period of time. (3) The CSI they obtain only comes from a single channel, and the presence of noise will affect the accuracy rate of the classifier. Moreover, in the CSI system, not only is there environmental noise, but the carrier frequency offset inside the device will also cause burst noise. And as shown in "Wireless Communications" written by Professor A. Goldsmith, co-channel interference will also cause unstable mutations in the amplitude and phase of the CSI. Therefore, even after filtering, there will still be some noise remaining. These uncontrollable factors will seriously affect the recognition accuracy rate of the behavior recognition system working on a single channel. Especially for behaviors that are already similar, if the amplitude or phase of any subcarrier in the CSI undergoes a jump due to interference, it may lead to a misjudgment of the final behavior category.

[0005] Based on obtaining CSI using WiFi signals, to solve the above problems, in the present invention, WI-FHOP: a behavior recognition system based on WiFi cross-band communication is proposed, and it can be implemented on a commercial WiFi device ESP32. WI-FHOP can maintain a high recognition accuracy rate when identifying similar behaviors and with a small number of training samples. First, a high-speed frequency hopping communication scheme suitable for the commercial WiFi device ESP32 is proposed, and CSI data of different channels is obtained; then, the obtained CSI data is processed through steps such as filtering, behavior data interception, and data division according to channels; after that, principal component analysis, wavelet transform, feature extraction, and standardization steps are performed on the data divided according to channels for feature analysis and extraction; then, a channel selection algorithm based on feature weights is proposed. This algorithm combines traditional classification algorithms (KNN, SVM, Bagging), selects the optimal channel according to the feature data of multiple channels, and uses the recognition result of this channel as the final recognition result. This avoids the influence of burst noise on a single channel and improves the recognition accuracy of the system in scenarios of identifying similar actions and with a small number of training samples; finally, the performance of WI-FHOP is evaluated through experiments. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a port personnel behavior recognition system based on WiFi cross-band communication.

[0007] To achieve the above purpose, the present invention provides the following technical solutions:

[0008] The behavior recognition system based on WiFi cross-band communication according to the present invention includes the following steps:

[0009] Step 1: The present invention selects channels 1, 6, and 11 in the 2.4G WiFi band. In cross-band communication, it is crucial to ensure that the channel settings at the transmitting and receiving ends are consistent when transmitting and receiving data. However, due to the asynchronous clocks of the devices at the transmitting and receiving ends, the time for switching channels is different, resulting in failed data transmission, which is the main reason affecting the cross-band communication rate. Based on this, a cross-band communication scheme is proposed, and this scheme is used to collect behavior data samples.

[0010] Step 2: Data processing is performed on the continuous CSI data received by the receiving end on multiple channels during the entire test period, including filtering out high-frequency noise in the original CSI using a Butterworth low-pass filter, extracting row data using a sliding window technique, and dividing data by channel.

[0011] Step 3: Behavioral feature analysis and extraction are performed on the amplitude data on the channels, including reducing the dimension of the CSI amplitude using the principal component analysis (PCA) algorithm, then performing wavelet transform (DWT) on each PCA principal component to extract time-frequency features, extracting statistical features from the feature matrix after DWT (selecting 6 different statistical features: the mean, standard deviation, interquartile range, 0.5 percentile, 0.683 percentile, and 0.95 percentile of the probability distribution), processing the feature data using the z-score normalization method, and converting all features to the same scale.

[0012] Step 4: After the data processing in the steps described above, the present invention proposes a channel selection algorithm based on feature weights. No matter which benchmark classification algorithm is combined, such as KNN, SVM, or Bagging, it can significantly improve the recognition accuracy.

[0013] The beneficial effects of the present invention are as follows:

[0014] 1. The present invention first adopts the cross-band communication method in behavior recognition. The same action data can reflect different features on multiple different channels respectively, so as to expand the feature library.

[0015] 2. The present invention proposes a channel selection algorithm based on feature weights. Based on the feature data on multiple channels, it combines traditional classification algorithms to select an optimal channel, and uses the recognition result on this channel as the final recognition result, so as to avoid the influence of sudden noise on individual channels on the final recognition result and at the same time improve the recognition accuracy of the system in scenarios where similar behaviors are recognized and the number of training samples is small.

[0016] 3. The present invention proposes a high-rate cross-frequency communication scheme that can be implemented on a commercial WiFi device ESP32, with a packet transmission rate of up to 200 Hz, and ensures that the channels experienced by the CSI on each group of multi-channels in the WI-FHOP system are non-time-varying.

[0017] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0019] Figure 1 is a flowchart of the present invention;

[0020] Figure 2 is a schematic diagram of the cross-frequency communication principle;

[0021] Figure 3 is a schematic diagram of the deployment of this system in the terminal yard;

[0022] Figure 4 is a warning flowchart. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The following uses specific specific examples to illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention schematically, and the following embodiments and the features in the embodiments can be combined with each other without conflict.

[0024] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and cannot be understood as a limitation to the present invention; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged, or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0025] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, it is based on the orientation or positional relationship shown in the accompanying drawings. This is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and cannot be construed as a limitation of the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0026] Figure 1 This is a flowchart of the present invention, and the steps are as follows:

[0027] Step 1: Consider the CSI signal propagation environment. The environment includes static reflectors (walls) and dynamic reflectors (people). The CSI data calculated by the receiving end should be the superposition of multiple path signals, which can be expressed by formula (1) as:

[0028]

[0029] Where H(f j , t) and arg(H(f j , t)) respectively represent the amplitude and phase of the jth (j = 1, 2,..., 51) subcarrier at time t, f j is the frequency of the jth subcarrier, e -j2πΔft represents the phase offset caused by the inconsistent carrier frequencies, N represents a total of N multipaths, a k (f j , t) is a complex number, which represents the propagation attenuation and initial phase offset of the kth multipath, represents the phase offset caused by the propagation delay τ k (t), and n(f j , t) represents noise. When there is a behavioral action of a dynamic target in the link, the propagation of a certain multipath (assumed to be k°) in the environment changes, and the corresponding a k° (f j , t) and τ k°(t) also changes, resulting in a change in the CSI amplitude of the received signal. By extracting the CSI data blocks corresponding to human activities and analyzing their characteristics, different behavioral activities can be distinguished based on different characteristics. The WI-FHOP system uses cross-frequency communication to combine and utilize the different amplitude characteristics of the same behavior reflected in different frequency bands to improve the recognition accuracy. Therefore, it is hoped that the channels experienced by each group of multi-channel CSI data received by the receiver are time-invariant. Because only when the multi-channel CSI data corresponds to the same time series, the different characteristics of the same action reflected in different frequency bands are comparable. In human action recognition, all data packets received within the channel coherence time can be considered to arrive at the receiver simultaneously. Therefore, the total time interval of each group of CSI data on multiple channels received by the receiver should be less than the channel coherence time. The channel coherence time T c can be expressed as where f d is the Doppler frequency shift. Previous studies have investigated the Doppler frequency shift caused in typical mobile environments. Generally, the Doppler frequency shift caused by human activities is less than 12 Hz, and the corresponding T c is approximately 40 ms. Assuming that the arrival times of a group of CSI data containing 3 channels at the receiver are t1, t2, and t3 respectively, the above content can be expressed by formula (2) as:

[0030] (t3 - t2) + (t2 - t1) ≤ 40 ms (2)

[0031] Therefore, this paper proposes a high-rate cross-frequency communication scheme. The frequencies of the same subcarriers on different channels are different, that is, f in formula (1) j is different. According to formula (1), this makes the CSI amplitudes of different frequency bands different. In addition, the reason why channels 1, 6, and 11 in the 2.4GWiFi frequency band are selected in this paper is that in the 2.4GWiFi frequency band, the spectra of channels 1, 6, and 11 do not overlap with each other, and the burst noises they receive are also uncorrelated. This results in a reduction in the correlation of the CSI amplitude characteristics of different frequency bands. Because for the same human behavior, it reflects different characteristics in different frequency bands. This is equivalent to communicating on 3 different channels on the basis of a single frequency band, which can not only avoid the influence of burst noises on certain frequency bands on the recognition accuracy, but also expand the number of its characteristics to 3 times the original. The more the number of characteristics, the more fully a behavior can be described. In cross-frequency communication, the most important thing is to ensure that the channel settings at the transmitting and receiving ends are consistent when sending and receiving data. However, due to the clock asynchronization of the devices at the transmitting and receiving ends, the time for switching channels is different, which leads to the failure of sending data. This is the main reason affecting the cross-frequency communication rate. To address this problem, the proposed cross-frequency communication scheme is as Figure 2. First, the sender sends data on Channel 1. The receiver receives the data at time t1 and almost immediately sends an ACK signal back to the sender. Then, at time t2, the channel is switched to the next channel, Channel 6. The sender receives the ACK signal at time t3, switches the channel to the next Channel 6 at time t4, and sends data on Channel 6 at time t5. It can be seen that after the sender completes the channel switch, it does not immediately send data but waits for a delay t d before sending. This is because in reality, due to API limitations, there is an additional delay in the actual start time of the channel switch. This means that the actual time t4 when the sender completes the channel switch may be less than the actual time t2 when the receiver completes the channel switch. Therefore, a delay t d is artificially imposed after the sender completes the channel switch to ensure that t5 > t2. The subsequent channel switch process is the same.

[0032] Step 2: First, use a Butterworth low-pass filter to filter out the high-frequency noise in the original CSI data for each CSI data. Then, according to Equation (1), when there is a human behavior action, H(f j ,t) will change. Therefore, it is possible to analyze and determine whether a human behavior is detected at the current time based on the change in H(f j ,t), and intercept the CSI amplitude data block corresponding to each behavior action for subsequent behavior recognition. In this paper, the sliding window technique is used, and the variance within the sliding window is selected as the metric to characterize the signal jitter, and the change of the data stream within the sliding window is extracted to analyze the signal jitter situation in a short time. The sliding window mean Δm of the amplitude of the j-th subcarrier in the t-th sliding window j,t The calculation formulas of Equation (3) and variance x j,t are as follows:

[0033]

[0034]

[0035] where l represents the sliding window length. For the CSI amplitude corresponding to a certain time period, when the variances of k consecutive sliding windows are all greater than a certain threshold s, it is considered that there is a human behavior within this time period, and k is defined as the cache band size. In the WI-FHOP system, after multiple actual verifications, when l = 80, k = 100, s = 0.4, the interception effect is the best. The intercepted behavior data contains amplitude data on multiple channels, and the data can be divided through the channel labels on the CSI data packet. The data on each channel after division respectively represent the amplitude jitter characteristics of the same behavior in each frequency band.

[0036] Step 3: The CSI signal contains information of 51 subcarriers. For the behavior recognition system, much of this information is redundant. Therefore, this paper first uses the PCA (Principal Component Analysis) algorithm to reduce the dimension of the CSI amplitude data. After performing eigenvalue decomposition on the covariance matrix and arranging the eigenvalues from largest to smallest, we get [λ1, λ2,..., λ 51 . This paper selects 3 principal components corresponding to [λ2, λ3, λ4]. This is because the first principal component of PCA contains a large amount of burst noise caused by factors of the device itself, and [λ2, λ3, λ4] already satisfies λ2 + λ3 + λ4 ≥ 0.99. Through PCA, the original CSI amplitude matrix of 51×N is reduced to 3×N, where N represents the number of CSI packets. Immediately afterwards, this paper performs Discrete Wavelet Transformation (DWT) on each PCA principal component to extract time-frequency characteristics. The present invention sets the number of wavelet layers to 9. Through DWT, the matrix dimension changes from 3×N to [3×(9 + 1)]×N. Define H W as the feature matrix after wavelet transformation, where the j-th row H W of H W,j represents the time-frequency components of a certain layer of wavelet of a certain principal component. For the time-frequency components obtained through DWT, each component represents a different speed interval. By the magnitudes of each frequency component, it can help analyze the speed and intensity of different behaviors. Although the time-frequency characteristics can well reflect the speed and intensity of behavioral actions, since the speed or rhythm of each behavior is inconsistent, resulting in inconsistent time lengths, that is, the length N of the feature sequence is inconsistent, which causes most pattern recognition methods to be inapplicable. Therefore, the present invention makes the feature sequence lengths consistent by extracting the statistical information of H W . And 6 different statistical features of the mean, standard deviation, interquartile range, 0.5 percentile, 0.683 percentile, and 0.95 percentile of the probability distribution are selected;

[0037] ① Mean: The mean is an index used in statistics to represent the center point of data. It can characterize the central tendency of the time-frequency components at different moments of the same behavior. Its calculation formula (5) is as follows:

[0038]

[0039] ② Standard deviation: The standard deviation is an index used in statistics to represent the degree of dispersion of data. It can characterize the fluctuation of the time-frequency components at different moments of the same behavior. Its calculation formula (6) is as follows:

[0040]

[0041] ③ The interquartile range is a robust statistic that characterizes the dispersion of data. Its calculation formula (7) is as follows:

[0042] IQR(H W,j )=Q3(H W,j )-Q1(H W,j )(7)

[0043] Among them, Q3(H W,j ) and Q1(H W,j ) represent H W,j The 3 / 4 quantile and 1 / 4 quantile of .

[0044] 3d. The 0.5 lower quantile, 0.683 lower quantile, and 0.95 lower quantile of the probability distribution: These three parameters are based on H W,j The cumulative probability distribution of They are three parameters commonly used in probability statistics, and their calculation formula (8) is as follows:

[0045]

[0046] There are 6 statistical features in total. Therefore, after statistical feature extraction, the amplitude jitter characteristics of any behavior on each channel can be described by a vector of length 6×3×(9+1). The feature data is processed using the z-score normalization method, and all features are converted to the same scale.

[0047] Step 4. This paper proposes a channel selection algorithm based on feature weights. Taking SVM as the benchmark classification algorithm and the binary classification problem as an example, the channel selection algorithm is introduced in detail below.

[0048] Define two different categories of behavior labels as -1 and 1. c,v Represents the 1×30 feature matrix corresponding to the vth statistical feature of a test behavior containing an unknown category on channel c, where c=1,6,11 represents different channels and v=1,2,3,4,5,6 corresponds to the six statistical features mentioned in the previous section. c,v Represents the m×30 feature matrix corresponding to the vth statistical feature of m training behaviors of two categories on channel c.

[0049] The algorithm consists of the following four steps:

[0050] 4a: By Ftr c,v Get the weight W of each feature v .

[0051] 4b: Ftr c,v As a training feature database, Fts c,vThe classification result RS is obtained by using the SVM classification algorithm c,v and P c,v , where RS c,v is the classification result label and P c,v is the probability corresponding to this classification result. Repeat this step until the classification results of all statistical features on all channels are obtained.

[0052] 4c: Calculate the total weight Q of each channel according to the formula c . Q c When Q < 0, it means that under channel c, the preliminary judgment of the label of this behavior is -1. On the contrary, when Q c ≥ 0, it is judged as 1.

[0053] 4d: Based on the size distribution of the Q1, Q6, Q 11 values, make the final decision on this behavior.

[0054] 4a can be understood as the work in the offline stage, finding the weight W of the first type of feature v , taking each row of Ftr c,1 as the test data and the remaining m - 1 as the training data, and also using SVM for classification and recognition. Repeat this step until the classification results of this statistical feature on all channels are obtained. During this period, record the number num1 of test behaviors with correct classification of this feature, then W1 = num1 / (m×3), and W2, W3,..., W6 can be obtained in the same way.

[0055] 4b and the subsequent steps are the work in the online stage. In 4b, a total of 6×3 groups of RS c,v and P c,v can be obtained. Based on this, in 4c, multiply the corresponding RS c,v , P c,v and W v to get q c,v , that is, q c,v = RS c,v ×P c,v ×W v . Based on the positivity or negativity of RS c,v , q c,v may be positive or negative. The total weight Q of each channel c is obtained by the following formula (9):

[0056]

[0057] Q c When Q < 0, it means that after weighting under channel c, the q c,v corresponding to the label of -1 in the recognition result dominates, that is, the behavior label of this test behavior under channel c is more likely to be -1. On the contrary, when Q cWhen it is ≥0, it indicates a higher likelihood of being 1.

[0058] In 4d, first for Q1, Q6, Q 11 Take the absolute value respectively to get |Q1|, |Q6|, |Q 11 |, which this paper calls the confidence of each channel, and 0 ≤ |Q1|, |Q6|, |Q 11 | ≤ 6. Find the one with the largest absolute value. The larger the absolute value, the greater the probability that it is considered to be a certain behavior label under this channel. Here, assume |Q1| is the largest, and its corresponding channel number is used as the preliminary optimal channel c op = 1. Then count the number of channels in Q1, Q6, Q 11 that have the same positive or negative nature as Q1 and denote it as c sm , so the value range of c sm is [1, 2, 3]. Then calculate the difference between |Q1| and the other two absolute weights respectively, and denote it as When c sm ≥2, maintain the original judgment of c op . When c sm = 1, set a threshold Q min related to the maximum total weight and a threshold related to the difference of absolute weights. If |Q1| < Q min , and as long as one of them satisfies , then modify c op to any other channel, because the recognition results of the other two channels are the same. The above situation means that on the channel with the largest total weight, it doesn't have a high certainty about which category this behavior belongs to, but there are channels on which the confidence is close to that of this channel, and the recognition results of the other two channels are opposite to this channel. So at this time, if we still adhere to the principle of the largest confidence, the misjudgment rate will be relatively high. In other sub - cases when c sm = 1, still maintain the original judgment of c op . Finally, according to Q op corresponding to c c , if Q c <0, the final output recognition result label is - 1, otherwise output 1.

[0059] The above content completely describes the specific process of the channel selection algorithm when using SVM as the benchmark classification algorithm. Similarly, when using KNN or Bagging as the benchmark classification algorithm, the channel selection algorithm can also be implemented, because for any recognition result, they can both obtain RS c,v and P c,v . In addition, in the WI - FHOP system, according to the actual performance, set Q min= 2.8, The average recognition accuracy rate is the highest at this time.

[0060] Application in the wharf scenario:

[0061] In the tally operation of the general cargo wharf, the movement and stacking of large timber are involved, which are likely to cause safety accidents and personal injuries, and there are certain potential safety hazards (for example, when a timber forklift / transport vehicle unloads timber, there are other staff members in the operation area but not noticed). Therefore, safety management of the tally operation is required. Through the implementation of this technology, accidents occurring during the tally process of the timber wharf can be prevented.

[0062] Traditional safety protection measures generally rely on video surveillance for protection. However, due to the large amount of wood stored in the yard, there are large visual blind spots, and at the same time, wharf operations often occur at night. Affected by low light, video surveillance cannot provide effective early warnings.

[0063] Therefore, deploying a behavior recognition system based on WiFi cross-band communication in the yard can not only identify whether there are people, but also recognize human behaviors such as squatting and sitting, which further improves the protection measures. The schematic diagram of deploying this system in the wharf yard is as Figure 3 shown.

[0064] The early warning process is as Figure 4 shown, and the steps are as follows:

[0065] 1. WIFI devices deployed near the timber yard collect CSI data;

[0066] 2. Each WIFI device transmits the collected data back to the central server;

[0067] 3. The central server processes the data and conducts behavior recognition of cross-band communication according to the method of this system;

[0068] 4. Since the CSI data transmitted back by each WIFI device is marked with an IP or MAC address or has some kind of association, the recognized results can be corresponding to the WIFI devices;

[0069] 5. The installation displacement of the WIFI device can also be corresponding to the IP or MAC of this WIFI, that is, it can be known that the result of recognizing someone is in the coverage area of that WIFI device.

[0070] 6. Based on the result of step 5, early warnings are issued to avoid the occurrence of safety accidents.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and all of them should be covered by the scope of the claims of the present invention.

Claims

1. A port personnel behavior recognition system based on WiFi cross-band communication, characterized by: The system includes: A data acquisition module is used to collect channel state information (CSI) data from multiple different channels and send the collected CSI data to the data processing module; The data processing module is used to filter the received CSI data, intercept the behavior data and divide the channels, and send the processed behavior data blocks to the feature analysis module; A feature analysis module is used to perform principal component analysis, wavelet transform and statistical feature extraction on the received behavior data block, and send the extracted feature vector to the channel selection module; The channel selection module is used to select a channel according to the received feature vector and output the final recognition result.

2. According to claim 1, the port personnel behavior recognition system based on WiFi cross-band communication is characterized in that: The CSI data collected by the data collection module comes from channel 1, channel 6 and channel 11 in the 2.4 GHz WiFi frequency band.

3. According to claim 1, the port personnel behavior recognition system based on WiFi cross-band communication is characterized in that: The data processing module comprises: The filtering unit is used to filter out high-frequency noise in the original CSI using a Butterworth low-pass filter, as shown in the following formula: Among them, Δm j,t represents the sliding window mean of the amplitude of the jth subcarrier in the tth sliding window; H(f j ,t) represents the amplitude of the jth subcarrier at time t, f j represents the frequency of the jth subcarrier, f j ,t is the parameter of H() function; l represents the sliding window length. For the CSI amplitude corresponding to a certain time period, when the variance of d consecutive sliding windows is greater than a certain threshold s, it is considered that there is human behavior in this time period; d consecutive sliding windows are defined as the buffer band; j represents the index number of the subcarrier, j=1,2,...,51; The behavior data interception unit is used to extract the behavior data using the sliding window technology, and calculate the sliding window variance according to the following formula to determine whether there is human behavior; Among them, x j,t represents variance; Represents a channel division unit, which is used to divide the data into multiple channels according to the channel label on the CSI data packet; the intercepted behavior data contains amplitude data on multiple channels, and the data is divided according to the channel label on the CSI data packet. The data on each divided channel represents the amplitude jitter characteristics of the same behavior in each frequency band.

4. According to claim 1, the port personnel behavior recognition system based on WiFi cross-band communication is characterized in that: The feature analysis module comprises: The principal component analysis unit is used to reduce the dimension of the CSI amplitude data, as shown in the following formula: Among them, H(f j ,t) represents the amplitude of the jth subcarrier at time t, arg(H(f j ,t)) represents the phase of the jth subcarrier at time t, j = 1, 2, ..., 51, f j represents the frequency of the jth subcarrier, e -j2πΔft represents the phase offset caused by inconsistent carrier frequency, K represents the number of multipaths, a k (f j ,t) is a complex number, which represents the propagation attenuation and initial phase offset of the kth multipath. It means that due to the propagation delay τ k (t) caused by the phase shift, n(f j ,t) represents noise; when the dynamic target in the link has a behavior action, a multipath in the environment, set as k°, propagates and changes, and the corresponding a k° (f j ,t) and τ k° (t) changes, resulting in a change in the CSI amplitude of the received signal, and by intercepting the CSI data block corresponding to human activity and analyzing the characteristics of the CSI data block, different behavioral activities are distinguished based on different characteristics; A wavelet transform unit is used to perform wavelet transform on each principal component of the principal component analysis PCA to extract time-frequency features; A statistical feature extraction unit is used to extract statistical features from the feature matrix after wavelet transform DWT, including mean, standard deviation, interquartile difference, 0.5 percentile, 0.683 percentile and 0.95 percentile; The mean is an indicator used statistically to represent the center point of the data, representing the central trend of the time-frequency components of the same behavior at different times. The calculation formula is: Among them, mean(H W,j ) represents the mean, N represents the number of CSI packets, and n represents the number of CSI packets; H W,j Represents the time-frequency component of a certain layer of wavelet of a certain principal component; The standard deviation is a statistical indicator used to indicate the degree of dispersion of data, which characterizes the fluctuation of the time-frequency components of the same behavior at different times. The calculation formula is: Among them, std(H W,j ) represents the standard deviation; The interquartile range is a robust statistic that characterizes the dispersion of data. The calculation formula is: IQR(H W,j )=Q3(H W,j )-Q1(H W,j ) Among them, IQR (H W,j ) represents the interquartile range; Q3(H W,j ) and Q1(H W,j ) represent H W,j The 3 / 4 quantile and 1 / 4 quantile of; The 0.5 lower quantile, 0.683 lower quantile, and 0.95 lower quantile of the probability distribution; these three parameters are based on H W,j The cumulative probability distribution of What we get are three parameters commonly used in probability statistics, and the calculation formula is: Among them, α 0.5 represents the 0.5 lower quantile of the probability distribution, α 0.683 represents the 0.683 lower quantile, α 0.95 Represents the 0.95 lower quantile.

5. According to claim 1, the port personnel behavior recognition system based on WiFi cross-band communication is characterized in that: The channel selection module adopts a channel selection algorithm based on feature weights, which combines the KNN classification algorithm, the SVM classification algorithm and the Bagging classification algorithm to select the optimal channel according to the feature data of multiple channels, and uses the recognition result of the channel as the final recognition result.

6. The port personnel behavior recognition system based on WiFi cross-band communication according to claim 5 is characterized by: The channel selection algorithm based on feature weights comprises the following steps: Offline stage: Calculate the weight of each feature as shown below: Among them, Q c represents the total weight of each channel; Q c <0, indicating that after weighting under channel c, the label of the recognition result is -1 corresponding to q c,v dominates, that is, the behavior label of the test behavior under channel c is more likely to be -1, Q c ≥0, indicating that it is more likely to be 1; v=1,2,3,4,5,6; q c,v represents the calculation weight of the vth feature of the cth channel; Online stage: The total weight of each channel is calculated according to the weight of each feature, and the optimal channel is selected according to the size of the total weight.

7. The port personnel behavior recognition system based on WiFi cross-band communication according to claim 6 is characterized by: The offline phase includes the following steps: For each feature, the feature matrix containing one test behavior of the unknown category on the channel is used as the test data, and the rest is used as the training data. SVM is used for classification and recognition, and the number of correctly classified test behaviors is recorded to obtain the weight of this feature.

8. The port personnel behavior recognition system based on WiFi cross-band communication according to claim 6 is characterized by: The online phase includes the following steps: For each channel, the total weight is calculated based on the weight of each feature and the classification result; The optimal channel is selected according to the total weight, and the recognition result on this channel is used as the final recognition result.

9. A method for identifying port personnel behavior using the system according to any one of claims 1 to 8, characterized in that: The following steps are involved: Collect CSI data from multiple different channels; Filter the collected CSI data, extract behavioral data, and divide channels; Perform principal component analysis, wavelet transform and statistical feature extraction on the data after channel division; The optimal channel is selected according to the feature weight, and the recognition result on this channel is used as the final recognition result.

Images