Method for detecting and monitoring human intrusion based on CSI (Channel State Information) signal

Abstract

The invention discloses a detecting and monitoring method for human-shaped intrusion based on a CSI signal. The method comprises the following steps: S1, acquiring CSI signal training data; s2, inputting the CSI signal training data into a depth estimation model for training, discretizing a depth interval and setting a scaling scale of each interval as a learnable parameter, and in the training process, enabling the depth estimation model to adaptively optimize a distance interval through a dynamic width parameter, and mapping the CSI signal training data to a specific distance range; and S3, inputting CSI signal application data to the trained depth estimation model. According to the CSI signal human-shaped intrusion detection and monitoring method, the distance change between the human body and the signal end in the moving process can be detected in real time, the intrusion distance of the human body can be judged on the basis of the characteristics of the product, and after the human body intrudes to a certain distance, alarm information prompt is triggered, so that the intrusion detection and monitoring effect is achieved.

Description

Technical Field

[0001] The invention relates to a method for detecting and monitoring humanoid intrusion based on CSI signals. Background Art

[0002] In modern life, we usually use smart surveillance cameras to detect and monitor human intrusion. Such equipment often needs to be combined with machine vision systems, including lens modules, monitoring housings, pan / tilt and other electronic and mechanical structures, which have a certain cost. Therefore, for such scenarios, we proposed an intrusion detection system based on CSI signals. The system can judge the intrusion distance of a person based on the difference in CSI signals received by the wifi chip, and trigger an alarm message when the person intrudes to a certain distance. This type of product does not need to rely on cameras and image analysis algorithms, but can be realized by relying on WIFI chips that obtain CSI. WIFI chips are already widely used, so they can greatly reduce the cost of hardware and software. Summary of the invention

[0003] The purpose of the present invention is to provide a method for detecting and monitoring human intrusion based on CSI signals to solve the problems raised in the above background technology. To achieve the above purpose, the present invention provides the following technical solutions:

[0004] A method for detecting and monitoring human intrusion based on CSI signals comprises the following steps:

[0005] S1: Obtain CSI signal training data;

[0006] S2: Input the CSI signal training data into the depth estimation model for training, wherein the depth interval is discretized and the scaling scale of each interval is set as a learnable parameter. During the training process, the depth estimation model is allowed to adaptively optimize the distance interval through a dynamic width parameter to map the CSI signal training data to a specific distance range;

[0007] S3: Input CSI signal application data to the trained depth estimation model.

[0008] Furthermore, the CSI signal training data should ensure that 40-50 groups of CSI subcarrier information are received per second, each group of subcarrier information should contain at least 60 subcarrier signals, and each subcarrier signal includes a real part and an imaginary part.

[0009] Furthermore, the depth estimation model optimization distance interval algorithm includes the following steps:

[0010] S21: define an initial distance interval range D = [dmin, dmax], where dmin and dmax are the minimum and maximum values ​​of the distance respectively;

[0011] S22: Introduce a learnable parameter vector b, where each element bi of b represents the width of the i-th interval; the width is adaptive and will be automatically adjusted according to the training data;

[0012] S23: For each interval, calculate its center c(bi), which is the potential distance value predicted by the model, where the calculation formula for the interval center is:

[0013]

[0014] Where N is the total number of intervals;

[0015] S24: According to the interval center and parameter vector b, during the training process, whenever reasoning is completed, the label is mapped to each interval according to b returned by the network, and then a softmax layer is used to predict whether the bin to which the label falls is consistent with the maximum probability given by softmax, and the network weight is adjusted by reverse derivation.

[0016] Beneficial effects:

[0017] The system can determine the intrusion distance of a person based on the difference in CSI signals received by the wifi chip, and trigger an alarm message when the person intrudes to a certain distance. This type of product does not need to rely on cameras and image analysis algorithms, but can be realized by relying on WIFI chips that obtain CSI. WIFI chips are already widely used, so they can greatly reduce the cost of hardware and software. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Yes: CSI neural network structure diagram including dynamic width correction layer;

[0019] Figure 2 Yes: Dynamic width diagram. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] The present invention provides a method for detecting and monitoring human intrusion based on CSI signals, comprising the following steps:

[0022] S1: Obtain CSI signal training data;

[0023] The CSI signal training data should ensure that 40-50 groups of CSI subcarrier information are received per second (40 for example). Each group of subcarrier information should contain at least 60 subcarrier signals. Each subcarrier signal contains a real part and an imaginary part. Therefore, the data information contained in the CSI per second should be 40*60*2.

[0024] S2: Input the CSI signal training data into the depth estimation model for training, wherein the depth interval is discretized and the scaling scale of each interval is set as a learnable parameter. During the training process, the depth estimation model is allowed to adaptively optimize the distance interval through a dynamic width parameter to map the CSI signal training data to a specific distance range;

[0025] In conventional AI algorithm technology, the prediction distance is generally predicted by regression, that is, a set of features (or a picture) is input, and a regression loss function is used to train the model to regress a distance value. However, the distance value generated by the regression method is not strictly constrained, and the accuracy of this method is usually not as good as classification; this leads to another prediction method, that is, to divide the distance into n groups evenly, such as 1-1.2, 1.2-1.4, ..., 2.8-3.0, 3+, and then let the model predict the probability of each category, and take the category with the highest probability for output. This method may be able to adapt to a certain type of problem, but it cannot adapt to the characteristics of CSI signals: the sensitivity of CSI signal characteristics to changes in distance is not uniform. For example, when a person stands at a distance of 1.8-2m, CSI may be able to produce a relatively stable signal, but the sensitivity of the signal within 1-1.2m may decrease, resulting in the signal being unable to effectively distinguish the difference between the signal within 1-1.2m and the signal within 1.2-1.4m. Therefore, at this time, it is more appropriate to set the distance interval to 1-1.4m. Due to the differences in the specifications of CSI signal transmitters and receivers, we cannot preset a set of distance intervals to ensure the accuracy of the distance monitoring model. Therefore, we proposed an adaptive distance interval to discretize the depth interval and set the scaling of each interval as a learnable parameter. During the training process, the depth estimation model can adaptively optimize the distance interval (through a dynamic width parameter) to ensure that the CSI signal features are accurately mapped to a specific distance range.

[0026] The depth estimation model optimization distance interval algorithm includes the following steps:

[0027] S21: define an initial distance interval range D = [dmin, dmax], where dmin and dmax are the minimum and maximum values ​​of the distance respectively;

[0028] S22: Introduce a learnable parameter vector b, where each element bi of b represents the width of the i-th interval; the width is adaptive and will be automatically adjusted according to the training data;

[0029] S23: For each interval, calculate its center c(bi), which is the potential distance value predicted by the model, where the calculation formula for the interval center is:

[0030]

[0031] Where N is the total number of intervals;

[0032] S24: According to the interval center and parameter vector b, during the training process, whenever reasoning is completed, the label is mapped to each interval according to b returned by the network, and then a softmax layer is used to predict whether the bin to which the label falls is consistent with the maximum probability given by softmax, and the network weight is adjusted by reverse derivation.

[0033] Based on the above algorithm, in the model, we need to introduce an adaptive width output layer after the feature layer to perform width regression, and use the label layer to integrate the true label. Therefore, the feature layer will directly flow into two directions, namely the adaptive width output layer on the left and the softmax output below, see Figure 1 .

[0034] In order to train the model, we need to define a complete loss function. This loss function should consider both the accuracy of dynamic width prediction and the rationality of predicted distance. Therefore, we use the following form of loss function: L = Ld + βLbins where Ld is the softmax cross loss function, which is used to deal with the difference loss between the true label and the softmax predicted label. Lbins is the dynamic width loss, which is calculated as Lbins = (ci-d) 2 , where ci is the center point calculated by the above step S23:, the one whose center point is closer to the predicted distance has a smaller judgment loss. However, in order to ensure the rationality of the prediction, we do not let Lbins return to the true position, so we add δ as a margin. If (ci-d) 2 <δ, then Lbins is 0, that is, the loss is considered reasonable. In addition, ci means the center point ci that is closest to the true distance, and no optimization is required when other cj≠i. This setting not only avoids the problem of the model infinitely optimizing the width distance to the center point during training, but also limits the problem of the model infinitely increasing the width of the area when encountering insensitive areas.

[0035] S3: Input CSI signal application data to the trained depth estimation model

[0036] After using the above model to calculate, we can accurately distinguish the distance between the person and the CSI receiver. Figure 2In actual use, we can let users select the interval range that they need to trigger an alert. During the model reasoning process, the model will predict in real time which interval of the above interval the person is in when reaching the CSI receiver. Once the person reaches a certain interval, an alert will be issued. This is easy to implement in software.

[0037] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. The protection scope of the present invention shall be subject to the protection scope of the claims. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A method for detecting and monitoring human intrusion based on CSI signals, characterized in that: The following steps are involved: S1: Obtain CSI signal training data; S2: Input the CSI signal training data into the depth estimation model for training, wherein the depth interval is discretized and the scaling scale of each interval is set as a learnable parameter. During the training process, the depth estimation model is allowed to adaptively optimize the distance interval through a dynamic width parameter to map the CSI signal training data to a specific distance range; S3: Input CSI signal application data to the trained depth estimation model.

2. The method for detecting and monitoring human intrusion according to claim 1, characterized in that: The CSI signal training data should ensure that 40-50 groups of CSI subcarrier information are received per second. Each group of subcarrier information should contain at least 60 subcarrier signals, and each subcarrier signal contains both real and imaginary parts.

3. The exercise monitoring method according to claim 1, characterized in that: The depth estimation model optimization distance interval algorithm includes the following steps: S21: define an initial distance interval range D = [dmin, dmax], where dmin and dmax are the minimum and maximum values ​​of the distance respectively; S22: Introduce a learnable parameter vector b, where each element bi of b represents the width of the i-th interval; the width is adaptive and will be automatically adjusted according to the training data; S23: For each interval, calculate its center c(bi), which is the potential distance value predicted by the model, where the calculation formula for the interval center is: Where N is the total number of intervals; S24: According to the interval center and parameter vector b, during the training process, whenever reasoning is completed, the label is mapped to each interval according to b returned by the network, and then a softmax layer is used to predict whether the bin to which the label falls is consistent with the maximum probability given by softmax, and the network weight is adjusted by reverse derivation.

Images