Action recognition method and system based on CSI (Channel State Information) and development board

Abstract

The invention provides an action recognition method and system based on CSI (Channel State Information) and a development board. The action recognition method comprises the following steps: receiving real CSI data acquired by Wi-Fi equipment in real time; generating virtual data by using a generator of the trained semi-supervised generative adversarial network; real CSI data collected in real time and the generated virtual data are preprocessed, and preprocessed CSI data are obtained; adopting a local anomaly factor LOF algorithm to judge that an abnormal change position of the preprocessed CSI data is in action, and intercepting a preset number of data packets before and after taking the abnormal change position of the data as a center to serve as action signal segments; the method comprises the following steps: after acquiring an action signal segment, performing feature extraction on the action signal segment by using a fisher score algorithm, and selecting a preset number of features with the fisher score in the front; and inputting the selected features into a CNN convolutional neural network to classify the action features to obtain action categories. According to the invention, the hardware problem of CSI acquisition, the problem of poor adaptive capacity of action identification and the problem of selecting action features in the identification process in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of motion recognition, and in particular to a CSI-based motion recognition method, system and development board. Background Art

[0002] With the continuous development of synaesthesia integration, Wi-Fi wireless sensing technology has been widely used in the detection of daily activities of people. Compared with other activity detection solutions such as wearable devices and cameras, it has the advantages of no equipment, no privacy infringement, low cost, easy deployment, etc., and can be used on a large scale in real life.

[0003] Existing Wi-Fi personnel activity detection is mainly used in motion recognition, indoor positioning, etc. The main implementation solutions are as follows.

[0004] The wireless router is used as the transmitter, and a computer equipped with an Intel 5300 wireless network card is used as the receiver of the Wi-Fi signal to collect CSI data (Channel State Information). The collected CSI data is then denoised and reduced in dimension through PCA (principal component analysis). The CSI fingerprint database is trained, and machine learning or neural networks are used for motion recognition and indoor positioning.

[0005] However, the above solutions have the following shortcomings:

[0006] First, data collection requires a computer equipped with an Intel 5300 wireless network card and the Linux 802.11n CSI toolkit, which has stringent requirements on both hardware and software, is difficult to operate, and is highly dependent on external devices, making it difficult to apply in actual daily life;

[0007] Second, current motion recognition is based on manually divided motion data segments, which are identified through machine learning. It is unable to independently judge and intercept motion segments, or cannot accurately capture motion features, and cannot be applied to actual scenarios of real-time detection;

[0008] Third, the above experimental scheme is suitable for specific users in specific environments. For new users with different behavioral habits and new environments with different layouts and displays, the recognition accuracy will drop significantly.

[0009] The root cause of the above problems is that there are few ways to capture CSI data in Wi-Fi signals. The current mainstream is to extract CSI data through the Intel 5300 network card. Using a network card to collect CSI means that a computer is needed for processing, which makes it difficult to truly apply wireless sensing technology to actual scenarios. Secondly, real-time edge detection requires the firmware to have a certain ability to autonomously identify action features, accurately extract valid data, and improve recognition efficiency. Finally, due to the lack of labeled data for new scenarios and new users, the model recognition ability will decrease with only a small amount or even almost no data. Summary of the invention

[0010] In order to solve the above technical problems, the present invention utilizes the advantages of the development board such as low power consumption and stable and high-speed wireless data throughput to truly realize edge detection; a semi-supervised generative adversarial network is used for data generation, and a supplementary generator is added to the original basic generator to generate sufficient labeled data from the source domain to the target domain, and generate virtual data containing each type of action label; an action feature selection scheme combining the local outlier factor LOF algorithm with the fisher score algorithm is adopted to reduce the amount of data that needs to be analyzed for machine learning, and effectively improve the scientificity and accuracy of action feature selection, and improve the accuracy and efficiency of action recognition.

[0011] The present invention adopts the following technical solution:

[0012] On the one hand, a CSI-based motion recognition method is applied on a development board, including:

[0013] Receive real CSI data collected by Wi-Fi devices in real time;

[0014] Input the pre-collected real data with action category labels into the generator of the trained semi-supervised generative adversarial network to generate virtual data containing each type of action label;

[0015] Preprocess the real CSI data collected in real time and the generated virtual data to achieve denoising and dimension reduction, and obtain preprocessed CSI data;

[0016] The local anomaly factor LOF algorithm is used to determine the abnormal changes in the pre-processed CSI data as the occurrence of an action; the preset data packets before and after the abnormal data change are intercepted as action signal segments;

[0017] After obtaining the action signal segment, the Fisher score algorithm is used to extract features of the action signal segment, and the preset features with the highest Fisher score are selected;

[0018] The selected features are input into the CNN convolutional neural network to identify and classify the action features and obtain the action category.

[0019] Preferably, the semi-supervised generative adversarial network includes a generator and a discriminator; the generator includes a basic generator and a supplementary generator;

[0020] The basic generator generates virtual data close to real data based on the random noise vector;

[0021] The supplementary generator generates virtual data with all action category labels from real data with action category labels;

[0022] The discriminator receives input virtual data close to real data, virtual data with all action category labels, real data with action category labels and unlabeled real data, and generates a virtual data set with labels after iterative training.

[0023] Preferably, the real CSI data collected in real time and the generated virtual data are preprocessed, including: hampel filtering, moving average filtering and PCA principal component analysis.

[0024] Preferably, a local outlier factor LOF algorithm is used to determine the abnormal changes in the preprocessed CSI data as actions that occur, specifically including:

[0025] The distance between point p and point q in the CSI data is defined as d(p,q), and the kth distance d of data point p is k (p) is defined as d k (p) = d(p, q), the kth distance of point p and the set of all points contained in the kth distance are the kth neighborhood N of point p k (p); point q is a point in the kth neighborhood of point p, d k (q) is the k-neighbor distance of point q; calculate the k-th reachable distance rd from point p to point q k (p,q) takes point q as the object, compares the kth distance of q with the actual distance between q and p, and then selects the larger distance as the reachable distance from point p to point q. The calculation formula is as follows:

[0026] rd k (p,q)=max{d k (q),d(p,q)}

[0027] Calculate the local reachability density lrd k (p), as follows:

[0028]

[0029] Calculate the local outlier factor LOF of data point pk (p), as follows:

[0030]

[0031] Local Outlier Factor LOF k (p) value to judge, if LOF k If the deviation between the value of (p) and 1 is within the preset numerical range, p is judged to be data in a static environment. Otherwise, p is judged to be abnormal data, indicating that an action change has occurred at this moment, which is an abnormal change.

[0032] Preferably, the Fisher score algorithm is used to extract features from the action signal segment, and a preset number of features with high Fisher scores are selected, specifically including:

[0033] Given a feature set {f1, f2, ..., f i ,…,f n}, for any CSI data sample x j ∈R m , calculate the feature f i The inter-class divergence S B (f i ) and the lth class sample in feature f i The intra-class divergence under as follows:

[0034]

[0035] Among them, i = 1, 2, ..., n, is the feature subscript; j = 1, 2, ..., m, is the sample category subscript; c represents the number of sample categories; d l is the number of samples in the lth class; Indicates that the l-th category sample has feature f i The mean value on i Indicates that all samples have feature f i The mean on ; Indicates that the jth sample in the lth class sample is in f i The value of the following;

[0036] Define feature f i The Fisher score is J fisher (f i ), according to the feature f i The inter-class divergence and intra-class divergence of J fisher (f i ) is expressed as:

[0037]

[0038] After calculating the Fisher score of each feature, the most discriminative feature is selected according to the value of the Fisher score. Specifically, the Fisher scores are sorted from high to low, and preset features are selected in turn for action recognition.

[0039] On the other hand, a development board includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute any one of the CSI-based action recognition methods.

[0040] On the other hand, a CSI-based motion recognition system includes a development board and a Wi-Fi device; the Wi-Fi device collects CSI data and sends the collected CSI data to the development board; the development board includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute any one of the CSI-based motion recognition methods.

[0041] Preferably, the Wi-Fi device includes a router.

[0042] The present invention has the following beneficial effects:

[0043] (1) At present, the research on CSI-based wireless sensing mainly uses network cards and computers, ignoring the practicality of the wireless sensing function. In practical applications, devices are often required to perform edge detection instead of over-relying on computers. The present invention develops the MT7931 development board, that is, a small development board can be used to realize the wireless sensing function, which reduces the cost and can be deployed on a large scale, meeting the practicality;

[0044] (2) Existing research often cannot cope with changes in the environment and changes in the identification object. For example, actions such as falling do not often occur in daily life, so it is difficult to collect all action category data in a new environment. The present invention uses a semi-supervised generative adversarial network and adds a supplementary generator to generate virtual data including all action categories, which improves generalization and is conducive to the practical application of CSI action recognition technology.

[0045] (3) The present invention fully considers the problem that in practical applications, it is necessary to autonomously intercept the fragments of the signal containing action information and accurately extract the action features. The local outlier factor LOF algorithm is used to first intercept the signal fragments where the action occurs, thereby reducing the amount of data and the influence of noise in subsequent processing. The Fisher score algorithm is then used to obtain the action signal features with the most classification value. Subsequent action classification is performed based on these features, which further reduces the dimension of the data and improves the efficiency of subsequent machine learning classification. At the same time, it also reduces the interference of irrelevant information and improves the accuracy of recognition.

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a CSI-based action recognition method according to an embodiment of the present invention;

[0048] Figure 2 A structural block diagram of a board-supervised generative adversarial network according to an embodiment of the present invention;

[0049] Figure 3 It is an algorithm flow chart of the local outlier factor LOF according to an embodiment of the present invention;

[0050] Figure 4 4 is a structural block diagram of a CSI-based action recognition system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments, and all other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work are within the scope of protection of the present invention.

[0052] In the description of the present invention, it should be noted that the terms "comprises", "includes" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0053] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, step identifiers S101, S102, S103, etc. are only used for convenient expression and do not represent the execution order. The corresponding execution order can be adjusted as needed.

[0054] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0055] See also Figure 1 As shown, the present invention discloses a CSI-based action recognition method, which is applied on a development board and includes the following steps.

[0056] S101, receiving real CSI data collected in real time by a Wi-Fi device.

[0057] In this embodiment, the development board may be an MT7931 development board, and the Wi-Fi device includes a router.

[0058] Specifically, a motion recognition system based on channel state information (CSI) is built. The system consists of an MT7931 development board and a router. The MT7931 development board is used as a receiving device, and the router is used as a transmitting device to collect CSI data. You can choose to connect to 2.4G or 5G according to actual needs.

[0059] S102, inputting the pre-collected real data with action category labels into the generator of the trained semi-supervised generative adversarial network to generate virtual data containing each type of action label.

[0060] In specific implementation, the training process of the semi-supervised generative adversarial network can be carried out on the computer. After the CSI data is collected, it will be uploaded to the computer for automatic recording. After the collected data is labeled with action categories, it is input into the semi-supervised generative adversarial network for virtual data generation.

[0061] See also Figure 2 As shown, the semi-supervised generative adversarial network of the present invention has two generators: one is the basic generator G v , a virtual signal close to the real data is generated by a random noise signal; the second is to supplement the generator G c, virtual data with all action category labels are generated from real data with action category labels. In addition, there are real data with action category labels and a small amount of unlabeled real data from the target domain. These four parts of data are input into the discriminator D for discrimination, and a large number of virtual data sets with labels are generated after iterative training. Figure 2 In the figure, the upper right number of each parameter is the action category, such as y 1 refers to the label of the first type of action. ′ refers to the generated virtual label, y refers to x labeled The label of the lower right corner is c, which means that it is generated by the generator G c Generate dummy data such as Refers to the generator G c The first type of action data generated. Similarly, y ′1 Indicates the virtual first-category action label. Real c=1 indicates the first real-category action, and similarly fake c=1 ′ Represents the first virtual class action.

[0062] The training process of the semi-supervised generative adversarial network is described in detail as follows.

[0063] (1) Collect real data with action category labels from known scenes and input them into the supplementary generator G c , generate virtual data close to the unknown scene. Since the real data covers all action categories, virtual data containing all action category labels can be generated here. And the generated virtual data is input into the discriminator D.

[0064] (2) Input the real data with action category labels collected from known scenes into the discriminator D.

[0065] (3) Input a small amount of unlabeled real data in the unknown scene into the discriminator D.

[0066] (4) Use the randomly generated noise vector z as the basic generator G v , generates virtual data, and inputs the generated virtual data into the discriminator D.

[0067] (5) Input the above four parts of data into the discriminator. Assuming that there are k action labels to be classified, the discriminator outputs 2k+1 categories, of which the real data is placed in the first k categories (1~k), supplementing the generator G c The generated labeled data is placed into the next k categories (1 ’ ~k ’ ), basic generator G v The generated virtual samples are placed into 1 ’ to k+1 ’ The categories in .

[0068] Assume x l and x u denote the labeled and unlabeled CSI data samples respectively, x c and x v Respectively represent the complementary generator G c and the basic generator G v The generated labeled virtual data and basic virtual data. Then the objective function of the discriminator is expressed as:

[0069]

[0070]

[0071] Among them, γ1, γ2, and γ3 are parameters for adjusting the weights of unlabeled data, labeled virtual data, and basic virtual data, respectively. The first term of indicates that the labeled data in the training set are correctly placed in the corresponding categories in the top k categories. The second term is designed for unlabeled samples and tries to reduce the probability of samples being placed in virtual categories, while the third term tries to put labeled virtual samples into corresponding virtual categories, and the fourth term encourages the discriminator to distinguish between virtual samples and real data.

[0072] S103, preprocessing the real CSI data collected in real time and the generated virtual data to achieve denoising and dimensionality reduction, and obtain preprocessed CSI data.

[0073] Considering that there is a certain amount of noise influence in the process of sending and receiving data, it is necessary to perform data preprocessing operations on it. The main processing algorithms include hampel filtering, moving average filtering and PCA principal component analysis, which are used to denoise and reduce the dimension of the signal on the development board.

[0074] S104, using the local outlier factor LOF algorithm, determining the abnormal change of the pre-processed CSI data as the occurrence of an action; taking the abnormal change of the data as the center, intercepting a preset number of data packets before and after as action signal segments.

[0075] The preprocessed CSI data contains 56 subcarriers (a wireless communication channel is modulated into 64 subcarriers using OFDM technology, IEEE 802.11n uses 56 of them, 52 of which are used to transmit data signals and 4 are used to transmit pilot signals. Therefore, the 56 subcarriers are determined by the 802.11n protocol). In order to extract the segments involving action changes in the waveform, the local anomaly factor LOF algorithm is first used to determine the abnormal change of the waveform as an action (when activity occurs in a static space, the CSI amplitude will change significantly, and its data points will begin to shift, creating a certain distance from the data points in the static state. At this time, the local reachable density of the data points becomes smaller, and the LOF algorithm can separate the data points at these moments). A total of 450 data packets are intercepted before and after the abnormal point as the center as the action signal segment (according to the set sampling rate of 50 Hz, that is, 50 data packets are collected per second, and a total of 9 seconds of data before and after the action are intercepted with the abnormal point as the center to prevent the effective data generated during the action from being cut off).

[0076] For details, see Figure 3 As shown, the local outlier factor LOF algorithm is used to determine the abnormal changes in the preprocessed CSI data as the occurrence of an action, which specifically includes the following steps.

[0077] S1041, calculate the reachable distance.

[0078] The distance between point p and point q in the CSI data is defined as d(p,q), and the kth distance d of data point p is k (p) is defined as d k (p) = d(p, q), the kth distance of point p and the set of all points contained in the kth distance are the kth neighborhood N of point p k (p); point q is a point in the kth neighborhood of point p, d k (q) is the k-neighbor distance of point q; calculate the k-th reachable distance rd from point p to point q k (p,q) takes point q as the object, compares the kth distance of q with the actual distance between q and p, and then selects the larger distance as the reachable distance from point p to point q. The calculation formula is as follows:

[0079] rd k (p,q)=max{d k (q),d(p,q)}.

[0080] It should be noted that there are at least k points q′ in the set that do not include p, and the distance from point p is smaller than the distance from point q, that is, d(p,q ′ )≤d(p,q), and at most k-1 points, not including p, satisfy d(p,q ′) The k-th distance of <d(p,q)> is the distance to the k-th farthest data point from p. Then, the number of the k-th neighborhood points of point p satisfies |N k (p)| ≥ k.

[0081] S1042. Calculate the local reachability density.

[0082] The local reachability density of data point P is the reciprocal of the average reachability distance based on the nearest neighbors of P. The greater the distance, the

[0083] smaller the density. The calculation formula is as follows:

[0084]

[0085] S1043. Calculate the local outlier factor.

[0086] The LOF algorithm measures the degree of abnormality of a data point, not by its absolute local density, but by its relative density compared with the surrounding neighboring data points. The local relative density (local outlier factor) of data point p is the ratio of the average local reachability density of the points in the neighborhood of point p to the local reachability density of data point p. The calculation formula is as follows:

[0087]

[0088] By analyzing the LOF value, the degree of abnormality of the data point can be judged. The closer the local outlier factor LOF value of the data point is to 1, the more similar the local density of this point is to the local density of the data points in its neighborhood, indicating that this point is data in a static environment; on the contrary, if the LOF value is much greater than 1, it means that this data point deviates from other points more and may be abnormal data, that is, it indicates that an action change may occur at this moment. By setting a threshold, the LOF algorithm can be used to detect the part containing action information in the CSI amplitude data.

[0089] S105. After obtaining the action signal segment, use the fisher score algorithm to extract features from the action signal segment, and select a preset number of features with the top fisher scores. The specific steps are as follows.

[0090] (1) Given the feature set {f1, f2, …, f i , …, f n}, for any CSI data sample x j ∈ R m , calculate the between-class scatter S i (f B ) of feature f i and the within-class scatter of the l-th class samples (for the present invention, the types of samples are the collected real data and the virtual data generated by the generator) under feature f i ​ as follows:

[0091]

[0092] Among them, i = 1, 2, ..., n, is the feature subscript; j = 1, 2, ..., m, is the sample category subscript; c represents the number of sample categories; d l is the number of samples in the lth class; Indicates that the l-th category sample has feature f i The mean value on i Indicates that all samples have feature f i The mean on ; Indicates that the jth sample in the lth class sample is in f i The value below.

[0093] (2) Define feature f i The Fisher score is J fisher (f i ), according to the feature f i The inter-class divergence and intra-class divergence of J fisher (f i ) is expressed as:

[0094]

[0095] (3) After calculating the Fisher score of each feature, select the most discriminative feature subset based on the numerical value. The higher the numerical value, the stronger the ability of the feature subset to distinguish different categories. Sort the Fisher scores from high to low, and select multiple feature subsets for action recognition in turn, so as to improve the accuracy of action recognition.

[0096] S106, input the selected features into the CNN convolutional neural network to identify and classify the action features to obtain the action category. The specific processing process is as follows.

[0097] In summary, this embodiment is a CSI-based action recognition method. First, the MT7931 development board is used to connect to commercial Wi-Fi devices to collect CSI data, and the action recognition algorithm is directly run on the development board, which reduces the dependence on external devices and can realize Wi-Fi perception on an independent edge system. Secondly, a semi-supervised generative adversarial network is used to generate virtual data with labels in the target domain. This part of virtual data contains all the action types in the real data, thereby improving the environmental adaptability of the entire recognition system. Thirdly, the LOF algorithm and the fisher score algorithm are used to reduce the dimension and extract the action signal, realize the autonomous selection of the action signal, and improve the accuracy of the action signal capture. At the same time, the fisher score algorithm reduces the dimension of the data, reduces the power consumption of the firmware, and reduces the amount of calculation. Finally, the CNN convolutional neural network is used to classify the action features to realize action recognition in the firmware.

[0098] Furthermore, the development board of the present invention includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute any one of the CSI-based action recognition methods.

[0099] See also Figure 4 As shown, the present invention also discloses a CSI-based motion recognition system, including a development board and a Wi-Fi device; the Wi-Fi device collects CSI data and sends the collected CSI data to the development board; the development board includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute any one of the CSI-based motion recognition methods.

[0100] The above is only a preferred specific implementation of the present invention; however, the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solution and its improved conception within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A CSI-based motion recognition method, applied on a development board, characterized in that: include: Receive real CSI data collected by Wi-Fi devices in real time; Input the pre-collected real data with action category labels into the generator of the trained semi-supervised generative adversarial network to generate virtual data containing each type of action label; Preprocess the real CSI data collected in real time and the generated virtual data to achieve denoising and dimension reduction, and obtain preprocessed CSI data; The local anomaly factor LOF algorithm is used to determine the abnormal changes in the pre-processed CSI data as the occurrence of an action; the preset data packets before and after the abnormal data change are intercepted as action signal segments; After obtaining the action signal segment, the Fisher score algorithm is used to extract features of the action signal segment, and the preset features with the highest Fisher score are selected; The selected features are input into the CNN convolutional neural network to identify and classify the action features and obtain the action category.

2. The CSI-based action recognition method according to claim 1, characterized in that: The semi-supervised generative adversarial network includes a generator and a discriminator; the generator includes a basic generator and a supplementary generator; The basic generator generates virtual data close to real data based on the random noise vector; The supplementary generator generates virtual data with all action category labels from real data with action category labels; The discriminator receives input virtual data close to real data, virtual data with all action category labels, real data with action category labels and unlabeled real data, and generates a virtual data set with labels after iterative training.

3. The CSI-based action recognition method according to claim 1, characterized in that: The real CSI data collected in real time and the generated virtual data are preprocessed, including: hampel filtering, moving average filtering and PCA principal component analysis.

4. The CSI-based action recognition method according to claim 1, characterized in that: The local outlier factor (LOF) algorithm is used to determine the abnormal changes in the preprocessed CSI data as actions that have occurred, including: The distance between point p and point q in the CSI data is defined as d(p,q), and the kth distance d of data point p is k (p) is defined as d k (p) = d(p, q), the kth distance of point p and the set of all points contained in the kth distance are the kth neighborhood N of point p k (p); point q is a point in the kth neighborhood of point p, d k (q) is the k-neighbor distance of point q; calculate the k-th reachable distance rd from point p to point q k (p,q) takes point q as the object, compares the kth distance of q with the actual distance between q and p, and then selects the larger distance as the reachable distance from point p to point q. The calculation formula is as follows: rd k (p,q)=max{d k (q),d(p,q)} Calculate the local reachability density lrd k (p), as follows: Calculate the local outlier factor LOF of data point p k (p), as follows: Local Outlier Factor LOF k (p) value to judge, if LOF k If the deviation between the value of (p) and 1 is within the preset numerical range, p is judged to be data in a static environment. Otherwise, p is judged to be abnormal data, indicating that an action change has occurred at this moment, which is an abnormal change.

5. The CSI-based action recognition method according to claim 1, characterized in that: The Fisher score algorithm is used to extract features from the action signal fragments, and the preset features with the highest Fisher score are selected, including: Given a feature set {f1, f2, ..., f i ,…,f n }, for any CSI data sample x j ∈R m , calculate the feature f i The inter-class divergence S B (f i ) and the lth class sample in feature f i The intra-class divergence under as follows: Among them, i = 1, 2, ..., n, is the feature subscript; j = 1, 2, ..., m, is the sample category subscript; c represents the number of sample categories; d l is the number of samples in the lth class; Indicates that the l-th category sample has feature f i The mean value on i Indicates that all samples have feature f i The mean on ; Indicates that the jth sample in the lth class sample is in f i The value of the following; Define feature f i The Fisher score is J fisher (f i ), according to the feature f i The inter-class divergence and intra-class divergence of J fisher (f i ) is expressed as: After calculating the Fisher score of each feature, the most discriminative feature is selected according to the value of the Fisher score. Specifically, the Fisher scores are sorted from high to low, and preset features are selected in turn for action recognition.

6. A development board, characterized in that: It includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the CSI-based action recognition method described in any one of claims 1 to 5.

7. A CSI-based action recognition system, characterized in that: It comprises a development board and a Wi-Fi device; the Wi-Fi device collects CSI data and sends the collected CSI data to the development board; the development board comprises at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the CSI-based motion recognition method described in any one of claims 1 to 5.

8. The CSI-based action recognition method according to claim 7, characterized in that: The Wi-Fi device includes a router.

Images