Indoor non-contact human activity recognition method and system

Through antenna array acquisition and principal component analysis, and combined with transfer learning CNN model, the problems of overfitting and data processing efficiency in indoor human activity recognition are solved, and the contactless human body with high precision and low complexity are achieved Activity recognition.

CN113920577BActive Publication Date: 2024-11-12NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110900089.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-08-06
Publication Date
2024-11-12
Estimated Expiration
2041-08-06

AI Technical Summary

Technical Problem

The prior art has problems such as overfitting and single features in indoor contactless human activities recognition, and the data processing time is expensive. How to use as few samples as possible to achieve high-precision recognition is a challenge.

Method used

The antenna array is used to collect reflected signals, remove noise through principal component analysis, and input it into a pre-trained CNN network model based on transfer learning algorithm. The transfer learning ability of the CNN model is used to apply knowledge from one location to other locations, reducing the system Complexity and training sample requirements.

Benefits of technology

It realizes high-precision position-independent human activities recognition in indoor equipment-free environments, reduces the impact of noise, improves recognition accuracy and efficiency, and can effectively monitor the activities of the elderly in home-based elderly care scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113920577B_ABST
    Figure CN113920577B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for indoor non-contact human activity recognition, including: using an antenna array to collect indoor reflected signals; filtering the reflected signals to obtain reflected signals after noise removal; inputting into a pre-trained human activity recognition model to determine the human activity category, wherein the human activity recognition model is a pre-trained CNN network model based on a transfer learning algorithm. Advantages: using an antenna array to collect human movements and perform activity recognition indoors, which can be applied to scenarios of home-based elderly care; denoising the original data to remove most of the high-frequency noise and retain the phase change of the signal; using a CNN structure for training is intended to reduce system complexity, and at the same time using the idea of ​​transfer learning to give the system the ability to learn independently and use the knowledge obtained from one location for other locations. Ultimately, the purpose of achieving high-precision location-independent perception using as little data as possible is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to an indoor non-contact human activity recognition method and system, belonging to the field of information technology. Background Art

[0002] Realizing high-level human-computer interaction is the future development direction. Accurate perception and understanding of human behavior is an indispensable technical support. Traditional human behavior recognition uses various sensors, such as physical sensors, cameras, etc. In recent years, with the rapid development of wireless networks, the role of radio frequency signals has expanded from a single communication medium to a non-invasive environmental sensing tool. Using wireless signals for human behavior recognition has liberated users from wearing restrictions and has broad development prospects.

[0003] The basic principle is that RF signals propagate through multiple paths in a wireless medium, reflect off different objects and reach the receiver, thus carrying information about the environment. The human body is a good reflector, and by analyzing the received RF signal patterns and features, human activities and behavioral states, such as breathing rate, gestures, and falls, can be detected. In recent years, deep learning, which has received widespread attention, has been successfully applied to various fields such as speech recognition and graphic recognition, and has achieved remarkable results in feature extraction. By extracting features from wireless signals in indoor environments through deep learning, human activity recognition can be performed.

[0004] However, if redundant data is used to train a model, it may lead to overfitting, and the single feature may lead to inaccurate recognition. In addition, it takes a lot of time to process a large amount of data and use them for training. In practical applications, how to use as few samples as possible to achieve high-precision recognition is a challenging key issue. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide an indoor non-contact human activity recognition method and system.

[0006] In order to solve the above technical problems, the present invention provides an indoor non-contact human activity recognition method, comprising:

[0007] Use antenna array to collect reflected signals indoors;

[0008] Filtering the reflected signal to obtain a reflected signal after noise removal;

[0009] The human activity classification is input into a pre-trained human activity recognition model, where the human activity recognition model is a pre-trained CNN network model based on a transfer learning algorithm.

[0010] Furthermore, the reflection signal is filtered by using a principal component analysis method to obtain a reflection signal after noise removal, and the reflection signal after noise removal is input into a pre-trained human activity recognition model.

[0011] Furthermore, the filtering of the reflected signal by using a principal component analysis method comprises:

[0012] The indoor activity space is divided into several location areas, and the data packets of the fixed-time reflection signals of each activity in each location area are collected, and the static path part (the static path part corresponding to the static reflection signal of the indoor furniture, etc.) in the data packets of the fixed-time reflection signals of each activity is removed;

[0013] Cut the data packets of the fixed-time reflected signals of each activity except the static path part into a number of time blocks of fixed time lengths, and form a matrix H corresponding to the activity according to all the time blocks;

[0014] According to the matrix H, we can get the covariance matrix H T ×H, calculate the covariance matrix H T ×H correlation matrix, which is composed of the correlation coefficients of the covariance matrix. The size of the correlation matrix is ​​N×N, where N is the number of data streams;

[0015] Perform eigendecomposition on the obtained correlation matrix and calculate the eigenvector;

[0016] According to the formula h a =H×q a Perform signal reconstruction to obtain the reflected signal after noise removal, where q a is the ath eigenvector, h a is the ath principal component.

[0017] Furthermore, the training process of the human activity recognition model includes:

[0018] Obtain a pre-built CNN model, wherein the structure of the CNN model includes several convolutional layers, pooling layers and a fully connected layer, the convolutional layers are sequentially connected through the pooling layers, and the last convolutional layer is connected to the fully connected layer through the pooling layer;

[0019] Determine a suitable training position based on the conclusion that adjacent positions have similar statistical characteristics and the result of the maximum mean difference, and obtain the reflection signal of the training position after noise removal to form an initial training data set;

[0020] Use the initial training data set to train the CNN model and obtain the initial optimal model parameters;

[0021] Initialize the CNN model using the initial optimal model parameters;

[0022] Obtain the noise-removed reflection signals at inappropriate training positions to form a secondary training data set;

[0023] Freeze the parameters of the convolutional layer and pooling layer before the fully connected layer, and then use the secondary training dataset to train the fully connected layer until the upper limit of the number of iterations is reached or the preset accuracy is achieved, to obtain a trained human activity recognition model.

[0024] An indoor non-contact human activity recognition system, comprising:

[0025] A collection module, used for collecting indoor reflected signals using an antenna array;

[0026] A filtering module, used for filtering the reflected signal to obtain a reflected signal after noise is removed;

[0027] The processing module is used to input the reflected signal after noise removal into a pre-trained human activity recognition model to determine the human activity category. The human activity recognition model is a pre-trained CNN network model based on a transfer learning algorithm.

[0028] Furthermore, the filtering module is used to filter the reflection signal by using a principal component analysis method to obtain the reflection signal after noise is removed.

[0029] Furthermore, the filtering module includes:

[0030] A division module is used to divide the indoor activity space into several location areas, collect the data packets of the fixed-time reflection signals of each activity in each location area, and remove the static path part (the static path part corresponding to the static reflection signal of the indoor furniture, etc.) from the data packets of the fixed-time reflection signals of each activity;

[0031] A cutting module, used for cutting the data packets of the fixed-time reflection signals of each activity except the static path part into a number of time blocks of fixed time lengths, and forming a matrix H corresponding to the activity according to all the time blocks;

[0032] The first calculation module is used to obtain the covariance matrix H according to the matrix H T ×H, calculate the covariance matrix H T ×H correlation matrix, which is composed of the correlation coefficients of the covariance matrix. The size of the correlation matrix is ​​N×N, where N is the number of data streams;

[0033] The second calculation module is used to perform eigendecomposition on the obtained correlation matrix and calculate the eigenvector;

[0034] Reconstruction module, used to calculate the a=H×q a Perform signal reconstruction to obtain the reflected signal after noise removal, where q a is the ath eigenvector, h a is the ath principal component.

[0035] Furthermore, the processing module includes a model training module for

[0036] Obtain a pre-built CNN model, wherein the structure of the CNN model includes several convolutional layers, pooling layers and a fully connected layer, the convolutional layers are sequentially connected through the pooling layers, and the last convolutional layer is connected to the fully connected layer through the pooling layer;

[0037] Determine a suitable training position based on the conclusion that adjacent positions have similar statistical characteristics and the result of the maximum mean difference, and obtain the reflection signal of the training position after noise removal to form an initial training data set;

[0038] Use the initial training data set to train the CNN model and obtain the initial optimal model parameters;

[0039] Initialize the CNN model using the initial optimal model parameters;

[0040] Obtain the noise-removed reflection signals at inappropriate training positions to form a secondary training data set;

[0041] Freeze the parameters of the convolutional layer and pooling layer before the fully connected layer, and then use the secondary training dataset to train the fully connected layer until the upper limit of the number of iterations is reached or the preset accuracy is achieved, to obtain a trained human activity recognition model.

[0042] The beneficial effects achieved by the present invention are:

[0043] Using antenna arrays to collect human movements and perform activity recognition indoors can be applied to home-based elderly care scenarios; using principal component analysis to reduce noise on raw data can remove most high-frequency noise and retain the phase change of the signal; using a simple CNN structure for training is intended to reduce system complexity, and using the idea of ​​transfer learning can give the system the ability to learn independently and use the knowledge obtained from one location for other locations. Ultimately, the goal of achieving high-precision location-independent perception using as little data as possible is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a signal indoor propagation model during the implementation of the method of the present invention;

[0045] Figure 2 A flow chart for implementing the method of the present invention;

[0046] Figure 3This is a schematic diagram of the location division in a room when the method of the present invention is implemented. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0048] An indoor contactless human activity recognition method uses an antenna array to send signals and another antenna to receive signals. The RF reflection at different locations is calculated using the following formula:

[0049]

[0050] Where S(d,θ) is the signal at distance d and azimuth θ, N is the number of antennas in the antenna array, M is the number of samples taken for the RF signal, and s n,t is the signal received by antenna n at the mth sample, l is the antenna spacing, λ is the wavelength, c is the speed of light, j is the imaginary part of the complex number, and k is the slope of the RF signal.

[0051] Since the original data is noisy, principal component analysis is used to filter and remove noise. Principal component analysis is a statistical method that maps data to a low-dimensional space by extracting principal components, converting linearly correlated variables into a set of linearly uncorrelated variables.

[0052] After data preprocessing, feature selection is required for extraction and classification. Discrete wavelet transform is used to separate the features of the signal frequency domain and time domain for feature selection. CNN is used as a feature extractor. In order to achieve location-independent activity recognition, the classifier should have the ability that the knowledge learned from one location can be used in another location, and transfer learning can meet this requirement. In addition, the use of transfer learning can achieve the goal of using as little training data as possible. Most of the network parameters are obtained from sufficient source domain data, while a small part is learned from target domain data.

[0053] Next, consider how to select the training positions in the experiment. After dividing the room into 25 positions, select the same activity (such as sitting) to calculate the MMD (maximum mean difference) distance between them. MMD is the most commonly used effective indicator in transfer learning. The kernel two-sample test compares the distribution between two data sets n1 and n2. The empirical estimate of MMD is as follows:

[0054]

[0055] In the formula, Indicates that this distance is measured by φ() mapping the data into the reproducing kernel Hilbert space; φ represents the kernel function that maps the original data to, and here the Gaussian kernel is used; x b ,y d The bth and dth data in n1 and n2 are the same dimension. The smaller the MMD value, the more similar the data sets n1 and n2 are. The typical position for transfer learning is selected by calculating the MMD distance.

[0056] The network structure of the CNN model in this embodiment is: 3 convolutional layers are used, the size of the convolution kernel in each convolutional layer is 3×3, each convolutional layer is followed by a 2×2 pooling layer, and the last is a fully connected layer. CNN is trained by minimizing the cross entropy loss:

[0057]

[0058] Where probability distribution p is the expected output, and probability distribution q is the actual output. The smaller the value of H(p,q), the closer the actual output is to the expected output, and x represents a random variable.

[0059] In order to reduce the training samples, a method for indoor contactless human activity recognition based on transfer learning is proposed. Based on the above pre-trained CNN structure, the aim is to apply the results trained from the first few layers of the network to other locations. Different activities have different characteristics at the same location, and for the same activity, the received signals are significantly different at different locations. Therefore, it is difficult to find common features that represent activities at all locations, especially statistical features. Through simulation, it can be seen that adjacent locations have similar statistical features, which can be used to guide the selection of training locations in the experiment.

[0060] Finally, the number of samples for transfer learning is considered. The recognition accuracy varies with the number of transfer samples. If the room is divided into 25 locations, simulation shows that only three to four transfer samples are needed to achieve an accuracy of more than 90%.

[0061] Figure 1 It is a signal indoor propagation model during the implementation of the method of the present invention.

[0062] Wireless signals are affected by the surrounding environment during transmission. In addition to human activities, there may be other environmental factors, such as electronic devices in different frequency bands and other network communications. The health of the network may also affect the data, such as packet loss in the case of network congestion. During the transmission process, when encountering obstacles (dynamic or static), the signal will be refracted, reflected and scattered, which will produce multipath superposition signals at the receiver.

[0063] If the influence of various obstacles is not considered, and it is assumed that the propagation of wireless signals will not be blocked by obstacles and thus cause reflection, the propagation of wireless signals can be regarded as propagation along the line-of-sight path in free space. Considering that the signal will suffer a certain path loss during the transmission and reception process, the power P from the transmitter to the receiver is r (d) can be expressed as a function of distance d according to the Freese free space equation:

[0064]

[0065] Where P t is the transmit power; P r (d) is the received power, which is a function of the distance d; G r is the receiving antenna gain; G t is the transmitting antenna gain; λ is the wavelength of the signal; d is the distance between the transmitting end and the receiving end.

[0066] Normally, there is only one line-of-sight path and some paths reflected by the surrounding environment in the wireless signal transmission path, such as paths reflected by the floor, ceiling, and wall. If the reflection due to the ceiling and floor is taken into account, the receiving end power can be expressed as:

[0067]

[0068] Where h is the distance from the ceiling or floor to the line-of-sight path.

[0069] The presence and movement of people will also affect the wireless propagation path. Assuming that when someone appears indoors, there will be a path reflected by the person, and these scattered powers should also be added to the final received power, so the received power should be expressed as:

[0070]

[0071] where Δ is a simplified representation of the path length due to reflection from the person.

[0072] Figure 2 Flow chart of the method of the present invention.

[0073] Step 1: Signal acquisition

[0074] An antenna array is used for signal collection to identify specific human activities. In home scenarios, it can timely monitor the situation of elderly people falling at home. At the same time, it uses non-invasive environmental sensing tools, without the risk of privacy leakage.

[0075] Figure 3 This is a schematic diagram of the location division in a room when the method of the present invention is implemented.

[0076] There are furniture in the room, represented by tables. Only two are drawn near the wall in the figure. The relative positions of all furniture and the room do not change during the experiment. The transmitter and receiver are placed at the two ends of the same side of the room, about 1.2 meters above the ground. The area of ​​the room as shown in the figure is divided into 25 positions, each with the same area. The blue arrow indicates the path of the signal reflected by a person or object.

[0077] Step 2: Preprocessing

[0078] Use principal component analysis for filtering. Principal component analysis is a statistical method that maps data to a low-dimensional space by extracting principal components, converting linearly correlated variables into a set of linearly uncorrelated variables. Since the signal changes caused by human actions are correlated in each link, we can use principal component analysis to remove noise. The steps for using principal component analysis for noise reduction are as follows:

[0079] (a) PCA preprocessing. First, we remove the static path part of each signal stream, that is, the line-of-sight path part. By calculating the mean of the signal amplitude over a period of time as a constant offset, we collect 5-second data packets and take the average of their signal amplitudes. Then the original signal is cut into time blocks of 1 second in length to form a matrix H. We choose the time length to be 1 second. Usually, the distance a person moves or the change in movement in 1 second is very small, so enough data can be collected to calculate the relevant estimation.

[0080] (b) Correlation estimation. Next, we calculate the matrix H T ×H correlation matrix, the correlation matrix is ​​composed of the correlation coefficients of the matrices, and the size of the correlation matrix is ​​N×N, where N is the number of data streams.

[0081] (c) Eigendecomposition: Perform eigendecomposition on the obtained correlation matrix and calculate the eigenvector.

[0082] (d) Signal reconstruction. According to formula h a =H×q a Signal reconstruction is performed, where q a is the ath eigenvector, h a is the ath principal component. That is, we project the original data onto the feature subspace. This paper selects the first principal component h1 and retains its principal component for subsequent activity interception and feature extraction.

[0083] Step 3: Feature Selection

[0084] Considering the characteristics of each action, it has different effects on CSI data, which are manifested in different aspects, such as different durations, amplitude changes, change speeds, and sizes. Related to activities are duration and frequency. Duration represents the time it takes a person to perform an action, and frequency represents the speed of change of different signal paths caused by the body movement caused by the person's activities. Different activities may have similar durations, but different frequencies. For example, the duration of sitting and falling is relatively short, but the speed of change of the path of falling is significantly higher than that of sitting. Therefore, the power of CFR (crest factor reduction) of falling is higher than that of sitting. Similarly, different actions may have the same frequency, but different durations. For example, running and falling have similar frequencies, but the duration of falling is much shorter. Therefore, to analyze the power of CFR of human activities, we need to extract frequencies with different resolutions from different time scales.

[0085] Discrete wavelet transform is often used to obtain the time-frequency relationship of signals. It has two main advantages: one is that it has the best resolution in the time domain and frequency domain; the other is that it can achieve a fine-grained multi-scale analysis. Discrete wavelet transform can show the distribution of signals at different frequencies and decompose them at different decomposition scales. Level 1 represents the first-level decomposition. The low-frequency signal of the first-level decomposition can continue to be decomposed, and finally a multi-layer decomposition is achieved. Discrete wavelet transform calculates the energy at different wavelet levels based on the given signal. Each level represents a corresponding frequency range, in which the two adjacent wavelet levels decrease exponentially. For example, if the DWT of wavelet level 1 represents a frequency range of 150 to 300 Hz, the frequency range represented by the DWT of wavelet level 2 is half of the frequency range of wavelet level 1, that is, 75 to 150 Hz. The higher the energy of the wavelet level, the closer the frequency of the change of the action impact path is to that range.

[0086] Dividing the denoised signal into different wavelet levels and reconstructing several layers of signals can depict the distribution of the signal at different frequency scales. These combined frequency domain and time domain features can better describe the action.

[0087] Step 4: Feature extraction and training

[0088] CNN has amazing capabilities in feature representation. Therefore, it is used as a feature extractor. How to reduce complexity and overfitting as well as improve accuracy should be considered.

[0089] In order to achieve location-independent activity recognition, the classifier should have the ability that the knowledge learned from one location can be used in another location. Transfer learning can meet this requirement. In addition, fine-tuning through transfer learning can achieve the use of as little training data as possible. Most of the network parameters are obtained from sufficient source domain data, while a small part is learned from target domain data.

[0090] After designing a simple CNN network architecture, the transfer learning scheme is based on the pre-trained CNN structure mentioned above. The features extracted from the first few layers of the network (the features learned from some positions) are applied to other positions. The specific process is as follows. First, the CNN model is trained using a training dataset consisting of some positions, and then the optimal model parameters are obtained. Next, these updated parameters are used to initialize the network. Finally, we freeze the layer before the fully connected layer, and then use the transfer dataset with only a few samples at each position to train the subsequent fully connected layers of the network. This will greatly reduce the demand for training samples and improve recognition accuracy and convergence speed. The forward and back propagation of CNN training are related to the entire network, while CNN transfer learning only involves the last two layers. The parameters required for fine-tuning are significantly reduced.

[0091] Next, consider the training locations and sample size for transfer learning. The relationship and distribution differences of data collected at different locations can be analyzed from a statistical perspective. It can be found that for the same activity, the data distribution is different at different locations. Therefore, it is difficult to find common features that represent activities at all locations, especially statistical features. Through simulation, it can be seen that adjacent locations have similar statistical features, which can guide us in selecting training locations in the experiment.

[0092] In addition, the maximum mean difference (MMD) is used to demonstrate data distribution differences. We can use the same activity at all locations to calculate the MMD distance. The smaller the MMD value, the more similar the datasets n1 and n2 are. For samples from the same location, the MMD is small, while for samples from different locations, the MMD is large. The above conclusions can also be used to guide us in choosing training locations in experiments.

[0093] When performing performance verification, we divided the experiments into three categories according to different training data. One is to mix the data of all locations together as training samples. Another is to use data from only some locations as training data, but the test data is not only located in these locations. The last one is to use a part of the locations for training the initial model, and then use very few samples from all locations for transfer learning to fine-tune the model. Finally, it was found that selecting three to four samples from 25 samples as training locations can achieve higher accuracy.

[0094] Step 5: Activity Identification

[0095] There are four actions that need to be recognized, namely walking, standing up, sitting down and falling. The precision index can be used as the evaluation standard for the accuracy of activity recognition:

[0096]

[0097] TP means that the data determined to be activity i is finally correctly classified as activity i, and FP means that the data that is not activity i is classified as activity i. The higher the precision, the higher the proportion of correctly classified data in the classified data.

[0098] The present invention also provides an indoor non-contact human activity recognition system, comprising:

[0099] A collection module, used for collecting indoor reflected signals using an antenna array;

[0100] A filtering module, used for filtering the reflected signal to obtain a reflected signal after noise is removed;

[0101] The processing module is used to input the reflected signal after noise removal into a pre-trained human activity recognition model to determine the human activity category. The human activity recognition model is a pre-trained CNN network model based on a transfer learning algorithm.

[0102] Furthermore, the filtering module is used to filter the reflection signal by using a principal component analysis method to obtain the reflection signal after noise is removed.

[0103] Furthermore, the filtering module includes:

[0104] A division module is used to divide the indoor activity space into several location areas, collect the data packets of the fixed-time reflection signals of each activity in each location area, and remove the static path part (the static path part corresponding to the static reflection signal of the indoor furniture, etc.) from the data packets of the fixed-time reflection signals of each activity;

[0105] A cutting module, used for cutting the data packets of the fixed-time reflection signals of each activity except the static path part into a number of time blocks of fixed time lengths, and forming a matrix H corresponding to the activity according to all the time blocks;

[0106] The first calculation module is used to obtain the covariance matrix H according to the matrix H T ×H, calculate the covariance matrix H T ×H correlation matrix, which is composed of the correlation coefficients of the covariance matrix. The size of the correlation matrix is ​​N×N, where N is the number of data streams;

[0107] The second calculation module is used to perform eigendecomposition on the obtained correlation matrix and calculate the eigenvector;

[0108] Reconstruction module, used to calculate the a =H×q a Perform signal reconstruction to obtain the reflected signal after noise removal, where q a is the ath eigenvector, h a is the ath principal component.

[0109] Furthermore, the processing module includes a model training module for

[0110] Obtain a pre-built CNN model, wherein the structure of the CNN model includes several convolutional layers, pooling layers and a fully connected layer, the convolutional layers are sequentially connected through the pooling layers, and the last convolutional layer is connected to the fully connected layer through the pooling layer;

[0111] Determine a suitable training position based on the conclusion that adjacent positions have similar statistical characteristics and the result of the maximum mean difference, and obtain the reflection signal of the training position after noise removal to form an initial training data set;

[0112] Use the initial training data set to train the CNN model and obtain the initial optimal model parameters;

[0113] Initialize the CNN model using the initial optimal model parameters;

[0114] Obtain the noise-removed reflection signals at inappropriate training positions to form a secondary training data set; freeze the parameters of the convolutional layer and the pooling layer before the fully connected layer, and then use the secondary training data set to train the fully connected layer until the upper limit of the number of iterations is reached or the preset accuracy is achieved, thereby obtaining a trained human activity recognition model.

[0115] The present invention extends the existing combination of equipment-free human activity recognition and deep learning, combines transfer learning with CNN, and selects training positions and the number of training samples through experiments, thereby achieving the purpose of high recognition accuracy with less data, and can promote the use of equipment-free human activity recognition in daily life.

[0116] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0117] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0120] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for indoor non-contact human activity recognition, characterized in that: include: Use antenna array to collect reflected signals indoors; Filtering the reflected signal to obtain a reflected signal after noise removal; Input into a pre-trained human activity recognition model to determine the human activity category, wherein the human activity recognition model is a pre-trained CNN network model based on a transfer learning algorithm; The reflected signal is filtered by a principal component analysis method to obtain a reflected signal after noise removal, and the reflected signal after noise removal is input into a pre-trained human activity recognition model; The step of filtering the reflected signal by using a principal component analysis method comprises: Divide the indoor activity space into several location areas, collect data packets of fixed-time reflection signals of each activity in each location area, and remove the static path part in the data packets of fixed-time reflection signals of each activity; Cut the fixed-time reflected signal data packets of each activity except the static path part into several time blocks of fixed time length, and form a matrix corresponding to the activity based on all the time blocks H ; According to the matrix H Get the covariance matrix H T ×H , calculate the covariance matrix H T ×H The correlation matrix is ​​composed of the correlation coefficients of the covariance matrix. The size of the correlation matrix is N×N ,in N is the number of data streams; Perform eigendecomposition on the obtained correlation matrix and calculate the eigenvector; According to the formula Perform signal reconstruction to obtain the reflected signal after removing noise, where It is a feature vectors, It is a principal components.

2. The indoor non-contact human activity recognition method according to claim 1, characterized in that: The training process of the human activity recognition model includes: Obtain a pre-built CNN model, wherein the structure of the CNN model includes several convolutional layers, pooling layers and a fully connected layer, the convolutional layers are sequentially connected through the pooling layers, and the last convolutional layer is connected to the fully connected layer through the pooling layer; Determine a suitable training position based on the conclusion that adjacent positions have similar statistical characteristics and the result of the maximum mean difference, and obtain the reflection signal of the training position after noise removal to form an initial training data set; Use the initial training data set to train the CNN model and obtain the initial optimal model parameters; Initialize the CNN model using the initial optimal model parameters; Obtain the noise-removed reflection signals at inappropriate training positions to form a secondary training data set; Freeze the parameters of the convolutional layer and pooling layer before the fully connected layer, and then use the secondary training dataset to train the fully connected layer until the upper limit of the number of iterations is reached or the preset accuracy is achieved, to obtain a trained human activity recognition model.

3. An indoor non-contact human activity recognition system, characterized in that: include: A collection module, used for collecting indoor reflected signals using an antenna array; A filtering module, used for filtering the reflected signal to obtain a reflected signal after noise is removed; A processing module, used for inputting the reflected signal after noise removal into a pre-trained human activity recognition model to determine the human activity category, wherein the human activity recognition model is a pre-trained CNN network model based on a transfer learning algorithm; The filtering module is used to filter the reflection signal by using the principal component analysis method to obtain the reflection signal after noise is removed; The filtering module comprises: A division module, used to divide the indoor activity space into a number of location areas, collect data packets of fixed-time reflection signals of each activity in each location area, and remove the static path part in the data packets of fixed-time reflection signals of each activity; The cutting module is used to cut the data packets of the fixed-time reflection signals of each activity except the static path part into several time blocks of fixed time length, and form a matrix corresponding to the activity according to all the time blocks. H ; The first calculation module is used to calculate the matrix H Get the covariance matrix H T ×H , calculate the covariance matrix H T ×H The correlation matrix is ​​composed of the correlation coefficients of the covariance matrix. The size of the correlation matrix is N×N ,in N is the number of data streams; The second calculation module is used to perform eigendecomposition on the obtained correlation matrix and calculate the eigenvector; Reconstruction module, used to Perform signal reconstruction to obtain the reflected signal after removing noise, where It is a feature vectors, It is a principal components.

4. The indoor non-contact human activity recognition system according to claim 3, characterized in that: The processing module includes a model training module for Obtain a pre-built CNN model, wherein the structure of the CNN model includes several convolutional layers, pooling layers and a fully connected layer, the convolutional layers are sequentially connected through the pooling layers, and the last convolutional layer is connected to the fully connected layer through the pooling layer; Determine a suitable training position based on the conclusion that adjacent positions have similar statistical characteristics and the result of the maximum mean difference, and obtain the reflection signal of the training position after noise removal to form an initial training data set; Use the initial training data set to train the CNN model and obtain the initial optimal model parameters; Initialize the CNN model using the initial optimal model parameters; Obtain the noise-removed reflection signals at inappropriate training positions to form a secondary training data set; Freeze the parameters of the convolutional layer and pooling layer before the fully connected layer, and then use the secondary training dataset to train the fully connected layer until the upper limit of the number of iterations is reached or the preset accuracy is achieved, to obtain a trained human activity recognition model.