A wireless signal-based method for detecting falls by the elderly

Abstract

The application discloses a wireless signal-based fall detection method for the elderly, and utilizes a CSITOOL tool to build a CSI data sending and receiving platform; a wireless signal-based fall detection process firstly removes outliers and abnormal values from collected CSI signals and carries out wavelet denoising on the collected signals; finally, a network is constructed, network features are extracted, and a fall fingerprint library is established, so that evaluation and early warning information of fall detection is regressed out. Meanwhile, the application also proposes a multi-antenna CSI signal layered sampling strategy. The application can greatly reduce actual operation cost and implement cost, thereby effectively realizing identification of accidental falls of the elderly and having certain application value in wireless sensing and behavior identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of human behavior recognition, and particularly relates to a method for detecting the falling of the old people based on wireless signals. BACKGROUND

[0002] The population aging in China is an objective population problem, and the young people have great work pressure, and there may be negligence or unobservable and uncontrollable risks in taking care of the old people and children, such as the abnormal falling of the old people in the room, which objectively exists, and if these hidden dangers can be monitored and reported, it will provide a great guarantee for the life safety.

[0003] The behavior recognition based on images is a very active research field in computer vision, and with the continuous development of computer technology and the deepening of deep learning research, the recognition method based on the camera is very mature, whether it is to use a monocular RGB camera to recognize human behavior, or to use multiple cameras to obtain multiple perspective RGB images to recognize three-dimensional human behavior, or to obtain depth distance images based on a depth camera to recognize three-dimensional human behavior, these methods have good performance in behavior recognition.

[0004] However, the traditional sensor technology through the camera is obviously limited by light, shielding and background, and there is a privacy problem; compared with the camera, the behavior perception and estimation technology based on wireless signals shows the potential to support more complex interactions between humans and physical environments, and has the potential to realize mature applications.

[0005] In recent years, wireless perception has made some achievements in posture perception, and experiments show that wireless signals can well complete the behavior task. With the continuous popularity of WiFi systems, WiFi signals for human perception can be well applied to the market, including the activity detection and life assistance of patients or old people who are inconvenient to move, and the use of WiFi signals can well avoid some privacy problems; in addition, WiFi has the advantages of low cost, simple and convenient application compared with night vision cameras. SUMMARY

[0006] In order to overcome the problems of the traditional camera sensor technology in human behavior recognition, especially the falling of the old people in private places, the sensor is limited by light, shielding, background and privacy problems, the present application provides a method for detecting the falling of the old people based on wireless signals, which can greatly reduce the actual operation cost and the realization cost, so as to effectively realize the recognition of the accidental falling of the old people, and has certain application value in wireless perception and behavior recognition.

[0007] The technical scheme adopted by the present application is:

[0008] A method for detecting the falling of the old people based on wireless signals, the method comprising the following steps:

[0009] Step 1: Use the CSITOOL tool to build a channel state information (CSI) data sending and receiving platform;

[0010] Step 2: Collect CSI data containing volunteer actions in an experimental environment, the actions are divided into 6 categories, including forward fall, backward fall, sitting on a chair and then falling forward, falling backward while sitting down, side fall and normal walking;

[0011] Step 3: Preprocess the collected CSI data, and use the conjugate multiplication method to solve the phase offset of the signal;

[0012] Step 4: Use Hampel outlier filtering to remove obvious outliers of the original signal, and replace the outliers with the sample median;

[0013] Step 5: Use discrete wavelet transform to filter the original signal and remove the static direct current component;

[0014] Step 6: Convolution-based input signal bilinear interpolation sampling;

[0015] Step 7: Build a neural network model and train it;

[0016] Step 8: Determine the evaluation index of the neural network model and evaluate it, with hit rate H as the evaluation index.

[0017] Further, in step 3, the phase offset elimination process is as follows:

[0018] First, build the CSI model formula:

[0019]

[0020] Wherein represents the phase offset caused by time alignment offset, sampling frequency offset, and carrier frequency offset, H s (f) represents the environmental changes and environmental object reflected signals on the path, L represents the human body reflected part of the signal link, a l (f, t) represents the amplitude noise of path l, represents the Doppler shift of path l;

[0021] The elimination method formula is as follows:

[0022]

[0023] Wherein, H cm (f, t) represents the result of conjugate multiplication, respectively, H 1,s (f), H 2,s(f) represents the static components of the conjugate signals H1(f, t), H2(f, t), k, l represent the dynamic link parts of the two signals, a l (f, t), a k (f, t) represents the amplitude noise of the two links, (f, t) represents the Doppler shift of the two links. In the elimination result, the first term represents the static path of the two antennas, the second term represents the product of the dynamic path components of the two antennas, and the other two terms represent the product between the static component of one antenna and the dynamic component of the other antenna. The static variable is eliminated by subtracting the average value of the signal, and the component not containing human body information is eliminated.

[0024] Further, the process of step 6 is as follows:

[0025] Each group of channel state information matrices of the input signal is determined by the number of receiving and transmitting antennas, where the number of transmitting antennas and receiving antennas is m = n = 3, and each group of state matrices has a fixed number of sampling points D = 150, so each group of human body key point information sequences can be represented by a multi-dimensional matrix:

[0026]

[0027] The multi-dimensional matrix i is processed using bilinear interpolation, the number of interpolations b = 32, and each matrix element i pq is replaced by the interpolation module B pq ;

[0028]

[0029] The function f(a) is the weight corresponding to the interpolation, that is, f(a) = (a + 1) / b; After interpolation, the channel state information matrix I' ∈ R M×N×D ; Further, the process of step 7 is as follows:

[0030] The convolution layer includes convolution kernels of different sizes, each size contains a large number of convolution kernels, and the number of channels of the convolution kernel is the same as that of the CSI matrix, which is D = 150;

[0031] Assuming that the length and width of the Kth convolution kernel are H, K, the convolution kernel is represented as W k = R H×K×D , that is:

[0032]

[0033] The convolution operation is the extraction of local features of I', and the process of the convolution operation is: when the convolution kernel W k meets I' 1,1 , the extracted feature is:

[0034]

[0035] In the formula, i' is the i-th row of the j-th column of the sub-element of I', i,j i,j is the weight of i', i,j is the bias, and Relu is a nonlinear activation function:

[0036] f(x)=max(0,x)

[0037] The convolution operation is W k respectively with step S H , S K From top to bottom of I', front to back, the generated feature combination is:

[0038]

[0039] After a series of convolution and pooling operations, all the extracted features are connected in an end-to-end manner, the length is normalized to 256, and finally input into several fully connected layers and an output layer, the number of units of the output layer is 2, and the falling detection result is obtained through the softmax activation function.

[0040] The loss function is defined as:

[0041] Loss=λL w (y predict )

[0042] Wherein λ is a regular coefficient L w is the mean square error of the sample, defined as:

[0043]

[0044] Wherein and respectively represent the true value label and the prediction result, and n represents the total number of all inputs.

[0045] The technical concept of the application is that: object and human body recognition based on wireless signal is a research field that has emerged in recent years, and many scholars have made a lot of research results in the field of wireless sensing. China's population aging is an objective population problem, and the death of the elderly due to unattended falling is a problem that must be addressed. Therefore, realizing contactless and privacy-free emergency alarm to protect the vulnerable groups such as the elderly is the current demand of the national aging. The falling recognition based on wireless signal of the application ingeniously combines channel state information and human motion information, realizes indoor old people falling detection and early warning.

[0046] ​The application has the advantages that actual operation cost and achievable cost can be greatly reduced, so that the identification of the accidental falling of the old people is effectively realized, and the application value in wireless sensing and behavior identification is possessed. BRIEF DESCRIPTION OF DRAWINGS

[0047] Fig. 1 is a work flow chart of the old people falling detection method based on the wireless signal of the application;

[0048] Fig. 2 is a working schematic diagram of the CSI data collection and detection of the application;

[0049] Fig. 3 is a hit rate obtained by the wireless signal falling detection evaluation method of the application. DETAILED DESCRIPTION

[0050] The advantages and features of the application can be more simply and quickly understood by the person skilled in the art by combining the preferred embodiments of the application with the drawings, so that the protection scope of the application can be more clearly and definitely divided and defined.

[0051] Referring to Figs. 1-3 A kind of old people falling detection method based on wireless signal, comprising the following steps:

[0052] Step 1: using CSI TOOL tool to build channel state information (CSI) data sending and receiving platform;

[0053] Step 2: simultaneously collect CSI data containing volunteer action in experimental environment, action is divided into 6 categories, including front fall, back fall, sit on chair and then fall forward, fall backward when sitting down, side fall and normal walking;

[0054] Step 3: the collected CSI data is preprocessed, the conjugate multiplication method is used to solve the phase offset of signal, and the formula is eliminated:

[0055]

[0056] Step 4: the obvious outliers of original signal are removed by Hampel outlier filtering, and the outliers are replaced by sample median value;

[0057] Step 5: the original signal is filtered by using discrete wavelet transform, and static direct current component is removed;

[0058] Step 6: the input signal bilinear interpolation sampling based on convolution, each group of human body key point information sequence is represented by a multidimensional matrix:

[0059]

[0060] The multi-dimensional matrix I is processed using bilinear interpolation, the interpolation number b = 32, and each matrix element i pq is interpolated by the interpolation module B pq instead

[0061]

[0062] Step 7: Constructing and training the neural network model, the convolutional layer includes convolution kernels of different sizes, each size contains a large number of convolution kernels, the channel number of the convolution kernel is the same as that of the CSI matrix, both are D = 150, the length and width of the Kth convolution kernel are H and K, and the convolution kernel is represented as W k = R H×K×D , the convolution operation extracts the local features of I', when the convolution kernel W k meets I' 1,1 , the extracted features are :

[0063]

[0064] The convolution operation is W k from top to bottom of I' with step S H , S K from front to back, and the generated feature combination is

[0065]

[0066] After a series of convolution and pooling operations, all the extracted features are connected in an end-to-end manner, the length is normalized to 256, and finally input into several fully connected layers and an output layer, the number of units of the output layer is 2, and the falling detection result is obtained through the softmax activation function;

[0067] The loss function is

[0068] Loss = λL w (y predict )

[0069] The mean square error L w of the sample is

[0070]

[0071] Step 8: Determine the evaluation index of the neural network model and evaluate it, taking the hit rate H as the evaluation index.

[0072] The processing process of the embodiment is as follows:

[0073] 1. As shown in the attached Fig. 2As shown, our experimental platform consists of two computers, both equipped with Intel5300 network cards and CSITOOL tools, one as the transmitting end to transmit CSI signals, and the other as the receiving end to receive CSI signals.

[0074] 2. The specific implementation site is an empty classroom, and the volunteers move in the activity area of about 5.5 meters around, as shown in the accompanying Fig. 2 ;

[0075] 3. At each collection, the human body is stationary and stands at the data collection point as shown in the accompanying Fig. 2 , and the CSI signal is collected, and 30 seconds are collected for each action. After the collection is completed, multiple.dat files from different volunteers can be obtained for each action;

[0076] 4. The collected CSI data is preprocessed, the conjugate multiplication method is used to solve the phase offset of the signal, and the Hampel outlier filter is used to remove the obvious outliers of the original signal, and the abnormal value is replaced with the sample median; after removing the abnormal value, the original signal is filtered by using the discrete wavelet transform, the static direct current component is removed, and the signal characteristics are retained;

[0077] 5. The processed CSI signal is input into the convolutional neural network, the error is regressed in the gradient minimum direction, and the model is trained;

[0078] 6. After the fall detection model is trained, the model is tested, and the hit rate of the fall detection is calculated (as shown in the accompanying Fig. 3 );

[0079] The above only describes the embodiments of the present application, and does not limit the patent range of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection range of the present application.

Claims

1. A wireless signal-based elderly fall detection method, characterized in that the method comprises the following steps: Step 1: Use the CSITOOL tool to build a channel state information (CSI) data sending and receiving platform; Step 2: Collect CSI data containing volunteer actions in an experimental environment, the actions are divided into 6 categories, including forward fall, backward fall, sitting on a chair and then falling forward, falling backward while sitting down, side fall and normal walking; Step 3: Preprocess the collected CSI data, and use the conjugate multiplication method to solve the phase offset of the signal; Step 4: Use Hampel outlier filtering to remove obvious outliers in the original signal, and replace the outliers with the sample median; Step 5: Use discrete wavelet transform to filter the original signal and remove static direct current components; Step 6: Bilinear interpolation sampling based on convolution of input signal, the process is as follows: Each group of channel state information matrix of input signal is determined by the number of receiving and transmitting antennas, where the number of transmitting antennas and receiving antennas is m=n=3, each group of state matrix has a fixed number of sampling points D=150, then each group of human key point information sequence is represented by a multi-dimensional matrix: The multi-dimensional matrix I is processed using bilinear interpolation, the interpolation number b = 32, each matrix element i pq is interpolated by the interpolation module B pq instead; The function f(a) interpolates the corresponding weight, that is, f(a) = (a + 1) / b; the channel state information matrix I' after interpolation is I' ∈ R M×N×D ; Step 7: Construct a neural network model and train it; Step 8: Determine the evaluation index of the neural network model and evaluate it, with hit rate H as the evaluation index.

2. The wireless signal-based fall detection method for the elderly according to claim 1, wherein The process of step 3 is as follows: The CSI model formula is constructed as follows: wherein denotes the phase offset caused by the time alignment offset, the sampling frequency offset, and the carrier frequency offset, H s (f) represents the environmental changes and environmental objects reflecting signals on the path, L represents the human body reflecting part of the signal link, a l (f, t) represents the amplitude noise of the path l, denotes the Doppler shift of the path l; The elimination method formula is as follows: where H cm (f,t) denotes the result of the conjugate multiplication, respectively H 1,s (f), H 2,s (f) denotes the static components of the conjugate signals H1(f,t), H2(f,t), k, l denote the dynamic link parts of the two signals, a l (f,t), a k (f,t) denotes the amplitude noise of the two links, denotes the Doppler shift of the two links; in the cancellation result, the first term denotes the product of the static paths of the two antennas, the second term denotes the product of the dynamic path components of the two antennas, the other two terms denote the products between the static component of one antenna and the dynamic component of the other antenna, the static variables are cancelled by subtracting the average of the signals, the components not containing the human body information are cancelled. 3.The wireless signal-based fall detection method for the elderly according to claim 1 or 2, characterized in that, The process of step 7 is as follows: The convolution layer includes convolution kernels of different sizes, each size contains a convolution kernel, and the number of channels of the convolution kernel is the same as that of the CSI matrix, both of which are D=150; Assuming that the length and width of the Kth convolution kernel is H, K, the convolution kernel is represented as W k = R H×K×D That is: The convolution operation is the extraction of local features of I'. The process of the convolution operation is as follows: when the convolution kernel W k meets I', 1,1 the extracted features are : In the above formula, i' i,j is the i'th row and j'th column sub-element of I', is the weight of i' i,j , is the bias, and Relu is a non-linear activation function: f(x) = max(0, x) The convolution operation is W k In steps S H , S K From top to bottom, front to back, the resulting feature combinations are: After the convolution and pooling operations of the channel state information, all the extracted features are connected in an end-to-end manner, the length is normalized to 256, and finally input into the fully connected layer and the output layer, the number of units of the output layer is 2, and the fall detection result is obtained through the softmax activation function; The loss function is defined as: Loss = λL w (y predict ) where λ is the regularization coefficient L w is the mean squared error of the sample, defined as: wherein and y and y^ represent the true value label and the prediction result, respectively, and n represents the total number of all inputs.

Images