Clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix

Abstract

A clutter elimination method for fall detection by millimeter-wave radar based on a time-frequency information matrix is ​​used to address the problem in the prior art that the radar echo signal contains both target information and ground clutter noise. Excessive clutter will cause the recognition network to focus more on clutter information and ignore the time-frequency information containing the target, leading to erroneous fall detection. Compared with the method of inputting a time-frequency graph into the network for recognition, the present application directly inputs the matrix containing the time-frequency information after the echo signal is transformed by the Fast Fourier Transform (FFT) into the network for recognition. The present application does not ignore the time-frequency information containing the target. Different from the method of removing "spots" in the time-frequency graph, the present application uses a search boundary algorithm to determine the clutter noise matrix in the time-frequency information matrix, and uses the Gaussian distribution probability to determine the threshold value for clutter elimination, effectively improving the recognition effect of the millimeter-wave radar for fall detection in different clutter noise environments.

Description

Technical Field

[0001] The present invention relates to the technical field of radar clutter elimination, in particular to a clutter elimination method for millimeter wave radar fall detection based on a time-frequency information matrix. Background Art

[0002] Falls are a particularly common cause of disability among the elderly, seriously affecting their quality of life and even causing death. Fall monitoring is very necessary for elderly people living alone and whose children are not around. Currently, the common fall monitoring methods are: (1) Surveillance cameras: Install surveillance cameras in each room and use video sequence information to detect whether the elderly have fallen. (2) Wearable devices: The elderly need to wear sensor devices such as smart watches, and use the changes in acceleration and gyroscope horizontal angle information when falling to determine whether they are currently in a fall state. (3) Millimeter wave radar: Install radar equipment in the room and identify the current fall behavior based on the significant changes in the time-frequency information in the radar echo when the elderly fall.

[0003] Fall detection methods based on surveillance cameras face challenges with privacy protection for the elderly and poor detection performance in strong and weak lighting conditions. Wearable device-based fall detection methods, on the other hand, face the inconvenience of elderly people forgetting to wear their devices. Furthermore, fall detection methods based on millimeter-wave radars significantly impact detection performance because radar echo signals contain both target information and ground clutter. The presence of excessive clutter causes the recognition network to prioritize clutter information while ignoring the time-frequency information of the target, leading to false fall detection.

[0004] At present, most fall recognition methods based on millimeter-wave radar use a time-frequency graph as input to the network recognition method, that is, converting the time-frequency information matrix into the form of an image, and on this basis, eliminating the "spots" other than the target information on the image, thereby achieving the purpose of eliminating clutter in the noise. Summary of the Invention

[0005] The purpose of the present invention is to propose a clutter elimination method for millimeter-wave radar fall detection based on a time-frequency information matrix to address the problem in the prior art that, because the radar echo signal contains both target information and ground clutter noise, the presence of excessive clutter will cause the recognition network to focus more on the clutter information and ignore the time-frequency information containing the target, resulting in fall detection errors.

[0006] The technical solution adopted by the present invention to solve the above technical problems is:

[0007] The clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix includes the following steps:

[0008] Step 1: Obtain the transmission signal of the millimeter wave radar within one frequency modulation cycle and the echo signal containing the human target;

[0009] Step 2: Mix the transmitted signal and the echo signal to obtain an intermediate frequency signal, and sample the intermediate frequency signal to obtain a discrete intermediate frequency signal;

[0010] Step 3: Perform a short-time Fourier transform on the discrete intermediate frequency signal to obtain a range Doppler map for each frame of data. A spectral peak search is performed on the range Doppler map for each frame of data to obtain the target's range. The number of range gates is set to 9 to obtain micro-Doppler information for the target's limbs, excluding the torso. The echo energy within the range gate number is summed and averaged to obtain the Doppler information for that frame.

[0011] Step 4: Divide the time domain of the intermediate frequency signal into time series of equal length through short-time Fourier transform, and perform fast Fourier transform on each time series to obtain the range-Doppler map of the frame data. Then, combine the range-Doppler map of the frame data with the Doppler information of the frame to obtain the time-frequency information matrix of the intermediate frequency signal segment.

[0012] Step 5: Smooth each row of data in the time-frequency information matrix to obtain the smoothed data of the row of data, specifically:

[0013] In each row of data, select data in sequence starting from the first data until the fifth data from the last is selected. For each selected data, obtain the data and the following four data, and then calculate the average of the five data to obtain the average value y mean , all the mean values ​​y in each row of data mean After the combination, a 1×(NL) matrix is ​​obtained. The 1×(NL) matrix is ​​the smoothed data of the row data, L is the slider size, and N is the number of repeated cycles;

[0014] Step 6: From the center of each row of smoothed data, that is, the zero frequency, search for the minimum value point that is smaller than the median of the row of data to the left and right ends respectively, and record the coordinates of the minimum value point on the left and right ends;

[0015] Step 7: Repeat step 6 to obtain the coordinates of the minimum points on the left and right ends of each frame of the time-frequency information matrix, and obtain the leftmost boundary point y left and the rightmost boundary point y right , and then we get the clutter noise matrix Noise, which is expressed as:

[0016] Noise=[a1,a2,…,a left ,a right ,…,a N-1 ,a N ]

[0017] Among them, a i is the i-th column data of the time-frequency information matrix, a left is the yth time-frequency information matrix left Column data, a right is the yth time-frequency information matrix right Column data, clutter noise matrix Noise obeys Gaussian distribution;

[0018] Get the maximum value of the data in the clutter noise matrix Noise, and calculate the probability p of the maximum value in the Gaussian distribution. Take the noise energy value corresponding to the Gaussian distribution probability p / 10 as the clutter noise elimination threshold, and then complete the clutter elimination.

[0019] Furthermore, the steps for obtaining the range Doppler map of each frame of data are as follows:

[0020] The signal obtained by the framing operation is subjected to windowing filtering, range-dimensional fast Fourier transform and Doppler-dimensional fast Fourier transform in sequence to obtain a range-Doppler map for each frame of data.

[0021] Furthermore, the transmission signal is expressed as:

[0022]

[0023] Among them, A T is the transmitted signal amplitude, f c is the central slope of the signal, t is the emission time, B is the bandwidth, T m is the frequency modulation period, τ is the time delay, and dτ is the differential of τ.

[0024] Furthermore, the echo signal is expressed as:

[0025]

[0026] Among them, A R is the amplitude of the echo signal, Δt is the time delay of the echo signal relative to the transmitted signal, Δf d is the Doppler frequency shift, and j is the imaginary unit.

[0027] Furthermore, the intermediate frequency signal is expressed as:

[0028] S IF (t) = S T (t)S R (t)≈A T A R exp{j2π[f c Δt+(f1-Δf d )t]}

[0029] in, Represents the frequency of the intermediate frequency signal at time t.

[0030] Furthermore, in the sampling of the intermediate frequency signal, the discrete expression of each sampling point is:

[0031] S I (n, m) = A I exp{j2π[f I (n)-Δf d (m)] / f s}

[0032] (1≤n≤N)(l≤m≤M)

[0033] Among them, N is the number of repetition cycles in each frame, that is, the number of Chirps in the transmitted signal, M is the number of sampling points in one frequency modulation cycle, S I (n,m) is the discrete expression of the mth sampling point in the nth cycle.

[0034] Furthermore, the short-time Fourier transform STFT is expressed as:

[0035]

[0036] Where W(t) is the window function, is the target's motion echo signal within this time range, t is time, and f is the signal frequency.

[0037] Furthermore, the slider mean y mean Expressed as:

[0038]

[0039] Among them, L is the slider size, y i is the energy at the i-th frequency point, and N is the number of repetition cycles.

[0040] Furthermore, the clutter noise matrix Noise obeys Gaussian distribution and is expressed as:

[0041]

[0042] Where x is the energy value of a point in the clutter noise matrix, μ is the mean of the clutter noise matrix, σ is the variance of the clutter noise matrix, and f(x) is the probability of occurrence of the energy value at the noise point.

[0043] Furthermore, the millimeter wave radar is a linear frequency modulation FMCW.

[0044] The beneficial effects of the present invention are:

[0045] Compared with the method of inputting a time-frequency graph into the network for identification, in this application, the matrix containing the time-frequency information after the echo signal is transformed by the Fast Fourier Transform (FFT) is directly input into the network for identification. This application does not ignore the time-frequency information of the target. Different from the method of removing "spots" in the time-frequency graph, this application uses a search boundary algorithm to determine the clutter noise matrix in the time-frequency information matrix, and uses the Gaussian distribution probability to determine the threshold value for clutter elimination, which effectively improves the recognition effect of millimeter-wave radar for fall detection in different clutter and noise environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is the flow chart for this application;

[0047] Figure 2 This is a comparison chart of the target Doppler information effect between the time-frequency information matrix and the time-frequency graph;

[0048] Figure 3 This is a noise filtering effect diagram of fixed threshold noise in a high noise environment;

[0049] Figure 4 This is the effect diagram of clutter noise elimination for this application. DETAILED DESCRIPTION

[0050] It should be noted that, unless there is any conflict, the various embodiments disclosed in this application can be combined with each other.

[0051] Specific implementation method 1: refer to Figure 1 Specifically describing this embodiment, the clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix described in this embodiment includes:

[0052] Stage 1: Time-frequency transformation based on echo information

[0053] The transmitted signal within one frequency modulation cycle of linear frequency modulation (FMCW) millimeter wave radar can be expressed as:

[0054]

[0055] Among them, A T is the transmitted signal amplitude, f c is the central slope of the signal, t is the emission time, B is the bandwidth, T m is the frequency modulation period.

[0056] After being reflected by the target and the environment, the echo signal received by the millimeter-wave radar receiving antenna is:

[0057]

[0058] Among them, A Ris the amplitude of the echo signal, Δt is the time delay of the echo signal relative to the transmitted signal, Δf d is the Doppler shift.

[0059] Step 1: Calculate the echo signal range Doppler map

[0060] The expression of the intermediate frequency signal obtained after mixing the transmitted signal and the echo signal is:

[0061] S IF (t) = S T (t)S R (t)≈A T A R exp{j2π[f c Δt+(f1-Δf d )t]} (3)

[0062] in, represents the frequency of the intermediate frequency signal at time t. The distance of the target at this time can be expressed as:

[0063]

[0064] Where c represents the speed of light, which is 3×10 8 m / s.

[0065] When the intermediate frequency signal is sampled, the discrete expression of the mth sampling point in its nth cycle is:

[0066]

[0067] Where N is the number of repetition periods in each frame, that is, the number of chirps in the transmitted signal, and M is the number of sampling points in one frequency modulation period.

[0068] At this time, if the M sampled discrete intermediate frequency signals with N linear frequency modulation cycles in each frame are first windowed and filtered to remove the high-resolution interference components, and then the range-dimensional FFT and Doppler-dimensional FFT are performed on them respectively, the range-Doppler map of the frame data can be obtained.

[0069] Step 2: Calculate the time-frequency information matrix

[0070] The time domain signal is framed, that is, the range Doppler map is calculated for each frame of data, and the range of the target is determined through the spectrum peak search algorithm. At the same time, an appropriate number of range gates is set to ensure that the micro-Doppler information of the target's limbs excluding the torso can be obtained (the Doppler component mainly consists of the torso Doppler component and the limb Doppler component). The echo energy within this range is summed and averaged, which is used as the Doppler information of the frame.

[0071] In order to process the time-frequency information of the continuously input time signal, the short-time Fourier transform (STFT) is used to divide the time domain of the entire signal into time series of equal length, and the FFT operation is performed on each small time series. Because the signal is considered to be of finite length in a short time range, the STFT can effectively extract the changes in micro-Doppler information within each time period. Its expression is:

[0072]

[0073] Where W(t) is the window function, is the target's motion echo signal within this time range.

[0074] According to the above steps, the target time-frequency information in the millimeter-wave radar echo information can be calculated.

[0075] Phase 2: Adaptive clutter removal based on time-frequency information matrix

[0076] Step 1: Determine the clutter noise matrix

[0077] In the time-frequency information matrix, the horizontal direction is usually the number of repeated cycles in each frame, which represents the size of the target Doppler frequency component, and the vertical direction is the number of frames, which represents time. To determine the clutter noise matrix, it is necessary to search the boundary points of the target Doppler information frame by frame in the time-frequency information matrix, that is, to search its motion feature envelope row by row. However, the energy of the target motion information and clutter noise information changes dramatically, which is not conducive to the selection of boundary points. Therefore, a 1×5 slider is set for each row, and the data in the slider is averaged and smoothed to obtain y mean , to better reflect the data energy trend:

[0078]

[0079] Among them, L is the slider size, y i is the energy at each frequency point, and N is the number of repetition cycles.

[0080] The mean value y of each slider after the above smoothing process mean The resulting matrix, 1×(NL), represents the smoothed data for that row of data. To determine the motion information boundary for that frame of data, we search for minimum points less than the median of that row of data, starting from the center of the smoothed data row (i.e., zero frequency) and moving toward the left and right ends, recording the coordinates of the current left and right endpoints. When the smoothed data is at the boundary between the target motion information and the noise, its value should gradually decrease toward both sides until the mean value within the slider is equal to the noise, causing fluctuations in the data. This is also a function of the slider smoothing data.

[0081] Repeat the above operation to obtain the left and right boundary points of each frame motion information of the entire time-frequency information matrix, and find the leftmost and rightmost boundary points yleft and y right , then the clutter noise matrix Noise can be expressed as:

[0082] Noise=[a1,a2,…,a left ,a right ,…,a N-1 ,a N ] (8)

[0083] Among them, a i Represents the i-th column data of the time-frequency information matrix.

[0084] Step 2: Determine the noise elimination threshold

[0085] After multiple verifications of data fitting, it was found that clutter noise in different environments basically obeys Gaussian distribution, with only different means and variances:

[0086]

[0087] Where x is the energy value of a point in the clutter noise matrix, μ is the mean of the clutter noise matrix, σ is the variance of the clutter noise matrix, and f(x) is the probability of occurrence of the energy value at the noise point.

[0088] Take the maximum value in the obtained clutter noise matrix and calculate its probability of occurrence p in the Gaussian distribution. Since this matrix is ​​the clutter noise sample of the time-frequency information matrix, the actual clutter noise value should be greater than or equal to the maximum value in the sample. Therefore, the noise energy value corresponding to the Gaussian distribution probability p / 10 is taken as the clutter noise elimination threshold.

[0089] Step 3: Use the Shapiro-Wilk normality test to detect the rationality of the selection of the clutter noise threshold.

[0090] The Shapiro-Wilk normality test is a commonly used Gaussian distribution test method. In the present invention, the correctness of the clutter noise matrix selection significantly affects the selection of the clutter noise threshold. If the clutter noise matrix is ​​too large and contains some target motion information, it will not conform to the Gaussian distribution. At this time, if the threshold value is still selected in the above manner, the threshold value will be too large, which will cause the loss of some target motion information while filtering out clutter, causing certain interference for subsequent detection and identification. Compared with other normality test methods, the Shapiro-Wilk normality test can effectively test whether the data conforms to the Gaussian distribution regardless of the number of samples selected for the clutter noise matrix.

[0091] Therefore, the Shapiro-Wilk normality test is used as the test index for the clutter noise removal effect. Check whether the selected clutter noise matrix passes the Gaussian distribution test. If the clutter noise matrix does not obey the Gaussian distribution, it is necessary to determine the leftmost and rightmost boundary points y of the target motion information. left and y right If the selected clutter noise matrix passes the normality test, it means that the selection of the matrix is ​​reasonable, that is, the selection of the clutter noise threshold is also reasonable.

[0092] Example:

[0093] The clutter elimination effect of the present invention is demonstrated by eliminating clutter noise from data of an elderly person falling down and obtaining the target information envelope:

[0094] Set the radar operating parameters: radar starting frequency 60 GHZ, FM bandwidth 2.1 GHZ, frame time 36 ms, number of repetition cycles 255, a total of 120 frames of data.

[0095] Specific process:

[0096] (1) Calculate the 120×255 time-frequency information matrix of the data;

[0097] (2) Smooth the matrix row by row and search for the left and right minimum points from the center to both sides as the boundaries, and finally obtain the leftmost boundary y of the target motion information left is 68, the rightmost boundary y right is 185;

[0098] (3) truncate columns 1-68 and 185-255 of the time-frequency information matrix and concatenate them into a 120 × 139 clutter noise matrix;

[0099] (4) Calculate the mean value 1001.4 and variance 220.5 in the clutter noise matrix, and fit the distribution function of the clutter noise matrix according to the Gaussian distribution probability density function;

[0100] (5) At a significance level of 0.05, the normality of the clutter noise matrix was tested using the Shapiro-Wilk normality test, and the probability that it conforms to the Gaussian distribution is 0.99, indicating that the selection of the clutter noise matrix is ​​reasonable;

[0101] (6) Calculate the maximum value of 2316.5 in the clutter noise matrix and calculate the Gaussian distribution probability p at this value as 8×10 -4 ;

[0102] (7) Take p / 10 as 8×10 -5 , the energy value corresponding to the Gaussian distribution under probability is calculated to be 2564.7;

[0103] (8) With the energy value of 2564.7 as the threshold, all energy values ​​below this value in the time-frequency information matrix are set to 0;

[0104] (9) At this point, correct adaptation eliminates clutter and noise in the environment and preserves the Doppler information of the target;

[0105] (10) The denoised time-frequency information matrix is ​​directly input into the network for recognition, and the correct recognition result is obtained: fall;

[0106] Effect comparison: Compared with the previous method of setting a single threshold to eliminate clutter noise, the present invention is suitable for indoor fall detection tasks in different clutter noise environments; it can accurately eliminate the interference of clutter noise while retaining the true target motion Doppler information.

[0107] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solutions of the present invention and cannot be used to limit the scope of protection. Any minor changes made based on the claims and description of the present invention shall still fall within the scope of protection of the present invention.

Claims

1. A clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix, characterized by The following steps are involved: Step 1: Obtain the transmission signal of the millimeter wave radar within one frequency modulation cycle and the echo signal containing the human target; Step 2: Mix the transmitted signal and the echo signal to obtain an intermediate frequency signal, and sample the intermediate frequency signal to obtain a discrete intermediate frequency signal; Step 3: Perform a short-time Fourier transform on the discrete intermediate frequency signal to obtain a range Doppler map for each frame of data. A spectral peak search is performed on the range Doppler map for each frame of data to obtain the target's range. The number of range gates is set to 9 to obtain micro-Doppler information for the target's limbs, excluding the torso. The echo energy within the range gate number is summed and averaged to obtain the Doppler information for that frame. Step 4: Divide the time domain of the intermediate frequency signal into time series of equal length through short-time Fourier transform, and perform fast Fourier transform on each time series to obtain the range-Doppler map of the frame data. Then, combine the range-Doppler map of the frame data with the Doppler information of the frame to obtain the time-frequency information matrix of the intermediate frequency signal segment. Step 5: Smooth each row of data in the time-frequency information matrix to obtain the smoothed data of the row of data, specifically: In each row of data, select data in sequence starting from the first data until the fifth data from the last is selected. For each selected data, obtain the data and the following four data, and then calculate the average of the five data to obtain the mean. , all the means in each row of data After combination, we get The matrix, The matrix is ​​the smoothing data of the row data, is the slider size, is the number of repetition cycles; Step 6: From the center of each row of smoothed data, that is, the zero frequency, search for the minimum value point that is smaller than the median of the row of data to the left and right ends respectively, and record the coordinates of the minimum value point on the left and right ends; Step 7: Repeat step 6 to obtain the coordinates of the minimum points on the left and right ends of each frame of the time-frequency information matrix, and obtain the leftmost boundary point and the rightmost boundary point , and then get the clutter noise matrix , clutter noise matrix Expressed as: in, is the first Column data, is the first Column data, is the first Column data, clutter noise matrix Obey Gaussian distribution; Get the clutter noise matrix The maximum value of the data and calculate the probability of the maximum value appearing in the Gaussian distribution , take the Gaussian distribution probability The corresponding noise energy value is used as the elimination threshold of the clutter noise, thereby completing the clutter elimination; The steps for obtaining the range Doppler map of each frame of data are as follows: The signal obtained by the framing operation is sequentially subjected to window filtering, range-dimensional fast Fourier transform, and Doppler-dimensional fast Fourier transform to obtain a range-Doppler map for each frame of data; The slider means Expressed as: in, is the slider size, is the energy at the i-th frequency point, is the number of repetition cycles.

2. The clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix according to claim 1 is characterized in that The transmission signal is expressed as: in, is the transmitted signal amplitude, is the central slope of the signal, is the launch time, is the bandwidth, is the frequency modulation period, τ is the time delay, dτ is the differential of τ, and j is the imaginary unit.

3. The clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix according to claim 2 is characterized in that The echo signal is expressed as: in, is the echo signal amplitude, is the time delay of the echo signal relative to the transmitted signal, is the Doppler shift, and j is the imaginary unit.

4. The clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix according to claim 3 is characterized in that The intermediate frequency signal is expressed as: in, Indicates The frequency of the intermediate frequency signal at this moment.

5. The clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix according to claim 4 is characterized in that In the sampling of the intermediate frequency signal, the discrete expression of each sampling point is: in, is the number of repetition cycles in each frame, i.e., the number of quantity, is the number of sampling points in one frequency modulation period, S I (n,m) is the discrete expression of the mth sampling point in the nth cycle.

6. The clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix according to claim 5 is characterized in that The short-time Fourier transform STFT is expressed as: in, is the window function, is the target's motion echo signal within this time range, For time, is the signal frequency, and j is the imaginary unit.

7. The clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix according to claim 1 is characterized in that The clutter noise matrix Obey Gaussian distribution, expressed as: in, is the energy value of a point in the clutter noise matrix, is the mean of the clutter noise matrix, is the variance of the clutter noise matrix, is the probability of occurrence of the energy value at the noise point.

8. The clutter elimination method for millimeter wave radar fall detection based on time-frequency information matrix according to claim 1 is characterized in that The millimeter wave radar is a linear frequency modulation FMCW radar.

Images