Millimeter wave radar-based indoor human target point cloud detection method and system

Abstract

The application discloses a kind of indoor human target point cloud detection method and system based on millimeter wave radar, and millimeter wave radar obtains the point cloud information of target in real time, including the distance gate index of target point, angle gate index, signal-to-noise ratio and background noise.Point cloud detection process includes:1, determine the target point detection threshold of each distance gate based on distance-angle heat map;2, the normalized gradient of two-dimensional heat map is calculated and the threshold is used to obtain target map;3, using closed operation updates target map;4, using scan line seed filling algorithm identifies the background area of target map;5, the information of non-background area point is calculated and used as point cloud output.The method can significantly reduce the interference of large size target itself and strong target, and can detect the morphology of target in heat map completely, and has strong practicability in indoor human target detection.

Description

Indoor Human Target Point Cloud Detection Method and System Based on Millimeter-Wave Radar Technical Field

[0001] This invention belongs to the field of target detection and positioning technology, specifically relating to a method and system for detecting indoor human target point clouds. Background Technology

[0002] In fields such as surveillance, home appliance control, and big data analytics, indoor human target detection has become an increasingly important functional requirement. Using millimeter-wave radar for indoor human target detection is a relatively new method that has emerged in recent years, while traditional methods primarily utilize 10GHz to 24GHz microwave radar, passive infrared (PIR), active infrared (LiDAR, TOF), and optical cameras. Although traditional methods are widely used, they all struggle to simultaneously meet the following requirements: 1. Detecting minute movements; 2. Minimal susceptibility to environmental influences; 3. Determining the direction of target movement; 4. Accurately locating the target in space; 5. Detecting targets even when they are obscured.

[0003] Millimeter-wave radar is a sensor that can provide a three-dimensional dataset of a target's distance, velocity, and angle. Based on the acquired three-dimensional data, millimeter-wave radar has the ability to accurately determine the position and direction of movement of a person. Furthermore, due to the characteristics of millimeter waves—low atmospheric loss, wide bandwidth, and narrow beamwidth—millimeter-wave radar can operate under various environmental conditions. Therefore, using millimeter-wave radar for detection is one of the important future development directions for indoor human target detection.

[0004] Because millimeter-wave radar technology is still in its early stages, the related radar signal processing algorithms are not yet fully mature. For example, a typical algorithm flow currently used for human target detection is as follows:

[0005] 1. Distance-dimensional FFT;

[0006] 2. Eliminate static clutter while preserving moving targets;

[0007] 3. Beamforming to obtain a distance-angle heatmap;

[0008] 4. Perform constant false alarm rate (CFAR) detection on the distance-angle heatmap and output point cloud information;

[0009] 5. Doppler FFT to obtain the velocity information of the point cloud;

[0010] 6. Cluster and track point clouds;

[0011] The CFAR used in step 4 is a primary method in traditional radar signal processing for detecting targets in heatmaps to obtain target point clouds. CFAR treats points near the target point as noise, processes the noise to obtain a threshold, and compares the threshold with the target point to determine whether the target point is a target. CFAR can detect targets that are relatively small, such as those occupying only one point on the heatmap. However, in indoor scenes, targets typically exhibit a near-large and far-small characteristic on the range-angle heatmap. When the target is close to the radar, it appears large on the heatmap, potentially occupying dozens of points. In such cases, when using CFAR, both the noise window and the target point fall within the target, leading to missed detections. Furthermore, in indoor scenes where multiple people are close together, the problem of both the noise window and the target point falling within the target also occurs, resulting in missed detections. Summary of the Invention

[0012] Technical Problem: Addressing the problems existing in current millimeter-wave radar technology for detecting indoor human targets, this invention proposes a method and system for indoor human target point cloud detection based on millimeter-wave radar. This method can significantly reduce interference from large targets themselves and strong targets, while simultaneously detecting the complete morphology of the target in the thermal image.

[0013] Technical solution: This invention proposes an indoor human target point cloud detection method based on millimeter-wave radar, the specific process of which is shown below;

[0014] The method is used to acquire point cloud information of a target in real time, including the target point's distance gate index, angle gate index, signal-to-noise ratio, and background noise. The method includes:

[0015] Step 1, Determine the target point detection threshold

[0016] One cycle of the electromagnetic wave emitted by the millimeter-wave radar is called a chirp. The required intermediate frequency (IF) data is obtained by sampling the mixed signal during a specific time period of the cycle. Range-dimensional FFT and beamforming are performed on the IF data to obtain a range-angle heatmap. The detection threshold of the target point is independently calculated for the heatmap data of each range gate. The detection threshold has upper and lower bounds and is calculated using a three-segment method.

[0017] Step 2, Detect the target

[0018] On the range-angle heatmap, the intensity of the target region exhibits a mountain-like shape, while the intensity of the background region follows a Gaussian distribution. The gradient of amplitude variation is used to determine whether it is a target point. Therefore, the Sobal operator is used to convolve with the range-angle heatmap to obtain the gradient information of the heatmap. Considering that the millimeter-wave radar of the frequency-modulated continuous wave system has high range resolution but low azimuth resolution, it is necessary to calculate the two-dimensional gradient by weighting the azimuth gradient and the range gradient. Then, the two-dimensional gradient is normalized to obtain the normalized gradient. The normalized gradient map is traversed, and the normalized gradient is compared with the detection threshold of the corresponding range gate to determine whether it is a target point, thus obtaining the target map E.

[0019] Step 3. Closing operation

[0020] The target image E obtained in step 2 may have unclosed boundaries or fragmented targets. Therefore, closing operations are used to process the target image, as shown below:

[0021] E′=E·B

[0022] Where E' is the processed target image and B is the convolution kernel for the closing operation. Due to the special nature of the distance-angle heatmap, different convolution kernels are used for the dilation and erosion operations in the closing operation.

[0023] Step 4. Determine the background area

[0024] The target image E' processed in step 3 is filled using the scanline seed filling algorithm to obtain the background area;

[0025] Step 5. Output point cloud

[0026] Step 4 has yielded the background region; the points outside the background region are the target points. Output the relevant information for each target point, including its distance gate index, angle gate index, background noise, and signal-to-noise ratio. The background noise for each target point uses the background noise N of the distance-angle heatmap at distance gate i. i Instead, the signal-to-noise ratio of the target point is obtained by dividing the intensity of the target point in the distance-angle heatmap by the background noise of the target point.

[0027] in,

[0028] Step 1 specifically includes:

[0029] S1.1, Determine the noise threshold N when the distance gate is i based on the background noise. t and target threshold S t As shown below:

[0030] S t =k s N i

[0031] N t =k n N i

[0032] Where, N i This is the background noise at a distance of i from the gate, k s k is the appropriate multiple required for the target threshold. n The appropriate multiple required for the noise threshold;

[0033] S1.2, the maximum value m of distance gate i in the distance-angle heatmap. i When the noise level is less than the threshold, the detection threshold is fixed at the upper bound of the threshold; m i When the noise threshold is greater than the noise threshold but less than the target threshold, the detection threshold decreases linearly from the upper limit to the lower limit as the intensity of the measured point increases; m i When the value exceeds the target threshold, the detection threshold is fixed at the lower bound of the threshold; the calculation formula is shown below.

[0034]

[0035] Where i is the distance gate index, and its value range is i∈[0,1]. max -1] and i∈Z, Z represents the set of integers, I max The number of sampling points in a chirp is raised to the power of 2; t i It is the detection threshold when the distance gate is i; t max It is the upper limit of the threshold; t min It is the lower limit of the threshold.

[0036] Step 2 specifically includes:

[0037] S2.1, the gradient information of the heatmap is obtained by convolving the Sobal operator with the distance-angle heatmap. The calculation formula is as follows:

[0038] G x =S x A

[0039] G y =S y A

[0040] The Sobel operator is divided into Si, which calculates the gradient in the horizontal direction (i.e., the azimuth direction). x Operator and calculation of the gradient S in the perpendicular direction, i.e., the distance direction. y The two operators are shown below:

[0041]

[0042] Where S xCompared to traditional operators, G is optimized for low azimuth resolution in millimeter-wave radar. x This is the azimuth gradient map; G y A is the distance-direction gradient map; A is the distance-angle heatmap.

[0043] S2.2, calculate the two-dimensional gradient by weighting the azimuth gradient and the range gradient, and then divide the two-dimensional gradient by the Sobel operator weights and m. i To obtain the normalized gradient, in order to avoid m i <N t If the time-normalized gradient is too large, during the calculation process, if m i <N t Let m i =N t The formula for calculating the normalized gradient is as follows:

[0044]

[0045] Where j is the angle gate index, and its value range is j∈[0,J]. max -1] and j∈Z, J max g is the number of steering vectors used during beamforming. ij G is the normalized gradient when the distance gate is i and the angle gate is j; xij It is the azimuth direction gradient when the distance gate is i and the angle gate is j; G yij is the distance gradient in the direction of the azimuth and the angular gradient when the distance gate is i and the angular gate is j; k is the weighted value of the azimuth gradient and the distance gradient; abs(X) is a matrix within parentheses, where X is the absolute value of each element; sum(X) is a matrix within parentheses, where X is the sum of each element.

[0046] S2.3, if the normalized gradient of a point is greater than the detection threshold of the corresponding distance gate, then the point is determined to be a target point; otherwise, it is a background point or a point in a flat region inside the target, i.e., a potential target point. The target map can be obtained by traversing each point of the normalized gradient matrix for decision-making. Furthermore, for ease of subsequent processing, it is assumed that there are no target points when the distance gate is 0. The specific calculation process is as follows:

[0047]

[0048] Among them, E ij This is the detection result when the distance gate is i and the angle gate is j. 1 represents the target point, and 0 represents the background point or a flat area point inside the target.

[0049] Step 3 specifically includes:

[0050] S3.1, Initialize E tempIt is a matrix of all zeros with the same number of rows and columns as E.

[0051] S3.2, Exchange E temp And E, in Repeat k times;

[0052] S3.3, Exchange E temp And E, in Repeat k times;

[0053] S3.4, E is output as E';

[0054] in To add matrix elements, Subtract the matrix elements.

[0055] Step 4 specifically includes:

[0056] S4.1 Select a point in the target image that is 0 distance from the gate and has an angle of 0, and store it as the first seed in the seed stack;

[0057] S4.2, Determine if the seed stack is empty;

[0058] If the seed stack is empty, the algorithm terminates.

[0059] If the seed stack is not empty, pop a seed from the top of the stack and extend that point left or right along the j-direction until it reaches the boundary or E'. ij =1, thus obtaining the interval [j l ,j r Points within the interval satisfy E' ij =0, determine the points in this interval as background points, and let E' = 0. ij =2;

[0060] S4.3, If the seed selected in step S4.2 satisfies i>0, then for the interval [j] in row i-1... l ,j r The connectivity within ] satisfies E' ij Push the right boundary of the interval = 0 onto the stack;

[0061] S4.4, if the seed selected in step S4.2 satisfies i max -1, for the interval [j] in row i+1 l ,j r The connectivity within ] satisfies E' ij Push the right boundary of the interval = 0 onto the stack;

[0062] S4.5, jump to step S4.2.

[0063] ​The present invention discloses an indoor human target point cloud detection system based on millimeter-wave radar, comprising a detection threshold calculation module, a target detection module, a closing operation module, a background region determination module, and a point cloud output module connected in sequence.

[0064] The detection threshold calculation module is used to calculate the detection threshold for each distance gate based on the distance-angle heatmap;

[0065] The target detection module is used to calculate the normalized gradient of the distance-angle heatmap and filter target points by a detection threshold;

[0066] The closing operation module is used to close the target boundary and aggregate the target blocks;

[0067] The background region determination module is used to separate background points and potential target points based on the connectivity characteristics of the background region.

[0068] The point cloud output module is used to count target points and output relevant information.

[0069] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0070] 1. Less affected by strong target points. In CFAR detection, it cannot be guaranteed that the points selected by the noise window are actually noise. If a target point falls within the noise window, it is easy to miss the detection. This invention detects target edges and areas with drastic intensity changes by using a normalized gradient method, then uses the connectivity properties of the background region to obtain the background region, and finally inversely selects the target region. Using gradients can avoid interference from other nearby targets and interference from large targets on their own internal points, thereby reducing the probability of missed detection.

[0071] 2. Capable of acquiring target morphology. This invention first obtains the background region, then inversely selects the target points. The target points constitute multiple connected regions, and each connected region is a target. Compared to CFAR detection, which detects each point independently, this invention can acquire the morphological information of the target. Attached Figure Description

[0072] Figure 1 is a flowchart of the human target point cloud detection method disclosed in this invention;

[0073] Figure 2 is a distance-angle heatmap in the embodiment;

[0074] Figure 3 shows a partial target image after threshold determination in the embodiment;

[0075] Figure 4 shows a local target image after the closing operation in the embodiment;

[0076] Figure 5 shows a comparison of algorithm performance in the embodiments; where (a) is the target area map, (b) is the CASO-CFAR detection result map, and (c) is the detection result map of the present invention.

[0077] Figure 6 is a schematic diagram of the composition of the human target point cloud detection system disclosed in this invention. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this invention clearer, specific embodiments of this invention will be described below in conjunction with the accompanying drawings.

[0079] Example:

[0080] This invention discloses an indoor human target point cloud detection method based on millimeter-wave radar. The method is used to acquire the target's point cloud information in real time, including the target point's distance index, angle index, signal-to-noise ratio, and background noise. As shown in Figure 1, the method includes the following steps:

[0081] (1) Determine the target point detection threshold

[0082] Data on a single person moving indoors was collected using the TI IWR1642BOOST radar. This radar is a 77GHz millimeter-wave radar with a modulation frequency of 100MHz / µs, and one cycle of the emitted electromagnetic wave is called a chirp. The number of sampling points in a chirp is configured to be 256, and the beamforming steering vector is configured to be 187 points ranging from -70° to 70°. After performing range-dimensional FFT and beamforming on the data, a range-angle heatmap is obtained, and one frame is shown in Figure 2. The detection threshold for each range-gated heatmap is calculated independently. The detection threshold has upper and lower bounds, and a three-segment method is used to calculate the detection threshold.

[0083] (1.1) Determine the noise threshold N when the distance gate is i based on the background noise. t and target threshold S t As shown below.

[0084] S t =k s N i

[0085] N t =k n N i

[0086] Where, N i This is the background noise when the distance gate is i. In this embodiment, it is approximated by the minimum intensity of distance gate i in the distance-angle heatmap; k s In this embodiment, k is a suitable multiple required for the target threshold. s =150; k n In this embodiment, k is a suitable multiple required for the noise threshold. n =30.

[0087] (1.2) The maximum value m of distance gate i in the distance-angle heatmap i When the noise level is less than the threshold, the detection threshold is fixed at the upper bound of the threshold; m i When the noise threshold is greater than the noise threshold but less than the target threshold, the detection threshold decreases linearly from the upper limit to the lower limit as the intensity of the measured point increases; m i When the value exceeds the target threshold, the detection threshold is fixed at the lower bound of the threshold. The calculation formula is shown below.

[0088]

[0089] Where i is the distance gate index, and its value range is i∈[0,1]. max -1] and i∈Z, Z represents the set of integers, I max The value is obtained by raising the number of sampling points in a chirp to the power of 2, which is 256 in this embodiment; t i It is the detection threshold when the distance gate is i; m i It is the maximum value in the distance-angle heatmap when the distance gate is i; t max It is the upper limit of the threshold; t min This is the lower bound of the threshold. The upper and lower bounds of the detection threshold can be configured according to the actual scenario; in this embodiment, they are 0.4 and 0.2, respectively.

[0090] Since we are not interested in targets at very close range (less than 0.5 meters from the radar) and targets at very long range (close to the radar's maximum detection range), this example ignores the threshold calculation for range gates less than 15 and greater than 240 in order to reduce processing time.

[0091] (2) Detection target

[0092] As shown in Figure 2, the intensity of the target area exhibits a mountain-like shape, while the intensity of the background area follows a Gaussian distribution. Therefore, the amplitude variation (gradient) can be used to determine whether it is a target point. Thus, the Sobal operator is used to convolve with the range-angle heatmap to obtain the gradient information of the heatmap. Considering that the millimeter-wave radar of the frequency-modulated continuous wave system has high range resolution but low azimuth resolution, it is necessary to calculate the two-dimensional gradient by weighting the azimuth gradient and the range gradient, and then normalize the two-dimensional gradient to obtain the normalized gradient. By traversing the normalized gradient map and comparing the normalized gradient with the detection threshold of the corresponding range gate to determine whether it is a target point, the target map E can be obtained.

[0093] (2.1) The gradient information of the heatmap is obtained by convolving the Sobal operator with the distance-angle heatmap. The calculation formula is shown below.

[0094] G x =S x *A

[0095] Gy =S y *A

[0096] The Sobel operator is divided into Si, which calculates the gradient in the horizontal direction (azimuth direction). x Operators and S-axis for calculating the gradient in the vertical direction (distance direction) y The two operators are shown below.

[0097]

[0098] Where S x Compared to traditional operators, this one is optimized for low azimuth resolution in millimeter-wave radar. x This is the azimuth gradient map; G y A is the distance-direction gradient map; A is the distance-angle heatmap.

[0099] (2.2) Calculate the two-dimensional gradient by weighting the azimuth gradient and the range gradient, and then divide the two-dimensional gradient by the Sobel operator weights and m. i Obtain the normalized gradient. To avoid m i <N t If the time-normalized gradient is too large, during the calculation process, if m i <N t Let m i =N t The formula for calculating the normalized gradient is as follows.

[0100]

[0101] Where j is the angle gate index, and its value range is j∈[0,J]. max -1] and j∈Z, J max This refers to the number of steering vectors used during beamforming; in this embodiment, it is 187. ij G is the normalized gradient when the distance gate is i and the angle gate is j; xij It is the azimuth direction gradient when the distance gate is i and the angle gate is j; G yij is the distance direction gradient when the distance gate is i and the angle gate is j; k is the weighted value of the azimuth direction gradient and the distance direction gradient, which is 1.2 in this embodiment; abs(X) is a matrix within parentheses, and the absolute value of each element of the matrix is ​​taken; sum(X) is a matrix within parentheses, and the sum of each element of the matrix is ​​taken.

[0102] (2.3) If the normalized gradient of a point is greater than the detection threshold of the corresponding distance gate, then the point is determined to be a target point; otherwise, it is a background point or a point in a flat region inside the target (a potential target point). The target map can be obtained by traversing each point of the normalized gradient matrix and making a decision. Since the threshold calculation of distance gates less than 15 and greater than 240 is ignored in step (1), the point corresponding to the distance gate is directly determined as a non-target point when determining the threshold. The specific calculation process of the threshold determination is as follows.

[0103]

[0104] Among them, E ij This is the detection result when the distance gate is i and the angle gate is j. 1 represents a target point, and 0 represents a background point or a point in a flat area inside the target. Figure 3 shows a local target map after threshold judgment, where target points are visible. Comparing it with Figure 2, it can be seen that the target points in Figure 3 are mostly located in the boundary area of ​​the target, and the target points do not form a unified whole.

[0105] (3) Closing operation

[0106] The image obtained in step (2) has unclosed boundaries and target blocks, so the closing operation is used to process the target image as shown below.

[0107] E′=E·B

[0108] Where E' is the processed target image, and B is the convolution kernel for the closing operation. In this embodiment, B has two values ​​depending on the stage: B1 in the dilation stage and B2 in the erosion stage.

[0109] (3.1) Initialize E temp It is a zero matrix with the same number of rows and columns as E.

[0110] (3.2) Exchange E temp And E, in Repeat k times.

[0111] (3.3) Exchange E temp And E, in Repeat k times.

[0112] (3.4) E is output as E'.

[0113] in To add matrix elements, Subtract the matrix elements.

[0114] Tests showed that k=1 or k=2 performed well. Considering the computation time, k=1 in this embodiment. The partial image of the target obtained after the closing operation is shown in Figure 4. It can be seen that the target has been aggregated into a whole and the boundary has been closed, but there are still hollow areas inside.

[0115] (4) Determine the background area

[0116] The target image E' processed in step 3 is filled using the scanline seed filling algorithm to obtain the background area.

[0117] (4.1) Initialize the seed stack and set its size. In this embodiment, the size is set to 30. Select a point in the target image that is 0 distance from the gate and has an angle of 0, and store it in the seed stack as the first seed.

[0118] (4.2) Determine if the seed stack is empty.

[0119] The algorithm terminates if the seed stack is empty.

[0120] If the seed stack is not empty, pop a seed from the top of the stack. Extend that point left or right along the j-direction until it reaches the boundary or E'. ij =1, so the interval [j] can be obtained. l ,j r Points within the interval satisfy E' ij =0, determine the points in this interval as background points, and let E' = 0. ij =2.

[0121] (4.3) If the seed selected in step (4.2) satisfies i>0, then for the interval [j] in row i-1... l ,j r The connectivity within ] satisfies E' ij Push the right boundary of the =0 interval onto the stack. If the stack is full, report an error and terminate the process.

[0122] (4.4) If the seed selected in step (4.2) satisfies i max -1, for the interval [j] in row i+1 l ,j r The connectivity within ] satisfies E' ij Push the right boundary of the =0 interval onto the stack. If the stack is full, report an error and terminate the process.

[0123] (4.5) Jump to step (4.2)

[0124] (5) Output point cloud

[0125] ​Figure 2 shows a target near the gate at a distance of 60° and an angle of 115°. Enlarging this area yields Figure 5(a). Using the CASO-CFAR algorithm to perform constant false alarm rate (CFAR) detection on the heatmap data in Figure 2, with a noise window length of 8, a guard window length of 4, and a threshold of 8, we obtain Figure 5(b). Step (4) has yielded the background area; the non-background points are the target points ultimately determined by this invention, as shown in Figure 5(c). It can be seen that this invention is less affected by strong target points within large targets, and all internal target points are detected; it also accurately obtains the target's morphological information. E' ij Output relevant information for points ≠ 2, including the distance gate index, angle gate index, background noise, and signal-to-noise ratio for each target point. The background noise for each target point is calculated using the background noise N of the distance-angle heatmap at the distance gate i. i Instead, the signal-to-noise ratio of the target point is obtained by dividing the intensity of the target point in the distance-angle heatmap by the background noise of the target point.

[0126] The system for implementing the above-mentioned indoor human target point cloud detection method is shown in Figure 6, including:

[0127] The detection threshold calculation module is used to calculate the detection threshold for each distance gate based on the distance-angle heatmap. The module takes a 2D heatmap as input and outputs an array of thresholds.

[0128] The target detection module calculates the normalized gradient of the distance-angle heatmap and filters target points based on detection thresholds. The module takes a 2D heatmap and a threshold array as input and outputs a target map.

[0129] The closing operation module is used to close the target boundary and aggregate the target blocks. The module input is the target graph, and the module output is the target graph after the closing operation.

[0130] The background region determination module is used to separate background points and potential target points based on the connectivity characteristics of the background region. The module input is the target image after closing operations, and the module output is the background image.

[0131] The point cloud output module is used to count target points and output relevant information. The module input is a background image, and the module output is a point cloud.

Claims

1. A method for indoor human target point cloud detection based on millimeter-wave radar, characterized in that, The method is used to acquire point cloud information of targets in real time, including the range gate index, angle gate index, signal-to-noise ratio, and background noise of the target points. The method includes: Step 1, determining the target point detection threshold. One period of the electromagnetic wave emitted by the millimeter-wave radar is called chirp. During a specific time period of the period, the mixed signal is sampled to obtain the required intermediate frequency data. Range-dimensional FFT and beamforming are performed on the intermediate frequency data to obtain a range-angle heatmap. The target point detection threshold is independently calculated for each range gate heatmap data. The detection threshold has upper and lower bounds and is calculated using a three-segment method. Step 2, the intensity of the target region on the range-angle heatmap shows a peak shape, while the intensity of the background region follows a Gaussian distribution. The gradient of amplitude change is used to determine whether it is a target point. Therefore, the Sobal operator is used to convolve with the range-angle heatmap to obtain the gradient information of the heatmap. Considering that the millimeter-wave radar of the frequency-modulated continuous wave system has high range resolution but low azimuth resolution, it is necessary to calculate the two-dimensional gradient by weighting the azimuth gradient and the range gradient. The gradient is then normalized to obtain the normalized gradient. The normalized gradient map is traversed, and the normalized gradient is compared with the detection threshold of the corresponding distance gate to determine whether it is a target point, thus obtaining the target map E. Step 3. Closing operation: The target map E obtained in step 2 may have unclosed boundaries and target blocks, so the closing operation is used to process the target map as follows: E′=E·B where E′ is the processed target map, and B is the convolution kernel of the closing operation. Due to the special nature of the distance-angle heatmap, different convolution kernels are used in the dilation and erosion of the closing operation. Step 4. Determine the background region: The target map E′ processed in step 3 is filled with the scan line seed filling algorithm to obtain the background region. Step 5. Output point cloud: The background region has been obtained in step 4. The non-background region points are the target points. The relevant information of the target points is output. The output information includes the distance gate index, angle gate index, background noise and signal-to-noise ratio of each target point. The background noise of the target point is the background noise N of the distance gate i of the distance-angle heatmap where the target point is located. i Instead, the signal-to-noise ratio of the target point is obtained by dividing the intensity of the target point in the distance-angle heatmap by the background noise of the target point.

2. The indoor human target point cloud detection method based on millimeter-wave radar according to claim 1, characterized in that, Step 1 specifically includes: S1.1, determining the noise threshold N when the distance gate is i based on the background noise. t and target threshold S t As shown below: S t =k s N i N t =k n N i Where, N i This is the background noise at a distance of i from the gate, k s k is the appropriate multiple required for the target threshold. n The appropriate multiple required for the noise threshold; S1.2, the maximum value m of the distance gate i in the distance-angle heatmap. i When the noise level is less than the threshold, the detection threshold is fixed at the upper bound of the threshold; m i When the noise threshold is greater than the noise threshold but less than the target threshold, the detection threshold decreases linearly from the upper limit to the lower limit as the intensity of the measured point increases; m i When the value exceeds the target threshold, the detection threshold is fixed at the lower bound of the threshold; the calculation formula is shown below. Where i is the distance gate index, and its value range is i∈[0,1]. max -1] and i∈Z, Z represents the set of integers, I max The number of sampling points in a chirp is raised to the power of 2; t i It is the detection threshold when the distance gate is i; t max It is the upper limit of the threshold; t min It is the lower limit of the threshold.

3. The indoor human target point cloud detection method based on millimeter-wave radar according to claim 1, characterized in that, Step 2 specifically includes: S2.1, using the Sobal operator to convolve with the distance-angle heatmap to obtain the gradient information of the heatmap, the calculation formula is as follows: G x =S x AG y =S y The ASobel operator is divided into S, which calculates the gradient in the horizontal direction, i.e., the azimuth direction. x Operator and calculation of the gradient S in the perpendicular direction, i.e., the distance direction. y The two operators are shown below: Where S x Compared to traditional operators, G is optimized for low azimuth resolution in millimeter-wave radar. x This is the azimuth gradient map; G y S2.2 is the range-direction gradient map; A is the range-angle heatmap; S2.2 calculates the two-dimensional gradient by weighting the azimuth and range-direction gradients, and then divides the two-dimensional gradient by the Sobel operator weights and m. i To obtain the normalized gradient, in order to avoid m i <N t If the time-normalized gradient is too large, during the calculation process, if m i <N t Let m i =N t； The formula for calculating the normalized gradient is as follows: Where j is the angle gate index, and its value range is j∈[0,J]. max -1] and j∈Z, J max g is the number of steering vectors used during beamforming. ij G is the normalized gradient when the distance gate is i and the angle gate is j; xij It is the azimuth direction gradient when the distance gate is i and the angle gate is j; G yij `x` represents the distance gradient in the azimuth direction when the distance gate is `i` and the angle gate is `j`; `k` is the weighted sum of the azimuth gradient and the distance gradient; `abs(X)` is a matrix where `X` is the absolute value of each element; `sum(X)` is a matrix where `X` is the sum of each element; S2.3, if the normalized gradient of a point is greater than the detection threshold of the corresponding distance gate, then the point is determined to be a target point; otherwise, it is a background point or a point in a flat area inside the target, i.e., a potential target point; by traversing each point of the normalized gradient matrix and making a decision, the target map can be obtained; in addition, for the convenience of subsequent processing, it is assumed that there are no target points when the distance gate is 0; the specific calculation process is as follows: Among them, E ij This is the detection result when the distance gate is i and the angle gate is j. 1 represents the target point, and 0 represents the background point or a flat area point inside the target.

4. The indoor human target point cloud detection method based on millimeter-wave radar according to claim 1, characterized in that, Step 3 specifically includes: S3.1, initializing E temp For a matrix with the same number of rows and columns as E, all zeros, in S3.2, swap E. temp And E, in Repeat k times; S3.3, swap E temp And E, in Repeat k times; S3.4, E is output as E'; where To add matrix elements, Subtract the matrix elements.

5. The indoor human target point cloud detection method based on millimeter-wave radar according to claim 1, characterized in that, Step 4 specifically includes: S4.1, selecting a point in the target image that is 0 units away from the gate and has an angle of 0 units, and storing it as the first seed in the seed stack; S4.2, determining whether the seed stack is empty; if the seed stack is empty, the algorithm ends; if the seed stack is not empty, taking a seed from the top of the stack, extending that point left or right along the j direction until it reaches the boundary or E'. ij =1, thus obtaining the interval [j l ,j r Points within the interval satisfy E' ij =0, determine the points in this interval as background points, and let E' ij =2; S4.3, If the seed selected in step S4.2 satisfies i>0, for the interval [j in row i-1] l ,j r The connectivity within ] satisfies E' ij Push the right boundary of the interval = 0 onto the stack; S4.4, if the seed selected in step S4.2 satisfies i max -1, for the interval [j] in row i+1 l ,j r The connectivity within ] satisfies E' ij Push the right boundary of the =0 interval onto the stack; S4.5, jump to step S4.2.​ 6. A system for implementing the indoor human target point cloud detection method based on millimeter-wave radar as described in claim 1, characterized in that, It includes a detection threshold calculation module, a target detection module, a closing operation module, a background region determination module, and a point cloud output module connected in sequence: the detection threshold calculation module is used to calculate the detection threshold of each distance gate based on the distance-angle heatmap; The target detection module is used to calculate the normalized gradient of the distance-angle heatmap and filter target points by a detection threshold; The closing operation module is used to close the target boundary and aggregate the segmented targets; the background region determination module is used to separate background points and potential target points based on the connectivity characteristics of the background region. The point cloud output module is used to count target points and output relevant information.

Images