A scene monitoring radar target extraction method based on video analysis

Abstract

The application discloses a scene monitoring radar target extraction method based on video analysis, comprising the following steps: obtaining target echo information of the i-th frame after radar front-end processing; performing CFAR processing on the target echo information of the i-th frame; marking the threshold EP point after CFAR processing on the two-dimensional image constituted by the target echo information of the i-th frame; extracting the contour information of the target on the two-dimensional image of the i-th frame, and numbering the target according to the distance and direction values of the target; obtaining the two-dimensional image of the i+1-th frame, repeating the above steps, and dynamically matching and updating the target in combination with the target information obtained from the historical frame image. The technical scheme provided by the application can improve the accuracy of target extraction and reduce the risk of splitting into multiple targets during target extraction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method for target extraction from scene surveillance radar based on video analysis. Background Technology

[0002] With the rapid development of my country's air transport industry, the number of aircraft flights at airports is increasing daily, and more and more airports are choosing to install surface surveillance radar systems. Due to the complex airport environment, when aircraft are turning, taxiing, or stationary, using traditional target extraction methods can easily lead to the separation of the aircraft fuselage and wings, affecting the accuracy of point convergence, resulting in unstable target tracking, increasing the workload of air traffic controllers, and ultimately affecting the safety of surface aviation.

[0003] In view of the shortcomings of the existing technology, the present invention provides a new technical means for target extraction, which improves the completeness of target extraction. Summary of the Invention

[0004] The purpose of this invention is to provide a target extraction method for scene surveillance radar based on video analysis, which solves the problem that traditional target extraction methods easily lead to the separation of the aircraft fuselage and wings when extracting targets, affecting the accuracy of point convergence and thus causing unstable target tracking.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A target extraction method for scene surveillance radar based on video analysis includes the following steps:

[0007] S1. Obtain the target echo information of the i-th frame after processing by the radar front end. Where i represents the number of frames scanned by the radar, i = 1, 2, 3...n.

[0008] S2. Perform CFAR processing on the target echo information of the i-th frame.

[0009] S3. Obtain the threshold EP point of the i-th frame after CFAR processing, and construct a two-dimensional image by superimposing the target echo information marked by the EP point.

[0010] S4. Extract the target contours from the two-dimensional image of the i-th frame to obtain the contour information of each target in the image.

[0011] S5. During the radar scanning of the next frame, the scene surveillance radar acquires the target echo information of the next frame after processing by the radar front end, and performs dynamic matching and updating of the target in accordance with the processing method of steps S2-S4, combined with the target information acquired in the i-th frame image, to obtain the contour information of the same target in the next frame.

[0012] As a further aspect of the present invention: in step S1, the target echo information includes:

[0013] The target's azimuth is the azimuth angle detected by the radar.

[0014] The target range library consists of range cells containing the distances of targets detected by radar relative to the radar's position. A range cell is a unit that quantifies the range of radar detection distances.

[0015] The target's Doppler velocity channel is the velocity unit containing the target's velocity relative to the radar, as detected by the radar. The velocity unit quantifies the range of velocities detected by the radar.

[0016] The target's timestamp is the time when the radar detected the target.

[0017] The target's amplitude value is the energy value of the target detected by the radar.

[0018] As a further aspect of the present invention: In step S2, the CFAR processing includes: sequentially comparing the target amplitude value in the target echo information of the i-th frame with the clutter amplitude threshold; if the target amplitude value in the target echo information is less than the clutter amplitude threshold, then the target echo information is filtered out as clutter to obtain the threshold EP point information of the i-th frame.

[0019] As a further aspect of the present invention: In step S3, the target's distance database information is used as the y-axis of the two-dimensional image, and the target's orientation information is used as the x-axis of the two-dimensional image. Based on the target's distance database and orientation in the target echo information of the i-th frame, the corresponding pixels on the two-dimensional image are found. The target's amplitude value is set to 0 as the second color value of the G channel, the first color value of the B channel, and the third color value of the R channel for each pixel, thus forming the two-dimensional color image of the i-th frame. The corresponding pixels on the two-dimensional image are found based on the distance and orientation of the EP point and marked with a crosshair. The RGB pixel values ​​of the marked pixels are all 255. The marked image is denoted as L.

[0020] As a further aspect of the present invention: Step S4, extracting the target contour from the two-dimensional image of the i-th frame includes the following specific steps:

[0021] S41. Use the DBSCAN clustering algorithm to cluster the threshold EP points obtained after CFAR processing in terms of range and orientation, and obtain the point clusters formed by the EP points contained in each target.

[0022] S42. Calculate the minimum bounding rectangle of each target point cluster, thereby obtaining the center point position (Cx, Cy) of each target in the two-dimensional image composed of distance and orientation. Set the width and height thresholds of the sub-image, and extract the sub-image of each target centered at (Cx, Cy) in the two-dimensional image.

[0023] S43. Extract the contours of each sub-image containing the target.

[0024] As a further aspect of this invention, the specific method of DBSCAN clustering is as follows: Two thresholds are set: the neighborhood radius Eps and the neighborhood density MinPts. An arbitrary point p to be clustered is selected from the Eps set. If the number of points within the neighborhood radius Eps of this point exceeds MinPts, then point p is marked as a core object. All points density-connected to point p are found, forming a cluster and assigned a new cluster label. If point p is not a core object, the search continues for the next point until all points have been processed.

[0025] As a further aspect of the present invention: the process of step S43 includes:

[0026] S431, perform mean filtering on the image, and then perform mean drift filtering on it to obtain the image to be processed, G1.

[0027] S432, perform multi-scale Canny edge extraction on the image to be processed G1 to obtain a multi-scale binary edge image G2, and extract the edge contours of the multi-scale binary edge image G2 to obtain a rectangular dataset C1 of the outer edge contours;

[0028] S433, the image to be processed G1 is converted into a grayscale image, the grayscale image is converted into a binary image using threshold segmentation technology, the edge contour is extracted from the binary image, and the rectangular dataset C2 of the outer edge contour is obtained.

[0029] S434: Merge the rectangular datasets C1 and C2 with the minimum bounding rectangle of the target point cluster to obtain the final target rectangle outline.

[0030] As a further aspect of the present invention, the dynamic matching and updating process includes: based on the contour information of each target in the i-th frame at the distance and orientation, template matching searches are performed within a certain image range at the same distance and orientation in the two-dimensional image of the (i+1)-th frame overlaid with EP point markers obtained in steps S2 and S3. If the target image is successfully matched, the structural similarity (SSIM) between the two images is calculated. If SSIM is greater than a preset threshold, the target is dynamically matched successfully, steps S41 and S42 are skipped, and the target contour information is extracted according to step S43. If SSIM is less than the preset threshold, the target is dynamically matched unsuccessfully, and the target contour information is extracted according to steps S41-S43.

[0031] As a further aspect of the present invention, the formula for calculating structural similarity (SSIM) is as follows:

[0032]

[0033] In the formula, X is the target image of the i-th frame, Y is the target update image matched by the template in the (i+1)-th frame, and u X u Y δ represents the mean of the X and Y axes of the image. XY δ represents the covariance of images X and Y. X 2 δ Y 2 Let c1 and c2 represent the variances of the images X and Y, respectively, where c1 and c2 are constants.

[0034] The beneficial effects of this invention are:

[0035] (1) The target image contour processing of the present invention combines multi-scale Canny edge extraction and outer edge contour extraction of binarized image after mean drift filtering, and also combines target EP point cluster information. The three information are effectively fused to improve the accuracy of target extraction.

[0036] (2) This invention combines historical frame target information and dynamically matches the target to prevent the target aggregation point from drifting back and forth when the target EP point is small, thus effectively improving the stability of target tracking.

[0037] (3) This invention improves the integrity of target extraction. Through a series of algorithms, it can reduce the risk of the same target splitting into multiple targets during target extraction, and provides a reliable foundation for subsequent point processing and track processing. Attached Figure Description

[0038] The invention will now be further described with reference to the accompanying drawings.

[0039] Figure 1 This is a flowchart of the method of the present invention;

[0040] Figure 2 This is a logic flowchart of the present invention;

[0041] Figure 3 The result of expanding the minimum bounding rectangle of the target point cluster in the i-th frame of this invention within a certain range to form a two-dimensional image;

[0042] Figure 4 The result of mean-off filtering of the target echo information of the i-th frame of this invention;

[0043] Figure 5This is the result of multi-scale Canny edge extraction on the mean-off filtered image of the i-th frame in this invention;

[0044] Figure 6 This is the result image of the target contour extraction in the i-th frame of the present invention;

[0045] Figure 7 This is a diagram showing the result of successful target template matching in the (i+1)th frame of this invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1-2 As shown, the present invention is a target extraction method for scene surveillance radar based on video analysis, including the following steps S1-S5.

[0048] S1. Obtain the target echo information of the i-th frame after radar front-end processing. Here, i represents the number of frames scanned by the radar, i = 1, 2, 3…n. The target echo information includes: target azimuth, target range database, target timestamp, and target amplitude value; wherein, the target azimuth is the azimuth angle detected by the radar; the target range database is the range unit containing the distance of the target relative to the radar position; the range unit quantizes the range of radar detection distances, and the quantized unit is the range unit; the target Doppler velocity channel is the velocity unit containing the velocity of the target relative to the radar; the velocity unit quantizes the range of radar detection velocities, and the quantized unit is the velocity unit. The target timestamp is the time the radar detects the target; the target amplitude value is the energy value of the target detected by the radar; the radar front-end processing is radar signal processing.

[0049] In this embodiment of the invention, during one frame of the scene surveillance radar scan, the radar monitors the azimuth range from 0° to 360°, with true north as 0°, and the azimuth sampling interval is 0.5°, quantizing the detected azimuth into 720 azimuths; the radar detects the distance range from 0 km to 10.95 km, with the detection distance relative to the radar position, and the distance sampling interval is 3.75 m, quantizing the detected distance into 2920 distance databases; the radar detects the speed range from 0 km / h to 463 km / h, quantizing the detected speed into 5 Doppler velocity channels; that is, starting from the 0° azimuth, the radar receives target echo information from the 2920 distance databases at the 0° azimuth, and then sequentially receives target echo information from the 2920 distance databases at the next azimuth according to the 0.5° azimuth sampling interval, with each distance database including 5 Doppler velocity channels.

[0050] S2. Perform CFAR processing on the target echo information of the i-th frame.

[0051] In step S2, the CFAR processing involves sequentially processing the target amplitude value from the target echo information of the i-th frame.

[0052] Compared with the clutter amplitude threshold, if the amplitude value of the target in the echo information is less than the clutter amplitude threshold, the echo information of this target is filtered out as clutter, and the threshold EP point information of the i-th frame is obtained.

[0053] Specifically, a clutter frequency threshold is set for the target echo information processed by the radar front end. The Doppler frequency information of the target in the target echo information sent by the radar front end is statistically analyzed to obtain the clutter frequency threshold. This threshold can detect 80% of the targets. The clutter frequency threshold value is related to the radar installation site and the radar front end processing. It needs to be obtained after the radar is installed and tested on-site to obtain an empirical value. This provides a clutter frequency threshold for step S2. Target echo information with Doppler frequency lower than the clutter frequency threshold is filtered out as clutter to obtain the filtered target echo information.

[0054] Because the same target generates multiple detection results in adjacent range and azimuth, and due to the influence of strong ground features, strong clutter, and weak clutter from similar targets, the target echo information processed by the radar front end includes interference target echo information. The false point filtering process of this invention filters out some interference target echo information to obtain filtered target echo information for subsequent processing.

[0055] S3. Obtain the threshold EP point of the i-th frame after CFAR processing, and construct a two-dimensional image by superimposing the target echo information marked by the EP point.

[0056] In step S3, the distance database information is used as the y-axis of the two-dimensional image, and the orientation information is used as the x-axis of the two-dimensional image. That is, the 2920 distance databases are used as the y-axis of the two-dimensional image, and the 720 orientations are used as the x-axis of the two-dimensional image. According to the target's distance database and orientation in the target echo information of the i-th frame, the corresponding pixel points on the two-dimensional image are found. The target amplitude value is used as the second color value of the G channel, the first color value of the B channel, and the third color value of the R channel of each pixel point and set to 0, thus forming a 2920-row × 720-column two-dimensional color image of the i-th frame. According to the distance and orientation of the EP point, the corresponding pixel points on the two-dimensional image are found and marked with a cross. The RGB pixel values ​​of the marked points are all 255. The marked image is denoted as L.

[0057] S4. Extract the target contours from the two-dimensional image of the i-th frame to obtain the contour information of each target in the image.

[0058] In step S4, the target contour extraction of the two-dimensional image L of the i-th frame includes the following specific steps:

[0059] S41, use the DBSCAN clustering algorithm to cluster the threshold EP points obtained after CFAR processing in terms of range and orientation to obtain point clusters formed by the EP points contained in each target;

[0060] The specific method of DBSCAN clustering is as follows: Based on empirical values, the neighborhood radius Eps and neighborhood density MinPts are set. An arbitrary point p to be clustered is selected from the EP point set. If the number of points contained within the neighborhood radius Eps of this point is more than MinPts, then point p is marked as a core object. All points connected to the density of this point P are found, forming a cluster and a new cluster label is assigned. If point p is not a core object, the search continues for the next point until all points have been processed.

[0061] S42, calculate the minimum bounding rectangle of each target point cluster, thereby obtaining the center point position (Cx, Cy) of each target in the two-dimensional image L composed of distance and orientation; based on empirical values, set the width threshold and height threshold of the sub-image, and extract the sub-image of each target centered at (Cx, Cy) in the image L.

[0062] S43, extract the contours of the sub-images containing each target. Taking one target as an example, such as... Figure 3 As shown, the same principle applies to other objectives, and the specific methods are as follows:

[0063] S431, perform mean filtering on the image, and then perform mean-shift filtering to obtain the image to be processed, such as... Figure 4 As shown;

[0064] S432, for Figure 4Multi-scale Canny edge extraction is performed to obtain a multi-scale binary edge image, such as... Figure 5 As shown; for Figure 5 Perform edge contour extraction to obtain a rectangular dataset C1 of the outer edge contour;

[0065] S433, for Figure 4 The image is converted to grayscale, and then threshold segmentation is used to convert the grayscale image to a binary image. The edge contours of the binary image are extracted to obtain a rectangular dataset C2 of the outer edge contours.

[0066] S434, merge the minimum bounding rectangle of the rectangular datasets C1 and C2 with the target point cluster to obtain the final target rectangle outline, such as... Figure 6 As shown

[0067] S5. During the radar scanning of the next frame, the scene surveillance radar acquires the target echo information of the next frame after processing by the radar front end, and performs dynamic matching and updating of the target in accordance with the processing method of steps S2-S4, combined with the target information acquired in the i-th frame image, to obtain the contour information of the same target in the next frame.

[0068] The dynamic matching update: Based on the contour information of each target in the i-th frame at the distance and orientation, template matching search is performed within a certain image range at the same distance and orientation in the (i+1)-th frame superimposed with EP point markers obtained in steps S2 and S3. If the target image is successfully matched, the structural similarity SSIM of the two images is calculated, as follows:

[0069]

[0070] Where X is the target image in the i-th frame, Y is the target update image matched by the template in the (i+1)-th frame, and u X u Y δ represents the mean of the X and Y axes of the image. XY δ represents the covariance of images X and Y. X 2 δ Y 2 Let c1 and c2 represent the variances of the images X and Y, respectively, where c1 and c2 are constants.

[0071] Furthermore, if SSIM is greater than a certain threshold, it indicates that the target has been dynamically matched successfully, such as... Figure 7 As shown, skip steps S41 and S42, and proceed directly to... Figure 7 Extract the target contour information according to step S43; if SSIM is less than the set threshold, it means that the target dynamic matching has failed, and extract the target contour information according to steps S41-S43.

[0072] The advantages of this invention are:

[0073] (1) The target image contour processing of the present invention combines multi-scale Canny edge extraction and outer edge contour extraction of binarized image after mean drift filtering, and also combines target EP point cluster information. The three information are effectively fused to improve the accuracy of target extraction.

[0074] (2) This invention combines historical frame target information and dynamically matches the target to prevent the target aggregation point from drifting back and forth when the target EP point is small, thus effectively improving the stability of target tracking.

[0075] (3) This invention improves the integrity of target extraction. Through a series of algorithms, it can reduce the risk of the same target splitting into multiple targets during target extraction, and provides a reliable foundation for subsequent point processing and track processing.

[0076] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for target extraction from scene surveillance radar based on video analysis, characterized in that, Includes the following steps: S1. Obtain the target echo information of the i-th frame after radar front-end processing; Where i represents the number of frames scanned by the radar, i = 1, 2, 3…n; S2. Perform CFAR processing on the target echo information of the i-th frame; The CFAR processing procedure includes: sequentially comparing the target amplitude value in the target echo information of the i-th frame with the clutter amplitude threshold; if the target amplitude value in the target echo information is less than the clutter amplitude threshold, then the target echo information is filtered out as clutter to obtain the threshold EP point information of the i-th frame. S3. Obtain the threshold EP points of the i-th frame after CFAR processing, and construct a two-dimensional image by superimposing the target echo information marked by the EP points; specifically including: Using the target's distance information as the y-axis and the target's orientation information as the x-axis, the corresponding pixels in the 2D image are found based on the target's distance information and orientation in the target echo information of the i-th frame. The target's amplitude value is set to 0 as the second color value of the G channel, the first color value of the B channel, and the third color value of the R channel for each pixel, thus forming the 2D color image of the i-th frame. The corresponding pixels in the 2D image are found based on the distance and orientation of the EP point and marked with a crosshair. The RGB pixel values ​​of the marked pixels are all 255. The marked image is denoted as L. S4. Extract the target contours from the 2D image of the i-th frame to obtain the contour information of each target in the image; specifically including: S41. Use the DBSCAN clustering algorithm to cluster the threshold EP points obtained after CFAR processing in terms of range and orientation to obtain point clusters formed by the EP points contained in each target. The specific method of DBSCAN clustering is as follows: Two thresholds are set, namely the neighborhood radius Eps and the neighborhood density MinPts. Randomly select a point p to be clustered in the EP point set. If the neighborhood radius Eps of the point contains more than MinPts points, then point p is marked as a core object. Find all points that are connected to the density of point P, form a cluster and assign a new cluster label. If point p is not a core object, continue to find the next point until all points have been processed. S42. Calculate the minimum bounding rectangle of each target point cluster, thereby obtaining the center point position (Cx, Cy) of each target in the two-dimensional image composed of distance and orientation; set the width threshold and height threshold of the sub-image, and extract the sub-image of each target centered at (Cx, Cy) in the two-dimensional image. S43. Extract the contours of each sub-image containing the target; S5. During the radar scanning of the next frame, the scene surveillance radar acquires the target echo information of the next frame after processing by the radar front end, and according to the processing method of steps S2-S4, combines the target information acquired in the i-th frame image to dynamically match and update the target, and obtain the contour information of the same target in the next frame. The dynamic matching update process includes: based on the contour information of each target in the i-th frame at the distance and orientation, template matching search is performed in the same distance and orientation within a certain image range in the two-dimensional image of the (i+1)-th frame with superimposed EP point markers obtained in steps S2 and S3. If the target image is successfully matched, the structural similarity SSIM of the two images is calculated. If SSIM is greater than a preset threshold, the target is dynamically matched successfully, steps S41 and S42 are skipped, and the target contour information is extracted according to step S43. If SSIM is less than the preset threshold, the target is dynamically matched unsuccessfully, and the target contour information is extracted according to steps S41-S43.

2. The method for target extraction from scene surveillance radar based on video analysis according to claim 1, characterized in that, In step S1, the target echo information includes: The target's azimuth is the angle at which the radar detects the target's azimuth. The target range library is a range unit containing the distances of targets detected by radar relative to the radar position; the range unit is a unit that quantifies the range of radar detection distances. The Doppler velocity channel of the target is the velocity unit in which the target's velocity relative to the radar is detected by the radar; the velocity unit is a unit that quantifies the range of radar-detected velocities. The target's timestamp, which is the time the radar detected the target; and The target's amplitude value is the energy value of the target detected by the radar.

3. The method for target extraction from scene surveillance radar based on video analysis according to claim 1, characterized in that, Step S43 includes the following process: S431, Perform mean filtering on the image, and then perform mean drift filtering on it to obtain the image to be processed G1; S432, perform multi-scale Canny edge extraction on the image G1 to be processed to obtain a multi-scale binary edge image G2, and extract the edge contours from the multi-scale binary edge image G2 to obtain a rectangular dataset C1 of the outer edge contours; S433, the image G1 to be processed is converted into a grayscale image, and a threshold segmentation technique is used to convert the grayscale image into a binary image. The edge contours of the binary image are extracted to obtain a rectangular dataset C2 of the outer edge contours; S434: Merge the rectangular datasets C1 and C2 with the minimum bounding rectangle of the target point cluster to obtain the final target rectangle outline.

4. The method for target extraction from scene surveillance radar based on video analysis according to claim 1, characterized in that, The formula for calculating the structural similarity (SSIM) is as follows: In the formula, X is the target image in the i-th frame, and Y is the target update image matched with the template in the (i+1)-th frame. , This represents the mean of the X and Y axes of the image. Represents the covariance of images X and Y. , Let c1 and c2 represent the variances of the images X and Y, respectively, where c1 and c2 are constants.

Images