A pre-identification method for UAV swarm target detection based on multi-temporal and multi-spatial scales

Abstract

The present invention discloses a pre-detection and recognition method for drone cluster targets based on multiple spatiotemporal scales. The method includes acquiring target echo data and feeding it into a multi-time-scale system. The method further includes setting sliding window parameters for a multi-time-scale constant false alarm detection method and detecting different types of targets in target windows at three time scales. The method further includes setting decision thresholds for three algorithms: dual-parameter constant false alarm, short-time statistical dual-parameter constant false alarm, and long-time statistical dual-parameter constant false alarm, and judging the detection results. The method further includes combining the three decision results to preliminarily identify and distinguish echo data within the detection window. The method further includes matching and searching the preliminarily detected targets with a preset multi-space type database using a nearest-neighbor peak search method, and outputting target classification results. The method further includes iteratively updating and expanding the types of the multi-space type database based on the matching results. The present invention realizes a closed-loop system for multi-target detection and recognition, effectively completing formation recognition of drone clusters.

Description

Technical Field

[0001] The present invention belongs to the field of radar signal processing, and in particular relates to a pre-detection recognition method for UAV cluster targets based on multiple spatiotemporal scales. Background Art

[0002] Nowadays, the detection environment faced by radars is becoming increasingly complex and changeable. The target detection of drone swarms is a multi-target detection problem in a complex environment with strong ground clutter interference. Since strong ground clutter is mostly concentrated near the Doppler zero frequency and overlaps with stationary targets in the Doppler domain, traditional moving target detection technology is difficult to apply. Moreover, the ground clutter intensity in complex terrains such as cities and mountains is usually greater than or much greater than the target reflection intensity. The target can easily be obscured by the clutter background, making traditional CFAR detection technology difficult to apply and unable to accurately remove clutter while retaining the target. In addition, the detection of drone formations is mostly based on distinguishing the number of targets, and there is less research on the formation type. However, in the context of airspace operations, the identification of the formation type of drone swarms is of great significance, which will directly affect subsequent combat strategies.

[0003] Therefore, it is of great significance to study the detection and recognition methods of drone clusters in complex backgrounds. Summary of the Invention

[0004] Purpose of the invention: The present invention aims to solve the problem of detection and identification of drone clusters in complex backgrounds, and provides a drone cluster target pre-detection identification method based on multiple spatiotemporal scales. It adopts a method combining multiple time scales and spatial scales to perform pre-detection identification of drone cluster targets; the multi-time scale system performs preliminary detection of target echoes and filters out some interfering targets and background clutter; the multi-spatial scale system matches the spatial distribution characteristics of the pre-processed target clusters by establishing a typical formation type database, thereby completing the formation identification of drone clusters.

[0005] Technical solution: The present invention provides a method for pre-detection recognition of drone cluster targets based on multiple spatiotemporal scales, comprising the following steps:

[0006] (1) acquiring target echo data and feeding it into a multi-time scale system; the multi-time scale system includes a constant false alarm detection system, a logic decision device, and a feedback system; wherein the constant false alarm detection system includes three detection algorithms: dual-parameter constant false alarm, short-time statistical dual-parameter constant false alarm, and long-time statistical dual-parameter constant false alarm;

[0007] (2) Setting the sliding window parameters of the multi-time-scale constant false alarm detection method and detecting different types of targets in the target windows at three time scales respectively;

[0008] (3) setting the decision thresholds of the three algorithms, namely, dual-parameter constant false alarm, short-time statistical dual-parameter constant false alarm, and long-time statistical dual-parameter constant false alarm, respectively, and making a decision on the detection result of step (2);

[0009] (4) The three time-scale detection algorithms are configured to output their respective judgment results for points in the same target window through a logic judge; the judgment results of the three algorithms are combined to preliminarily identify and distinguish the echo data within the detection window;

[0010] (5) After removing the areas corresponding to some real targets that affect the detection effect in the target window according to the target type, the mean and standard deviation of the remaining resolution units in the target window are calculated and then fed back to the calculation of the mean and standard deviation of the background window of the next frame in the short-time statistical dual-parameter and long-time statistical dual-parameter constant false alarm respectively through the feedback system;

[0011] (6) Establish a multi-space type database of UAV swarm formations, which includes echo simulation images of several typical formation types;

[0012] (7) Using the nearest peak search method, the target initially detected in step (4) is matched with the preset multi-space type database to output the target classification result;

[0013] (8) Iteratively update and expand the multi-space type database based on the matching results of step (7).

[0014] Furthermore, the echo data in step (1) includes moving targets, stationary targets, interfering targets and clutter.

[0015] Furthermore, the sliding window parameters in step (2) include: the mean of the target window, the mean and standard deviation of the background window in the current frame; the mean of the target window, the mean and standard deviation of the background window under a short time scale; and the mean of the target window, the mean and standard deviation of the background window under a long time scale.

[0016] Furthermore, the implementation process of step (3) is as follows:

[0017] For the DP-CFAR detection algorithm, the background window mean and standard deviation calculation method of the threshold value is:

[0018]

[0019]

[0020] Among them, n is the total number of points in the background window, x i is the numerical value of the i-th point, μ b is the background window mean, σ b is the standard deviation of the background window;

[0021] The mean of the target window under DP-CFAR detection is subtracted from the mean of the background window and divided by the standard deviation of the background window. The result is compared with the nominal factor to determine whether the point in the target window is the target point:

[0022]

[0023] Among them, μ c is the mean of the target window; μ b is the background window mean; σ b is the standard deviation of the background window; T is the nominal factor; if the above formula is satisfied, the output is 1 when it indicates that there is a target, otherwise the output is 0 when it indicates that there is no target;

[0024] For the SSDP-CFAR and LSDP-CFAR detection algorithms, the mean and standard deviation used for judgment are directly used for DP-CFAR detection in the first frame of image data. Starting from the second frame, the mean and standard deviation at the short time scale and long time scale need to be taken into account respectively. The mathematical model of the background window mean and standard deviation used to calculate the threshold value in the p-th frame is as follows:

[0025]

[0026]

[0027] in, is the mean value of the target window in the previous p frames (when calculating a long time scale, p tends to infinity); μ' b To calculate the background window mean of the threshold value at short time scale and long time scale respectively; is the standard deviation of the target window of the previous p frames (when calculating long time scales, p tends to infinity); σ' b To calculate the background window standard deviation of the threshold value at short time scale and long time scale respectively; k is a weight ranging from 0.7 to 0.99.

[0028] Furthermore, the implementation process of step (6) is as follows:

[0029] Extract data from multiple scanning cycles of the actual collected radar echo data and normalize the extracted echo data:

[0030]

[0031]

[0032] Among them, V imax The peak value in the echo data collected in each scanning cycle; For each cycle, Vimax Calculate the average value; V%(t) is the normalized echo data;

[0033] According to the radar equation, the echo strength at different times is calculated for the normalized echo data. Assuming that the distance between the target and the radar is R, the received power of the radar antenna cross section under ideal conditions is:

[0034]

[0035] Among them, P t is the peak power of the radar transmitted pulse, G is the antenna gain, and σ is the average radar cross-sectional area of the target; the target echo intensity is obtained as follows:

[0036] P r =P r0 F 4 L a

[0037] Among them, F is the propagation factor, L a Atmospheric attenuation.

[0038] Furthermore, the implementation process of step (7) is as follows:

[0039] The nearest neighbor search algorithm is used to search for the echo simulation graphs that are closest to the echo peak number in sequence, and the Euclidean distance between each peak point of the echo and each peak point of the simulation graph is calculated. The calculation formula is as follows:

[0040]

[0041] Among them, x i is the peak amplitude of each echo, y i is the peak amplitude of the simulation graph, i is the number of peaks;

[0042] Standardize the Euclidean distance, standardize each calculated value to the interval [0,1], find the echo simulation image corresponding to the minimum value, which is the optimal type of the target, and output the target recognition result.

[0043] Furthermore, the implementation process of step (8) is as follows:

[0044] The matched echo spatial distribution can be used as the characteristic information of the real target to enrich the type database, and the unmatched echo types can also be fed back to the multi-spatial type database as prior information to expand the echo types of the database.

[0045] Beneficial effects: Compared with the existing technology, the beneficial effects of the present invention are: 1. Most existing methods use echo information in the time domain or spatial domain to detect drone cluster targets. The present invention makes full use of time domain and spatial domain information to extend the target detection method based on multiple time scales to the spatial scale; 2. Most existing methods perform a series of detection and identification processing on target echoes. The present invention efficiently estimates the spatial distribution of detected targets based on a priori knowledge base of drone cluster formation shapes and numbers; 3. The present invention realizes a closed-loop system for multi-target detection and identification. The matching results of target echoes can be fed back as prior information to the knowledge base of drone cluster formation shapes and numbers for subsequent matching, thereby forming a self-learning and self-improvement closed-loop cognitive system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described in detail below with reference to the accompanying drawings.

[0048] The present invention provides a pre-detection identification method for drone swarm targets based on multiple spatiotemporal scales. This method uses a combination of multiple time and spatial scales to identify drone swarm targets before detection. The multi-time scale system performs preliminary detection of target echoes and filters out some interfering targets and background clutter. The multi-spatial scale system then establishes a typical formation type database to match the spatial distribution characteristics of the pre-processed target clusters, thereby completing drone swarm formation identification. Specifically, the method includes the following steps:

[0049] Step 1: Obtain target echo data and send it to the multi-time scale system.

[0050] The echo data contains moving targets, stationary targets, interfering targets, and clutter. The multi-time-scale system consists of three main components: a constant false alarm (CFAR) detection system, a logical decision maker, and a feedback system. The CFAR detection system includes three detection algorithms: dual-parameter (DP-CFAR), short-time statistical dual-parameter (SSDP-CFAR), and long-time statistical dual-parameter (LSDP-CFAR).

[0051] Step 2: Set the sliding window parameters of the multi-time scale (MTS) constant false alarm detection method, and detect different types of targets in the target windows at three time scales respectively.

[0052] The parameter setting steps are as follows:

[0053] 1) Calculate the mean of the target window of the current frame, the mean and standard deviation of the background window;

[0054] 2) Calculate the mean of the target window and the mean and standard deviation of the background window at a short time scale;

[0055] 3) Calculate the mean of the target window and the mean and standard deviation of the background window over a long time scale.

[0056] Step 3: Set the decision thresholds of the three algorithms, namely, the DP-CFAR constant false alarm, the short-time statistical dual-parameter (SSDP-CFAR) constant false alarm, and the short-time statistical dual-parameter (LSDP-CFAR) constant false alarm, and make a decision on the detection result of step 2.

[0057] For the DP-CFAR detection algorithm, the background window mean and standard deviation calculation method of the threshold value is:

[0058]

[0059]

[0060] Among them, n is the total number of points in the background window, x i is the numerical value of the i-th point, μ b is the background window mean, σ b is the standard deviation of the background window.

[0061] The mean of the target window under DP-CFAR detection is subtracted from the mean of the background window and divided by the standard deviation of the background window. The result is compared with the nominal factor to determine whether the point in the target window is the target point:

[0062]

[0063] Among them, μ c is the mean of the target window; μ b is the background window mean; σ b is the standard deviation of the background window; T is the nominal factor. If the above formula is satisfied, the output is 1 when it indicates the presence of a target, otherwise the output is 0 when it indicates the absence of a target.

[0064] For the SSDP-CFAR and LSDP-CFAR detection algorithms, the mean and standard deviation used for judgment are considered separately for short-time scale and long-time scale, except for the first frame of image data, which is directly used for DP-CFAR detection. Starting from the second frame, the mean and standard deviation of the background window used to calculate the threshold value in the p-th frame are as follows:

[0065]

[0066]

[0067] in, is the mean value of the target window in the previous p frames (when calculating a long time scale, p tends to infinity); μ' b To calculate the background window mean of the threshold value at short time scale and long time scale respectively; is the standard deviation of the target window of the previous p frames (when calculating long time scales, p tends to infinity); σ' b To calculate the background window standard deviation of the threshold value at short time scale and long time scale respectively; k is a weight ranging from 0.7 to 0.99.

[0068] Step 4: The three time-scale detection algorithms output their respective judgment results for points in the same target window through a logical decision maker. By combining the three judgment results, it is possible to identify and distinguish moving targets, stationary targets, interfering targets, and clutter within the detection window. The specific judgment rules are as follows:

[0069] Among them, 1 indicates that the point is a target, and 0 indicates that there is no target. The judgment results of the three are jointly judged to identify and distinguish the moving target, temporary stationary target, interference target and clutter in the detection section. Among them, the resolution unit where all three are judged as 1 is the moving target (111), the resolution unit where SSDP-CFAR judges as 0 and the rest are judged as 1 is the temporary stationary target (101), the resolution unit where DP-CFAR judges as 1 and LSDP-CFAR judges as 0 is the interference target (100), and the remaining resolution units are all clutter.

[0070] Step 5: After removing the areas corresponding to some real targets in the target window that will affect the detection effect according to the target type, the mean and standard deviation of the remaining resolution units in the target window are calculated and then fed back to the calculation of the mean and standard deviation of the background window of the next frame in the short-time statistical dual-parameter and long-time statistical dual-parameter constant false alarm systems, respectively.

[0071] Step 6: Establish a multi-space type database of UAV swarm formations, which includes echo simulation images of several typical formation types. The specific steps are as follows:

[0072] The actual collected radar echo data is extracted for multiple scanning cycles and the extracted echo data is normalized. The normalized calculation expression is:

[0073]

[0074]

[0075] Among them, V imax The peak value in the echo data collected in each scanning cycle; For each cycle, V imax Calculate the average value; V%(t) is the normalized echo data.

[0076] According to the radar equation, the normalized echo data is used to calculate its echo strength at different times. Assuming that the distance between the target and the radar is R, the received power of the radar antenna cross section under ideal conditions is:

[0077]

[0078] Among them, P t is the peak power of the radar transmitted pulse, G is the antenna gain, and σ is the average radar cross-section of the target.

[0079] Therefore, the target echo intensity can be obtained as follows:

[0080] P r =P r0 F 4 L a

[0081] Among them, F is the propagation factor, L a Atmospheric attenuation.

[0082] Step 7: Use the nearest neighbor peak search method to match the target initially detected in step 4 with the preset multi-space type database and output the target classification result.

[0083] Consider the following two scenarios: When multiple targets are located in different resolution units of the radar, their echo peaks are independent of each other, and the corresponding echo type can be matched by searching in ascending order by the number of peaks; when multiple targets are located in the same resolution unit of the radar, their echo peaks are coherently accumulated, and the accumulated peak value is much larger than the peak value of the echo of a single target. In this case, the nearest neighbor search algorithm is used to search in sequence for the echo simulation image that is closest to the number of echo peaks, and calculate the Euclidean distance between each peak point of the echo and each peak point of the simulation image. The calculation formula is as follows:

[0084]

[0085] Among them, x i is the peak amplitude of each echo, y i is the peak amplitude of the simulation graph, and i is the number of peaks.

[0086] Then, the Euclidean distance is standardized, and each calculated value is standardized to the interval [0,1]. The echo simulation graph corresponding to the minimum value is found, which is the optimal type of the target. At this time, the output is the recognition result of the target.

[0087] Step 8: Iteratively update and expand the multi-space type database based on the matching results of step 7.

[0088] The matched echo spatial distribution can be used as the characteristic information of the real target to enrich the database of this type, while the unmatched echo types can also be fed back to the multi-spatial type database as prior information to expand the echo types in the database, thereby accelerating the subsequent matching process and forming a self-learning and self-improving closed-loop cognitive system.

Claims

1. A multi-scale spatial and temporal UAV cluster target pre-detection recognition method, characterized by: The following steps are involved: (1) acquiring target echo data and feeding it into a multi-time scale system; the multi-time scale system includes a constant false alarm detection system, a logic decision device, and a feedback system; wherein the constant false alarm detection system includes three detection algorithms: dual-parameter constant false alarm, short-time statistical dual-parameter constant false alarm, and long-time statistical dual-parameter constant false alarm; (2) Setting the sliding window parameters of the multi-time-scale constant false alarm detection method and detecting different types of targets in the target windows at three time scales respectively; (3) setting the decision thresholds of the three algorithms, namely, dual-parameter constant false alarm, short-time statistical dual-parameter constant false alarm, and long-time statistical dual-parameter constant false alarm, respectively, and making a decision on the detection result of step (2); (4) The three time-scale detection algorithms are configured to output their respective judgment results for points in the same target window through a logic judge; the judgment results of the three algorithms are combined to preliminarily identify and distinguish the echo data within the detection window; (5) After removing the areas corresponding to some real targets that affect the detection effect in the target window according to the target type, the mean and standard deviation of the remaining resolution units in the target window are calculated and then fed back to the calculation of the mean and standard deviation of the background window of the next frame in the short-time statistical dual-parameter and long-time statistical dual-parameter constant false alarm respectively through the feedback system; (6) Establish a multi-space type database of UAV swarm formations, which includes echo simulation images of several typical formation types; (7) Using the nearest peak search method, the target initially detected in step (4) is matched with the preset multi-space type database to output the target classification result; (8) Iteratively update and expand the multi-space type database based on the matching results of step (7).

2. The method for pre-detection recognition of drone cluster targets based on multiple spatiotemporal scales according to claim 1 is characterized in that: The echo data in step (1) includes moving targets, stationary targets, interfering targets and clutter.

3. The method for pre-detection identification of drone cluster targets based on multiple spatiotemporal scales according to claim 1 is characterized in that: The sliding window parameters in step (2) include: the mean of the target window, the mean and standard deviation of the background window in the current frame; the mean of the target window, the mean and standard deviation of the background window under a short time scale; and the mean of the target window, the mean and standard deviation of the background window under a long time scale.

4. The method for pre-detection identification of drone cluster targets based on multiple spatiotemporal scales according to claim 1 is characterized in that: The implementation process of step (3) is as follows: For the DP-CFAR detection algorithm, the background window mean and standard deviation calculation method of the threshold value is: Among them, n is the total number of points in the background window, x i is the numerical value of the i-th point, μ b is the background window mean, σ b is the standard deviation of the background window; The mean of the target window under DP-CFAR detection is subtracted from the mean of the background window and divided by the standard deviation of the background window. The result is compared with the nominal factor to determine whether the point in the target window is the target point: Among them, μ c is the mean of the target window; μ b is the background window mean; σ b is the standard deviation of the background window; T is the nominal factor; if the above formula is satisfied, the output is 1 when it indicates that there is a target, otherwise the output is 0 when it indicates that there is no target; For the SSDP-CFAR and LSDP-CFAR detection algorithms, the mean and standard deviation used for judgment are directly used for DP-CFAR detection in the first frame of image data. Starting from the second frame, the mean and standard deviation at the short time scale and long time scale need to be taken into account respectively. The mathematical model of the background window mean and standard deviation used to calculate the threshold value in the p-th frame is as follows: in, is the mean value of the target window in the previous p frames (when calculating a long time scale, p tends to infinity); μ' b To calculate the background window mean of the threshold value at short time scale and long time scale respectively; is the standard deviation of the target window of the previous p frames (when calculating long time scales, p tends to infinity); σ' b To calculate the background window standard deviation of the threshold value at short time scale and long time scale respectively; k is a weight ranging from 0.7 to 0.

99.

5. The method for pre-detection identification of drone cluster targets based on multiple spatiotemporal scales according to claim 1 is characterized in that: The implementation process of step (6) is as follows: Extract data from multiple scanning cycles of the actual collected radar echo data and normalize the extracted echo data: Among them, V imax The peak value in the echo data collected in each scanning cycle; For each cycle, V imax Calculate the average value; V%(t) is the normalized echo data; According to the radar equation, the echo strength at different times is calculated for the normalized echo data. Assuming that the distance between the target and the radar is R, the received power of the radar antenna cross section under ideal conditions is: Among them, P t is the peak power of the radar transmitted pulse, G is the antenna gain, and σ is the average radar cross-sectional area of the target; the target echo intensity is obtained as follows: P r =P r0 F 4 L a Among them, F is the propagation factor, L a Atmospheric attenuation.

6. The method for pre-detection identification of drone cluster targets based on multiple spatiotemporal scales according to claim 1 is characterized in that: The implementation process of step (7) is as follows: The nearest neighbor search algorithm is used to search for the echo simulation graphs that are closest to the echo peak number in sequence, and the Euclidean distance between each peak point of the echo and each peak point of the simulation graph is calculated. The calculation formula is as follows: Among them, x i is the peak amplitude of each echo, y i is the peak amplitude of the simulation graph, i is the number of peaks; Standardize the Euclidean distance, standardize each calculated value to the interval [0,1], find the echo simulation image corresponding to the minimum value, which is the optimal type of the target, and output the target recognition result.

7. The method for pre-detection identification of drone cluster targets based on multiple spatiotemporal scales according to claim 1 is characterized in that: The implementation process of step (8) is as follows: The matched echo spatial distribution can be used as the characteristic information of the real target to enrich the type database, and the unmatched echo types can also be fed back to the multi-spatial type database as prior information to expand the echo types of the database.

Images