A method for identifying unmanned aerial vehicles based on multi-information fusion of Doppler radar

Abstract

The application discloses a kind of unmanned aerial vehicle identification methods based on Doppler radar multi-information fusion, including receiving radar front-end processing original echo data, and the original echo data is carried out two-dimensional FFT calculation, and the distance-speed two-dimensional Doppler spectrum is obtained.The distance-speed two-dimensional Doppler spectrum is accumulated between frames to form clutter map, and the target original EP point information is output based on clutter map CFAR detection.The target point track extraction is carried out to the original EP point information, and the information set of target is obtained.Information set includes n frames of EP point traces associated with history, condensed point trace and track information.Target knowledge base is established, and model training is carried out to target knowledge base.Information set is normalized and input into the trained model for identification, and the target type is obtained.The application improves the integrity of target extraction, simplifies the complexity of detection model, reduces the detection delay while improving the target recognition accuracy, and has stronger practicability.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) detection technology, and specifically to a UAV identification method based on Doppler radar multi-information fusion. Background Technology

[0002] The low-altitude surveillance radar system is a Doppler radar system, possessing an automated defense system integrating detection, identification, and response. It provides all-weather surveillance of ground targets and low-altitude targets, including drones and helicopters. Range / Doppler processing is accomplished by observing the pulse-to-pulse changes in the echo frequency, that is, extracting unambiguous range and Doppler velocity information through two-dimensional FFT calculation. By providing high-precision position and velocity information, it can detect various types of targets, including single individuals, multiple individuals, single vehicles, and aircraft, monitor and track target activities, and determine and display the target's position and trajectory.

[0003] Target identification typically utilizes radar detection information, including target echo intensity and target speed, to guide photoelectric imaging. This results in a comprehensive assessment combining photoelectric and video information to identify the target type and monitor its activity. Traditional photoelectric image-based target identification methods suffer from complex system models, high computational demands, significant latency, and the inability to simultaneously detect targets in multiple areas. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying unmanned aerial vehicles (UAVs) based on Doppler radar multi-information fusion.

[0005] The technical problem solved by this invention is:

[0006] Traditional target recognition methods based on photoelectric images have complex system models, high computational requirements, long latency, and cannot detect targets in multiple regions simultaneously.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A UAV identification method based on Doppler radar multi-information fusion includes the following steps:

[0009] S1. Receive the raw echo data processed by the radar front end, and perform two-dimensional FFT calculation on the raw echo data to obtain the range-velocity two-dimensional Doppler spectrum.

[0010] S2. Perform inter-frame accumulation of the range-velocity two-dimensional Doppler spectrum to form a clutter map, and perform CFAR detection based on the clutter map to output the original EP trace information of the target.

[0011] S3. Extract the target point track from the original EP point track information to obtain the target information set. The information set includes historically associated n-frame EP point tracks, aggregated point tracks, and track information.

[0012] S4. Establish the target knowledge base and train the model on the target knowledge base.

[0013] S5. Normalize the information set and input it into the trained model for recognition to obtain the target type.

[0014] As a further aspect of the present invention: step S1 includes:

[0015] S11, the I / Q data output by the radar front-end receiver after intermediate frequency input mixing and first-dimensional FFT processing.

[0016] S12. Perform inter-pulse FFT calculations on the I / Q data output by the first-dimensional FFT, observe the pulse-to-pulse changes in the echo frequency, extract unambiguous Doppler information, and obtain the range-velocity two-dimensional Doppler spectrum. The range-velocity two-dimensional Doppler spectrum is a two-dimensional RD power spectrum containing the range dimension and the frequency dimension.

[0017] As a further aspect of the present invention: the first-dimensional FFT processing operation is to observe the echo frequency through a distance-matched filter, suppress the difference frequency signal outside the range, derive the distance information in the frequency domain, and extract the unambiguous distance information.

[0018] As a further aspect of the present invention: step S2 includes:

[0019] S21. Initialize the clutter map by performing two-dimensional quantization on the range-velocity two-dimensional Doppler spectrum of each beam. The range quantization unit is Δr, and the frequency quantization unit is Δf. Calculate the mean and variance of the echo signal of each quantized map unit over N repetition periods. The clutter map contains Nr×Ns library units.

[0020] S22. Using the clutter map as the reference background map for CFAR detection, the detection threshold is obtained by using the power of the cells near the detection cell as a reference. The target cell is determined by comparing the echo power of the detection cell with the detection threshold, and then the target EP trace information is output.

[0021] S23. Update the clutter map cells to different degrees depending on whether each quantization map cell is the target.

[0022] As a further aspect of the present invention: the EP trace information includes: the target's timestamp, the target's distance dimension index, the target's velocity dimension index, and the target's amplitude value. Wherein,

[0023] The target's timestamp is the time when the radar detected the target.

[0024] The target's range dimension index is the index value of the range library cell of the target cell detected by the radar.

[0025] The velocity dimension index of the target is the index value of the velocity dimension cell of the target cell detected by the radar.

[0026] The target's amplitude value is the power value of the velocity dimension corresponding to the target element detected by the radar.

[0027] As a further aspect of the present invention: the target point trajectory extraction process includes:

[0028] Point-track convergence first performs single-beam mid-range and velocity convergence on the EP point-track information, then performs multi-beam correlation and azimuth convergence, and finally performs point-track filtering and outputs the converged point-track information. The target converged point-track information is the centroid-weighted value of the relevant EP point-track information.

[0029] The convergence point tracks are correlated and filtered to form track targets with unique batch numbers, and the information set of the batch of targets is output.

[0030] As a further aspect of the present invention: step S4 includes:

[0031] S41. Form a target knowledge base by normalizing the information set and then adding it to the target knowledge base.

[0032] S42. Train the model on the target knowledge base by grouping the completed target knowledge base and inputting it into the model to train and solve the support vector machine hyperplane function to obtain the model parameters.

[0033] As a further aspect of the present invention: step S5 includes:

[0034] S51. Normalize the information set.

[0035] S52. Input the normalized information set into the trained model, and output the most likely category of the target based on the newly observed feature data to identify the target and output the target type.

[0036] The present invention has at least one of the following beneficial effects:

[0037] 1. This invention preserves the amplitude information corresponding to each frequency index of the target unit in the same distance library index in the two-dimensional Doppler spectrum, and passes it as the spectral feature of the target to the next processing flow, thus providing a foundation for subsequent target tracking and recognition capabilities;

[0038] 2. The range of values ​​for the forgetting factor k and the exponential factor α in the clutter map update of this invention is a range obtained after balancing the mapping performance of clutter and real targets, so as to ensure the normal establishment and update of clutter map and thus guarantee the target detection performance of CFAR.

[0039] 3. Re-search for multi-frequency targets in the same distance library, process the output targets in the same distance library, limit the number of EP traces of the same target to avoid full-channel output similar to the rotor spectrum of UAVs, remove redundant feature values, and save resources for subsequent processing.

[0040] 4. The point trajectory aggregation and extraction in this invention further eliminates and extracts target features, providing more effective information for subsequent target recognition and improving the accuracy of target recognition;

[0041] 5. This invention combines multi-level feature information from the radar processing process to improve the completeness of target extraction with a simplified detection model and low detection latency, achieving a target recognition accuracy of over 99%. By backtracking the recognition results, it enhances the detection performance of small targets by UAVs, making it more practical. Attached Figure Description

[0042] The invention will now be further described with reference to the accompanying drawings.

[0043] Figure 1 This is a flowchart illustrating the UAV identification method based on Doppler radar multi-information fusion of the present invention.

[0044] Figure 2 This is a logic diagram of the UAV identification method based on Doppler radar multi-information fusion of the present invention;

[0045] Figure 3 It is a 128-point FFT two-dimensional spectrum of the UAV target in the two-dimensional Doppler spectrum;

[0046] Figure 4 It is a 256-point FFT two-dimensional spectrum of the UAV target in the two-dimensional Doppler spectrum;

[0047] Figure 5 This is a flowchart of the CFAR detection based on clutter maps of the present invention;

[0048] Figure 6 This is a schematic diagram of clutter map updating according to the present invention;

[0049] Figure 7 This is a result diagram of the target detection, tracking and recognition based on 10 frames of feature vectors in this invention. Detailed Implementation

[0050] 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.

[0051] Please see Figure 1-2 As shown, the present invention is a UAV identification method based on Doppler radar multi-information fusion, including the following steps S1-S5.

[0052] S1. Receive the raw echo data processed by the radar front end, and perform two-dimensional FFT calculation on the raw echo data to obtain the range-velocity two-dimensional Doppler spectrum.

[0053] S2. Perform inter-frame accumulation of the range-velocity two-dimensional Doppler spectrum to form a clutter map, and perform CFAR detection based on the clutter map to output the original EP trace information of the target.

[0054] S3. Extract the target point track from the original EP point track information to obtain the target information set. The information set includes historically associated n-frame EP point tracks, aggregated point tracks, and track information.

[0055] S4. Establish the target knowledge base and train the model on the target knowledge base.

[0056] S5. Normalize the information set and input it into the trained model for recognition to obtain the target type.

[0057] Specifically, in this invention, in step S1, the raw echo data processed by the radar front end is the I / Q data output by the first-dimensional FFT processor after the intermediate frequency input of the receiver is mixed.

[0058] The first dimension FFT is to observe the echo frequency through a distance-matched filter, suppress the difference frequency signal outside the range, derive the distance information in the frequency domain, and extract the unambiguous distance information.

[0059] The windowing method uses a carefully designed weighted window function to suppress sidelobe leakage in FFT and improve Doppler velocity resolution. An iterative search algorithm is used to design and optimize the window function, providing window functions with sidelobe suppression performance ranging from 30dB to 65dB.

[0060] The second-dimensional FFT is to perform inter-pulse FFT calculation on the I / Q data output by the first-dimensional FFT, observe the pulse-to-pulse change of the echo frequency, and extract unambiguous Doppler information;

[0061] The range-velocity two-dimensional Doppler spectrum is a two-dimensional RD power spectrum that includes the range dimension and the frequency dimension;

[0062] In a power spectrum calculation, the radar performs FFT processing on Nr consecutive pulses, and the range of detection distance on each pulse is quantized into Ns range libraries, that is, each Doppler velocity corresponds to Ns range libraries, and each range library corresponds to Nr Doppler velocities.

[0063]

[0064] In the formula, r is the range library index, f is the Doppler velocity library index, and x(m, l) is the raw echo I / Q data of the l-th range library of the m-th pulse.

[0065] In one embodiment of the present invention, the radar covers an azimuth range of 0° to 90°, with a total of 64 azimuths; the radar detects a range of distances of 0km to 24km in each azimuth, with a distance length of 12m, a distance number Ns of 2000, and a distance number Nr of 128.

[0066] In step S2, the inter-frame accumulation to form a clutter map for CFAR detection includes the following steps:

[0067] S21, Initialize the clutter map, perform two-dimensional quantization on the RD power spectrum of each beam, with distance quantization unit being Δr and frequency quantization unit being Δf, and calculate the mean and variance of the echo signal of each quantized map unit over N repetition cycles. The clutter map contains Nr×Ns library units.

[0068] ave n (f,r)=ave n-1 (f,r)×(1-k)+spec(f,r)×k

[0069]

[0070] Where aven(f,r) is the mean amplitude of frequency direction f and range direction r of a certain beam in the nth frame, varn(f,r) is the amplitude variance, k∈(0,1) is the forgetting factor, and spec(f,r) is the echo amplitude of the current frame.

[0071] After the radar is powered on, in order to quickly establish a clutter map, the forgetting factor in the clutter map update of the first 30 frames of scanning is set to 0.5.

[0072] S22 uses clutter map as reference background map for CFAR detection, uses the power of nearby cells as reference to obtain detection threshold, and determines whether it is a target cell by comparing the echo power of the detection cell with the detection threshold, and outputs target EP point trace information; compared with traditional range CFAR, it adds reference cell selection in Doppler velocity dimension, and the number of protection cells and reference cells in each direction can be changed;

[0073] For detection unit D, (x1...x n ) and (y1...y nThe reference cells (D) are located on both sides of the cell to be detected in the clutter diagram. The reference window contains 2n cells, and 2m guard cells are set on both sides of the cell to be detected. X and Y are local estimates of the cells within the two reference windows. By comparing the cell to be detected (D) with the adaptive threshold S = CFAR(X,Y) (the CFAR function can be average, maximum, minimum, sorting, etc.), the target can be automatically detected.

[0074] The EP point information includes: the target's timestamp, the target's distance dimension index, the target's velocity dimension index, and the target's amplitude value;

[0075] The timestamp of the target: the time when the radar detected the target;

[0076] The target's range dimension index: the index value of the range library cell of the target cell detected by the radar;

[0077] The velocity dimension index of the target: the index value of the velocity dimension unit of the target unit detected by the radar;

[0078] The target's amplitude value: the power value of the target element corresponding to the velocity dimension detected by the radar;

[0079] S23, the clutter map unit is updated to different degrees depending on whether each quantized map unit is a target. In order to avoid the missed detection and false alarm caused by drastic fluctuations in map building, the clutter map building process is divided into target map building and clutter map building.

[0080] In one embodiment of the present invention, after the radar scans 30 frames to build the map, it is converted into a clutter map update. When there is a target in the current detection unit, the echo data of the current unit participates in the clutter update with a small amplitude to avoid the clutter map amplitude from increasing rapidly.

[0081] ave n (f,r)=ave n-1 (f,r)×delta

[0082] delta=powf(spec(f,r) / ave(f,r),α)

[0083] In the formula, delta is the update coefficient, and α∈(0,1) is the exponential factor, which can be set to 0.125.

[0084] When the current detection unit is not the target, update the mean and variance of the clutter plot based on the amplitude of the unit echo data.

[0085] ave n (f,r)=ave n-1 (f,r)×delta

[0086]

[0087] In the formula, k∈(0,1) is the update forgetting factor, which is set to 0.025, and α∈(0,1) is the exponential factor, which can be set to 0.875.

[0088] In step S3, the target point track extraction includes the following steps:

[0089] S31, Point Track Convergence: First, single-beam mid-range and velocity convergence is performed on the EP point track information. Then, multi-beam correlation and azimuth convergence are performed. Finally, point track filtering is performed and the converged point track information is output.

[0090] The target aggregation point information includes: the target's timestamp, the target's distance dimension index, the target's velocity dimension index, the target's orientation index, and the target's amplitude value;

[0091] The target aggregation point information is the centroid weighted value of the relevant EP point information;

[0092] S32, perform track association and filtering on the convergence point traces to form track targets with unique batch numbers, and output the information set of the batch targets, including historically associated n frames of EP traces, convergence point traces and track information;

[0093] The target trajectory information includes: the target's timestamp, the target's distance, the target's speed, the target's bearing, and the target's heading;

[0094] The target information set is the EP point trace, convergence point trace and track information set of n historical frames associated with a track, which is an information set with a series of target characteristics extracted through experience.

[0095] The training data set T = {(X1,Y1),(X2,Y2),...(X...} m ,Y m )}, where m is the number of instances of the set, X i =(t1,v1,x1,y1,a1;t2,v2,x2,y2,a2;...;t n ,v n ,x n ,y n ,a n ), t, v, x, y, a are time, speed, distance, azimuth, and amplitude information vectors, respectively; n is the number of relevant frames in the flight path history, Y = {+1, -1}, where 1 represents UAV and -1 represents others (clutter or people and vehicles).

[0096] In one embodiment of the present invention, feature vectors from 10 historical frames are selected as model input data.

[0097] In step S4, the formation of the knowledge base for model training includes the following steps:

[0098] S41, Form a knowledge base by normalizing the information set and adding it to the target knowledge base;

[0099] The normalization process includes: normalizing the target's distance, speed, orientation, amplitude, etc. The most typical normalization method is the ratio of the current value to the maximum value of the relevant attribute.

[0100] S42, train the model on the knowledge base by grouping the completed knowledge base into the model for training and solving the support vector machine hyperplane function to obtain the model parameters;

[0101] Among them, the separating hyperplane:

[0102] w * ·x+b * =0

[0103] Where w* is the weight vector and b* is the intercept term.

[0104] In step S5, the multi-information fusion for target recognition includes the following steps:

[0105] S51, normalize the information set;

[0106] S52, input the normalized information set into the trained detection model, and output the most likely category of the instance for target recognition based on the new observed feature data, and output the target type;

[0107] Decision function: f(x) = sign(w) * ·x+b * ).

[0108] This invention introduces multi-level feature vectors of N-frame point tracks, uses SVM (Support Vector Machine) and BDT (Binary Decision Tree) for training and detection, and repeatedly measures the classification decision rate of the proposed algorithm. It was found that the detection performance of UAVs exceeded 99% and 96% respectively, and SVM was finally selected as the model for practical use.

[0109] Please combine Figure 3-7The identification results of the UAV identification method based on Doppler radar multi-information fusion of the present invention are presented. The experimental data are I / Q echo data after real-time acquisition of A / D sampling from the front end of low-altitude surveillance radar. Each beam contains 128 pulse repetition cycles and is divided into 2000 range library units. The present invention retains the amplitude information corresponding to each frequency index of the target unit in the same range library index in the two-dimensional Doppler spectrum as the target's spectral features and passes it to the next processing flow, providing a foundation for subsequent target tracking and identification capabilities. Multi-frequency targets in the same range library are re-searched and processed for output targets in the same range library, limiting the number of EP traces output for the same target to avoid full-channel output similar to UAV rotor spectrum, eliminating redundant feature values, and saving resources for subsequent processing flows. In the present invention, point track aggregation and extraction further eliminate and extract target features, providing more effective information for subsequent target identification and improving the target identification accuracy. The feature vector is normalized to provide a reliable data foundation for subsequent identification model processing. Figure 7 As shown, by combining multi-level feature information from the radar processing, the completeness of target extraction is improved with a simplified detection model and low detection latency, achieving a target recognition accuracy of over 99%. By backtracking the recognition results, the detection performance of small targets by UAVs is enhanced, making it more practical.

[0110] 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 identifying unmanned aerial vehicles (UAVs) based on Doppler radar multi-information fusion, characterized in that, Includes the following steps: S1. Receive the raw echo data processed by the radar front end, and perform two-dimensional FFT calculation on the raw echo data to obtain the range-velocity two-dimensional Doppler spectrum; S2. Perform inter-frame accumulation of the range-velocity two-dimensional Doppler spectrum to form a clutter map, and perform CFAR detection based on the clutter map to output the original EP trace information of the target; Step S2 includes: S21. Initialize the clutter map, perform two-dimensional quantization on the range-velocity two-dimensional Doppler spectrum of each beam, with the range quantization unit being Δr and the frequency quantization unit being Δf, and calculate the mean and variance of the echo signal of each quantized map unit over N repetition periods. The clutter map contains Nr×Ns library units. S22. Using the clutter map as the reference background map for CFAR detection, the detection threshold is obtained by using the power of the cells near the detection cell as a reference. The target cell is determined by comparing the echo power of the detection cell with the detection threshold, and then the target EP point trace information is output. S23. Update the clutter map cells to different degrees depending on whether each quantization map cell is the target; S3. Extract the target point track from the original EP point trace information to obtain the target information set; the information set includes historically associated n-frame EP point traces, aggregated point traces, and track information; The target point track extraction process includes: Point-track convergence first performs single-beam mid-range and velocity convergence on the EP point-track information, then performs multi-beam inter-correlation and azimuth convergence, and finally performs point-track filtering and outputs converged point-track information; wherein, the converged point-track information is the centroid weighted value of the relevant EP point-track information; The convergence point tracks are correlated and filtered to form track targets with unique batch numbers, and the information set of the batch of targets is output. S4. Establish the target knowledge base and train the model on the target knowledge base; S5. Normalize the information set and input it into the trained model for recognition to obtain the target type.

2. The UAV identification method based on Doppler radar multi-information fusion according to claim 1, characterized in that, Step S1 includes: S11. The I / Q data output by the radar front-end receiver after intermediate frequency input mixing is processed by the first-dimensional FFT. S12. Perform inter-pulse FFT calculation on the I / Q data output by the first-dimensional FFT, observe the pulse-to-pulse change of the echo frequency, extract the unambiguous Doppler information, and obtain the range-velocity two-dimensional Doppler spectrum; the range-velocity two-dimensional Doppler spectrum is a two-dimensional RD power spectrum containing the range dimension and the frequency dimension.

3. The UAV identification method based on Doppler radar multi-information fusion according to claim 2, characterized in that, The first-dimensional FFT processing operation observes the echo frequency through a distance-matched filter, suppresses the difference frequency signal outside the range, derives the distance information in the frequency domain, and extracts the unambiguous distance information.

4. The UAV identification method based on Doppler radar multi-information fusion according to claim 1, characterized in that, The EP trace information includes: the target's timestamp, the target's distance index, the target's velocity index, and the target's amplitude value; wherein, The timestamp of the target is the time when the radar detected the target; The range dimension index of the target is the index value of the range library cell of the target cell detected by the radar; The velocity dimension index of the target is the index value of the velocity dimension unit of the target unit detected by the radar; The amplitude value of the target is the power value of the velocity dimension corresponding to the target element detected by the radar.

5. The UAV identification method based on Doppler radar multi-information fusion according to claim 1, characterized in that, Step S4 includes: S41. Form a target knowledge base by normalizing the information set and then adding it to the target knowledge base; S42. Train the model on the target knowledge base by grouping the completed target knowledge base and inputting it into the model to train and solve the support vector machine hyperplane function to obtain the model parameters.

6. The UAV identification method based on Doppler radar multi-information fusion according to claim 1, characterized in that, Step S5 includes: S51. Normalize the information set; S52. Input the normalized information set into the trained model, and output the most likely category of the target based on the newly observed feature data to identify the target and output the target type.

Images