Unmanned aerial vehicle identification model training method and unmanned aerial vehicle identification method

Abstract

The invention relates to the technical field of target recognition, discloses an unmanned aerial vehicle recognition model training method and an unmanned aerial vehicle recognition method, and aims to solve the problems of high cost and low efficiency of an existing unmanned aerial vehicle recognition scheme, and the scheme mainly comprises the following steps: segmenting radar data of an unmanned aerial vehicle track to obtain a plurality of first unmanned aerial vehicle samples; correspondingly adding the feature data of each trace point on any two first unmanned aerial vehicle samples according to the same random weight, and then adding random perturbation to obtain a plurality of second unmanned aerial vehicle samples; segmenting the radar data of the non-unmanned aerial vehicle flight path to obtain a plurality of non-unmanned aerial vehicle samples; randomly selecting a second unmanned aerial vehicle sample as a third unmanned aerial vehicle sample; and constructing an identification sample set based on the first unmanned aerial vehicle sample, the third unmanned aerial vehicle sample and the non-unmanned aerial vehicle sample, and training an unmanned aerial vehicle identification model based on the identification sample set. The training efficiency and accuracy of the unmanned aerial vehicle identification model are improved, the cost is reduced, and the method is particularly suitable for radar detection.

Description

Technical Field

[0001] The present application relates to the field of target recognition technology, and in particular to a drone recognition model training method and a drone recognition method. Background Art

[0002] Unmanned aerial vehicles are referred to as UAVs. UAVs are unmanned aircraft that are controlled by radio remote control equipment and self-contained program control devices, or are completely or intermittently operated autonomously by onboard computers. UAVs have been widely used in military scenarios, civil monitoring, logistics distribution and other fields, showing their unique strategic value and broad application prospects.

[0003] In some application scenarios, especially military scenarios, drones need to be identified. The traditional solution is to use radar detection systems to identify drones. However, traditional radar detection systems mainly focus on measuring basic parameters such as the presence, height, and speed of targets, and lack the ability to accurately identify and classify targets, especially it is difficult to effectively distinguish drones from "low, small, and slow" targets such as birds and small drones. This limitation seriously restricts the effectiveness of radar in complex environments, allowing new aircraft such as drones to easily evade radar monitoring.

[0004] Application publication number CN106546975 discloses a method for classifying and identifying small and light UAVs and flying birds based on radar data. The main disclosed scheme is based on the target measurement information obtained by radar. Through four steps such as multi-model target tracking, multi-model probability extraction, target motion mode judgment, and target motion feature extraction, the target features such as the target motion model conversion frequency are finally extracted to distinguish small and light UAV targets from flying bird targets. However, this method requires a large amount of labeled data to train the classification model. The work of collecting a large amount of radar detection data and labeling is not only costly, but also time-consuming and labor-intensive, and inefficient, and it is difficult to meet the needs of practical applications. Summary of the invention

[0005] This application aims to solve the problems of high cost and low efficiency of existing drone identification solutions, and proposes a drone identification model training method and a drone identification method.

[0006] The technical solution adopted by this application to solve the above technical problems is:

[0007] In a first aspect, the present application provides a method for training a drone recognition model, the method comprising:

[0008] The radar data of the UAV track is divided according to a preset track length to obtain a plurality of first UAV samples;

[0009] The feature data of each point trace on any two first drone samples are added correspondingly according to the same random weight, and then random perturbations are added to obtain multiple second drone samples;

[0010] The radar data of the non-UAV track is divided according to the preset track length to obtain multiple non-UAV samples;

[0011] randomly selecting a second drone sample that is a preset multiple of the number of non-drone samples as the third drone sample;

[0012] An identification sample set is constructed based on the first drone sample, the third drone sample, and the non-drone sample, and a drone identification model is trained based on the identification sample set.

[0013] Furthermore, the radar data of the UAV track and the radar data of the non-UAV track respectively include the distance, pitch angle, azimuth and radial velocity of each point on the track.

[0014] Furthermore, the feature data of each point trace on any two first drone samples are added together according to the same random weight, specifically including:

[0015] Draw a random number from a Beta distribution as a random weight;

[0016] For any two first UAV samples, the radar data of each point trace thereon are resolved into feature data, and each feature data is summed up accordingly based on a random weight to obtain a plurality of weighted first UAV samples;

[0017] There are six types of characteristic data, including position components in the x, y, and z directions and radial velocity components in the x, y, and z directions in a spatial rectangular coordinate system.

[0018] Furthermore, the calculation formula of the weighted first drone sample is as follows:

[0019] X3(i)=[x3(i),y3(i),z3(i),Vx3(i),Vy3(i),Vz3(i)]=w*X1(i)+(1-w)*X2(i), 1≤i≤L;

[0020] X1(i)=[x1(i),y1(i),z1(i),Vx1(i),Vy1(i),Vz1(i)];

[0021] X2(i)=[x2(i),y2(i),z2(i),Vx2(i),Vy2(i),Vz2(i)];

[0022] Among them, X3(i) represents the feature data of the i-th point on the weighted first drone sample, x3(i), y3(i), z3(i) represent the position components of the i-th point on the weighted first drone sample in the x, y, and z directions, Vx3(i), Vy3(i), Vz3(i) represent the radial velocity components of the i-th point on the weighted first drone sample in the x, y, and z directions, X1(i) represents the feature data of the i-th point on the first drone sample, x1(i), y1(i), z1(i) represent the position of the i-th point on the first drone sample in the x, y, and z directions, respectively. Components, Vx1(i), Vy1(i), Vz1(i) represent the radial velocity components of the i-th point on the first UAV sample in the x, y, and z directions, respectively. X2(i) represents the feature data of the i-th point on the second first UAV sample. x2(i), y2(i), z2(i) represent the position components of the i-th point on the second first UAV sample in the x, y, and z directions, respectively. Vx2(i), Vy2(i), Vz2(i) represent the radial velocity components of the i-th point on the second first UAV sample in the x, y, and z directions, respectively. w represents the random weight, and L represents the preset track length.

[0023] Furthermore, the radar data of each point on the first UAV sample is solved into feature data, and the solution formula is as follows:

[0024]

[0025] Among them, x n (i),y n (i),z n (i) represents the position components of the ith point trace on the nth first drone sample in the x, y, and z directions, respectively, Vx n (i),Vy n (i),Vz n (i) represents the radial velocity component of the ith point on the nth first UAV sample in the x, y, and z directions, respectively, R n (i) represents the distance of the i-th point trace on the n-th first drone sample, θ n (i) represents the pitch angle of the i-th point on the n-th first drone sample, φ n (i) represents the position of the i-th point on the n-th first drone sample, V n (i) represents the radial velocity of the i-th point on the n-th first UAV sample.

[0026] Furthermore, random perturbations are added, including:

[0027] Obtain feature data of corresponding points from all first UAV samples respectively, and construct feature data sets corresponding to each point respectively;

[0028] Based on the feature data set, the sample mean and sample variance of each feature data of the corresponding point trace are calculated, and a normal distribution is constructed according to the sample mean and sample variance;

[0029] Sampling from a normal distribution, and calculating the perturbation corresponding to each characteristic data of the corresponding point trace according to the sampling results;

[0030] The perturbation is added to the corresponding feature data of the corresponding point trace of the weighted first drone sample to obtain a second drone sample.

[0031] Furthermore, the calculation formula of the perturbation is as follows:

[0032]

[0033] The perturbation is added to the corresponding feature data of the corresponding point trace of the weighted first drone sample, and the formula is as follows:

[0034] Y j (i) = X3 j (i)+δ j (i);

[0035] Among them, δ j (i) represents the perturbation corresponding to the jth characteristic data of the i-th point trace, ∈ j (i) represents the sampling result of the jth feature data of the i-th point, μ j (i) represents the sample mean of the jth feature data of the i-th point, X3 j (i) represents the value of the jth feature data of the i-th point on the weighted first drone sample, J represents the number of feature data types, J=6, k j Indicates the preset parameters, Y j (i) represents the value of the j-th feature data of the i-th point trace of the second drone sample.

[0036] Furthermore, the drone identification model is obtained by training a neural network model, the last layer of the neural network model is a fully connected classification layer, the classification layer is used to output the probability of belonging to the drone using a Sigmoid function, and the loss function of the neural network model is cross entropy loss, weighted cross entropy loss or FocalLoss.

[0037] In a second aspect, the present application provides a method for identifying a drone, the method comprising:

[0038] Acquire radar data of the track of the target to be identified in real time, and divide it into windows according to the preset track length to obtain multiple window samples to be identified;

[0039] Input each window sample into the drone recognition model in turn, determine the recognition result corresponding to each window sample, and judge whether the ratio of the number of window samples identified as drones in the recognition result to the total number of window samples is greater than a first preset ratio. If so, determine that the target to be identified is a drone; otherwise, determine that the target to be identified is not a drone.

[0040] The drone recognition model is obtained by training using the drone recognition model training method according to any one of claims 1 to 8.

[0041] Furthermore, the method further comprises:

[0042] After determining that the target to be identified is a drone, continue to obtain window samples corresponding to the target to be identified, and determine whether the proportion of the number of window samples identified as drones in the identification result to the total number of window samples is less than a second preset proportion. If so, the target to be identified is re-determined as a non-drone. Otherwise, the determination result remains unchanged, and the second preset proportion is less than the first preset proportion.

[0043] The beneficial effects of the present application are as follows: the drone identification model training method and drone identification method provided in the present application design a drone sample generation framework through random enhancement and random perturbation, etc., which solves the problem of difficulty in obtaining drone track sample data and the small quantity, reduces the cost of collecting and labeling radar data, improves the model training efficiency, and also fundamentally solves the problem of imbalance between positive and negative samples; the present application can judge and update online and in real time whether the target detected by the radar is a drone, and the longer the target is detected, the more accurate the result. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flowchart of a method for training a drone identification model provided in an embodiment of the present application;

[0045] Figure 2 A schematic diagram of the structure of a neural network model provided in an embodiment of the present application;

[0046] Figure 3 A flowchart of a method for identifying a drone provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0048] In some of the processes described in the specification of the present application and the above-mentioned figures, multiple operations appearing in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel, and the sequence numbers of the operations are only used to distinguish different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel.

[0049] The technical solution of the embodiment of the present application is applicable to application scenarios where drone identification is required, especially accurate identification of light and small drones and flying birds. Since the current solution for drone identification based on radar data requires the collection of a large amount of labeled sample data to train the model, the collection and labeling of sample data is costly and inefficient, resulting in high cost and low efficiency for model training and drone identification.

[0050] Based on this, the technical solution of the present application is proposed. In an embodiment of the present application, the radar data of the UAV track is divided according to a preset track length to obtain a plurality of first UAV samples; the feature data of each point track on any two first UAV samples are correspondingly added according to the same random weight, and then random perturbations are added to obtain a plurality of second UAV samples; the radar data of the non-UAV track is divided according to a preset track length to obtain a plurality of non-UAV samples; second UAV samples that are a preset multiple of the number of non-UAV samples are randomly selected as third UAV samples; an identification sample set is constructed based on the first UAV samples, the third UAV samples and the non-UAV samples, and a UAV identification model is trained based on the identification sample set.

[0051] Specifically, the present application generates a second drone sample by performing data enhancement on the radar data of the drone track, mainly including random enhancement and adding random perturbations, thereby reducing the amount of radar data collected for the drone track, reducing the cost of collecting and annotating radar data, and improving the efficiency of sample data collection. In addition, the present application extracts the third drone sample based on the number of non-drone samples, fundamentally solving the problem of imbalance between positive and negative samples. The present application uses the enhanced sample data to train the drone recognition model, thereby improving the efficiency and accuracy of building the drone recognition model.

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0053] Figure 1 A flowchart of a method for training a drone identification model provided in an embodiment of the present application is shown. Figure 1 , the method comprises the following steps:

[0054] S1. Divide the radar data of the UAV track according to a preset track length to obtain a plurality of first UAV samples.

[0055] In the embodiment of the present application, the radar data of the drone track can be obtained by collecting the drone target through radar. Specifically, the drone can be controlled to fly along a determined route, and then the drone track is accurately identified, and finally the corresponding radar data is collected by radar. The radar data of the drone track includes the distance R, pitch angle θ, azimuth φ and radial velocity V of each point on the drone track.

[0056] The track length indicates the number of point tracks included. If the track length is L, then it contains L point tracks. In the embodiment of the present application, the preset track length is L, and L is a preset positive integer. That is, the present application divides the radar data of the drone track to obtain radar data corresponding to multiple non-overlapping drone tracks with a track length of L. The radar data corresponding to each drone track constitutes a first drone sample, and each first drone sample contains radar data corresponding to L point tracks.

[0057] In the embodiment of the present application, the preset track length L=7, that is, each first drone sample contains radar data corresponding to 7 point tracks. When dividing, the radar data with a track length less than 7 is directly eliminated, and the track length greater than 7 is divided according to the length 7. If the length of the last track is less than 7, it is directly eliminated.

[0058] S2. According to the same random weight, the feature data of each point on any two first drone samples are correspondingly added, and then random perturbations are added to obtain multiple second drone samples.

[0059] In the embodiment of the present application, firstly, the feature data of each point trace on any two first drone samples are correspondingly added according to the same random weight, which specifically includes the following steps:

[0060] S201. Draw a random number from the Beta distribution as a random weight.

[0061] In practical applications, a random number w is drawn from the Beta distribution Beta(w1,w2) as a random weight, where w1 and w2 are parameters of the distribution, and w1,w2>0; in this embodiment, w1 and w2 are both 1, and the Beta distribution Beta(1,1) is equivalent to a uniform distribution.

[0062] S202: for any two first drone samples, the radar data of each point trace thereon are resolved into feature data, and each type of feature data is summed up accordingly based on a random weight to obtain a plurality of weighted first drone samples.

[0063] In practical applications, the radar data of each point on each first UAV sample is first solved into feature data. The feature data is six types, including position components in the x, y, and z directions and radial velocity components in the x, y, and z directions in the spatial rectangular coordinate system. The solution formula is as follows:

[0064]

[0065] Among them, x n (i),y n (i),z n (i) represents the position components of the ith point trace on the nth first drone sample in the x, y, and z directions, respectively, Vx n (i),Vy n (i),Vz n (i) represents the radial velocity component of the ith point on the nth first UAV sample in the x, y, and z directions, respectively, R n (i) represents the distance of the i-th point trace on the n-th first drone sample, θ n (i) represents the pitch angle of the i-th point on the n-th first drone sample, represents the position of the ith point on the nth first drone sample, V n (i) represents the radial velocity of the i-th point on the n-th first UAV sample.

[0066] The characteristic data of each point on each first drone sample can be obtained through the above formula. The characteristic data Xn(i) of the i-th point on the n-th first drone sample is as follows:

[0067] Xn(i)=[x n (i),y n (i),z n (i),Vx n (i),Vy n (i),Vz n (i)].

[0068] Then, for any two first drone samples X1 and X2, the feature data of each point trace are added together according to the random weight w to obtain the weighted first drone sample. The calculation formula is as follows:

[0069] X3(i)=[x3(i),y3(i),z3(i),Vx3(i),Vy3(i),Vz3(i)]=w*X1(i)+(1-w)*X2(i), 1≤i≤L;

[0070] X1(i)=[x1(i),y1(i),z1(i),Vx1(i),Vy1(i),Vz1(i)];

[0071] X2(i)=[x2(i),y2(i),z2(i),Vx2(i),Vy2(i),Vz2(i)];

[0072] Right now:

[0073] x3(i)=w*x1(i)+(1-w)*x2(i);

[0074] y3(i)=w*y1(i)+(1-w)*y2(i);

[0075] z3(i)=w*z1(i)+(1-w)*z2(i);

[0076] Vx3(i)=w*Vx1(i)+(1-w)*Vx2(i);

[0077] Vy3(i)=w*Vy1(i)+(1-w)*Vy2(i);

[0078] Vz3(i)=w*Vz1(i)+(1-w)*Vz2(i);

[0079] Among them, X3(i) represents the feature data of the i-th point on the weighted first drone sample, x3(i), y3(i), z3(i) represent the position components of the i-th point on the weighted first drone sample in the x, y, and z directions, Vx3(i), Vy3(i), Vz3(i) represent the radial velocity components of the i-th point on the weighted first drone sample in the x, y, and z directions, X1(i) represents the feature data of the i-th point on the first drone sample, x1(i), y1(i), z1(i) represent the position of the i-th point on the first drone sample in the x, y, and z directions, respectively. Components, Vx1(i), Vy1(i), Vz1(i) represent the radial velocity components of the i-th point on the first UAV sample in the x, y, and z directions, respectively. X2(i) represents the feature data of the i-th point on the second first UAV sample. x2(i), y2(i), z2(i) represent the position components of the i-th point on the second first UAV sample in the x, y, and z directions, respectively. Vx2(i), Vy2(i), Vz2(i) represent the radial velocity components of the i-th point on the second first UAV sample in the x, y, and z directions, respectively. w represents the random weight, and L represents the preset track length.

[0080] Repeat the above steps to obtain a corresponding weighted first drone sample for any two first drone samples, and then add random perturbations to obtain the second drone sample.

[0081] In the embodiment of the present application, adding random perturbations specifically includes the following steps:

[0082] S211, respectively obtaining feature data of corresponding points from all first drone samples, and respectively constructing feature data sets corresponding to each point.

[0083] In practical applications, for the i-th point trace of all first drone samples, a feature data set corresponding to the i-th point trace is constructed, and the feature data set contains feature data of the i-th point trace of all first drone samples.

[0084] S212. Calculate the sample mean and sample variance of each feature data of the corresponding point trace based on the feature data set, and construct a normal distribution according to the sample mean and sample variance.

[0085] In practical applications, the sample mean μ(i) of the feature data set corresponding to the i-th point is calculated as follows: j (i)] 1≤j≤J and sample variance Φ(i)=[Φ j (i)] 1≤j≤J , construct a normal distribution, the sample mean is used to determine the center position of the data distribution, and the sample variance is used to measure the discreteness of the data. Wherein, J represents the number of types of feature data, and each type of feature data is calculated independently. In this embodiment, J is 6, which respectively represents the position components in the x, y, and z directions, and the radial velocity components in the x, y, and z directions.

[0086] S213, sampling from the normal distribution, and calculating the perturbation corresponding to each characteristic data of the corresponding point trace according to the sampling result.

[0087] In practical applications, we can sample from the normal distribution N(μ(i),Φ(i)) to obtain ∈(i)=[∈ j (i)] 1≤j≤J , and calculate the perturbation as follows:

[0088]

[0089] The perturbation is added to the corresponding feature data of the corresponding point trace of the weighted first drone sample, and the formula is as follows:

[0090] Y j (i) = X3 j (i)+δ j (i);

[0091] Among them, δ j (i) represents the perturbation corresponding to the jth characteristic data of the i-th point trace, ∈ j (i) represents the sampling result of the jth feature data of the i-th point, μ j(i) represents the sample mean of the jth feature data of the i-th point, X3 j (i) represents the value of the jth feature data of the i-th point on the weighted first drone sample, J represents the number of feature data types, J=6, k j represents a preset parameter. In the embodiment of the present application, k j =0.01, Y j (i) represents the value of the j-th feature data of the i-th point trace of the second drone sample.

[0092] S214: Add the perturbation to the corresponding feature data of the corresponding point traces of the weighted first UAV sample to obtain a second UAV sample.

[0093] In practical applications, the perturbation δ j (i) is added to the j-th feature data of the i-th point trace of the weighted first drone sample to obtain the value of the j-th feature data of the i-th point trace of the second drone sample. The formula is as follows:

[0094] Y j (i) = X3 j (i)+δ j (i);

[0095] Among them, Y j (i) represents the value of the j-th feature data of the i-th point trace of the second drone sample.

[0096] S3. Divide the radar data of the non-UAV track according to the preset track length to obtain multiple non-UAV samples.

[0097] In the embodiment of the present application, the radar data of the non-UAV track can be obtained by radar collection of non-UAV targets, such as various flying birds. The radar data of the non-UAV track includes the distance R, pitch angle θ, azimuth φ and radial velocity V of each point on the non-UAV track.

[0098] The method of segmenting the radar data of the non-UAV track is the same as the method of segmenting the radar data of the UAV track, both of which are segmented according to the preset track length L, and the embodiment of the present application will not be repeated here.

[0099] S4. Randomly select a second drone sample that is a preset multiple of the number of non-drone samples as the third drone sample.

[0100] It is understandable that since non-UAV targets are easier to collect, the number of non-UAV tracks far exceeds the number of UAV tracks; in fact, when the radar collects radar data, all tracks except UAV tracks are non-UAV tracks. Therefore, the number of UAV tracks and non-UAV tracks is extremely unbalanced, which will have a negative impact on the recognition effect of the existing methods.

[0101] In order to solve the problem of imbalance between positive and negative samples, the embodiment of the present application randomly extracts m times the number of non-drone samples as the third drone samples, where m is a preset positive number. Random extraction refers to uniformly randomly extracting floor(m×Ω) second drone samples without replacement, where Ω is the number of non-drone samples and floor() represents a rounding function. In this embodiment, m is 1, so that the sample size of different categories is close to 1:1, thereby solving the problem of sample imbalance.

[0102] S5. Construct an identification sample set based on the first drone sample, the third drone sample, and the non-drone sample, and train a drone identification model based on the identification sample set.

[0103] In the embodiment of the present application, the drone identification model is obtained by training a neural network model. The last layer of the neural network model is a fully connected classification layer. The classification layer is used to output the probability of belonging to a drone using a Sigmoid function. The loss function of the neural network model is a cross entropy loss, a weighted cross entropy loss, or a FocalLoss. For the structure of the neural network model of the embodiment of the present application, please refer to Figure 2 , Figure 2 In the example, ResBlock is a residual block, and ResBlock(6,8) means that the input channel is 6 and the output channel is 8.

[0104] In practical applications, the neural network model is trained using the recognition sample set. When the loss function of the neural network model is minimized and the accuracy meets the requirements, the drone recognition model can be obtained.

[0105] See also Figure 3 Based on the above technical solution, the present application embodiment provides a method for identifying a drone, the method comprising the following steps:

[0106] S6. Acquire radar data of the track of the target to be identified in real time, and divide it into windows according to a preset track length to obtain multiple window samples to be identified.

[0107] In practical applications, the radar data of the track of the target to be identified acquired in real time can be slidingly windowed according to a preset track length L and a step length N, where N is a preset positive integer.

[0108] S7. Input each window sample into the UAV identification model in turn, determine the identification result corresponding to each window sample, and judge whether the ratio of the number of window samples identified as UAVs in the identification result to the total number of window samples is greater than a first preset ratio. If so, determine that the target to be identified is a UAV; otherwise, determine that the target to be identified is not a UAV.

[0109] The UAV recognition model is obtained by training with the UAV recognition model training method described in the embodiments of the present application. In practical applications, the obtained windowed samples are respectively recognized and classified using the trained UAV recognition model. Next, the recognition results corresponding to each windowed sample are counted. If the ratio of the number Q1 of windowed samples determined to be UAVs to the total number Q of windowed samples exceeds a set first preset ratio K1, the track label to be determined is a UAV; otherwise, it is a non-UAV.

[0110] The embodiments of the present application further include: after determining that the target to be recognized is a UAV, continuously obtaining the windowed samples corresponding to the target to be recognized, and determining whether the ratio of the number of windowed samples recognized as UAVs in the recognition results to the total number of windowed samples is less than a second preset ratio. If so, the target to be recognized is re-determined as a non-UAV; otherwise, the determination result remains unchanged. The second preset ratio is less than the first preset ratio.

[0111] Specifically, the recognition results of the embodiments of the present application are continuously updated as the radar operates, that is, the track of the target to be recognized will continuously lengthen as the radar operates. When a new windowed sample can be windowed from the lengthened track part, the windowed sample is recognized and classified using the trained UAV recognition model. If the recognition result of the windowed sample is a UAV, both the number Q1 of UAV windowed samples and the total number Q of windowed samples are incremented by 1; otherwise, only the total number Q of windowed samples is incremented by 1. If Q1 / Q is less than a second preset ratio K2 set to prevent the label from changing too quickly, where K2 < K1, the target to be recognized is updated to a non-UAV; otherwise, the target to be recognized remains a UAV unchanged.

[0112] In summary, the embodiments of the present application perform data augmentation on the radar data of the UAV track, mainly including random augmentation and adding random perturbations, to generate second UAV samples, thereby reducing the collection quantity of the radar data of the UAV track, lowering the cost of collecting and annotating the radar data, and improving the sample data collection efficiency. And the present application extracts third UAV samples according to the number of non-UAV samples, fundamentally solving the problem of unbalanced positive and negative samples. The embodiments of the present application use the augmented sample data to train the UAV recognition model, thereby improving the construction efficiency and accuracy of the UAV recognition model. Using the trained UAV recognition model for UAV recognition improves the efficiency and accuracy of UAV recognition. The embodiments of the present application can online and real-time determine and update whether the target detected by the radar is a UAV, and the result is more accurate as the target is detected for a longer time.

Claims

1. A method for training a drone recognition model, characterized in that: The method comprises: The radar data of the UAV track is divided according to a preset track length to obtain a plurality of first UAV samples; The feature data of each point trace on any two first drone samples are added correspondingly according to the same random weight, and then random perturbations are added to obtain multiple second drone samples; The radar data of the non-UAV track is divided according to the preset track length to obtain multiple non-UAV samples; randomly selecting a second drone sample that is a preset multiple of the number of non-drone samples as the third drone sample; An identification sample set is constructed based on the first drone sample, the third drone sample, and the non-drone sample, and a drone identification model is trained based on the identification sample set.

2. The method for training a drone identification model according to claim 1, characterized in that: The radar data of the UAV track and the radar data of the non-UAV track respectively include the distance, pitch angle, azimuth and radial velocity of each point on the track.

3. The method for training a drone recognition model according to claim 2, characterized in that: The feature data of each point trace on any two first drone samples are added together according to the same random weight, specifically including: Draw a random number from a Beta distribution as a random weight; For any two first UAV samples, the radar data of each point trace thereon are resolved into feature data, and each feature data is summed up accordingly based on a random weight to obtain a plurality of weighted first UAV samples; There are six types of characteristic data, including position components in the x, y, and z directions and radial velocity components in the x, y, and z directions in a spatial rectangular coordinate system.

4. The method for training a drone identification model according to claim 3, characterized in that: The calculation formula of the weighted first drone sample is as follows: X3(i)=[x3(i),y3(i),z3(i),Vx3(i),Vy3(i),Vz3(i)]=w*X1(i)+(1-w)*X2(i), 1≤i≤L; X1(i)=[x1(i),y1(i),z1(i),Vx1(i),Vy1(i),Vz1(i)]; X2(i)=[x2(i),y2(i),z2(i),Vx2(i),Vy2(i),Vz2(i)]; Among them, X3(i) represents the feature data of the i-th point on the weighted first drone sample, x3(i), y3(i), z3(i) represent the position components of the i-th point on the weighted first drone sample in the x, y, and z directions, Vx3(i), Vy3(i), Vz3(i) represent the radial velocity components of the i-th point on the weighted first drone sample in the x, y, and z directions, X1(i) represents the feature data of the i-th point on the first drone sample, x1(i), y1(i), z1(i) represent the position of the i-th point on the first drone sample in the x, y, and z directions, respectively. Components, Vx1(i), Vy1(i), Vz1(i) represent the radial velocity components of the i-th point on the first UAV sample in the x, y, and z directions, respectively. X2(i) represents the feature data of the i-th point on the second first UAV sample. x2(i), y2(i), z2(i) represent the position components of the i-th point on the second first UAV sample in the x, y, and z directions, respectively. Vx2(i), Vy2(i), Vz2(i) represent the radial velocity components of the i-th point on the second first UAV sample in the x, y, and z directions, respectively. w represents the random weight, and L represents the preset track length.

5. The method for training a drone identification model according to claim 3, characterized in that: The radar data of each point on the first UAV sample is solved into feature data. The solution formula is as follows: Among them, x n (i),y n (i),z n (i) represents the position components of the ith point trace on the nth first drone sample in the x, y, and z directions, respectively, Vx n (i),Vy n (i),Vz n (i) represents the radial velocity component of the ith point on the nth first UAV sample in the x, y, and z directions, respectively, R n (i) represents the distance of the i-th point trace on the n-th first drone sample, θ n (i) represents the pitch angle of the i-th point on the n-th first drone sample, represents the position of the ith point on the nth first drone sample, V n (i) represents the radial velocity of the i-th point on the n-th first UAV sample.

6. The method for training a drone identification model according to claim 3, characterized in that: Add random perturbations, including: Acquire feature data of corresponding points from all first UAV samples respectively, and construct feature data sets corresponding to each point respectively; Based on the feature data set, the sample mean and sample variance of each feature data of the corresponding point trace are calculated, and a normal distribution is constructed according to the sample mean and sample variance; Sampling from a normal distribution, and calculating the perturbation corresponding to each characteristic data of the corresponding point trace according to the sampling results; The perturbation is added to the corresponding feature data of the corresponding point trace of the weighted first drone sample to obtain a second drone sample.

7. The method for training a drone identification model according to claim 6, characterized in that: The calculation formula of the perturbation is as follows: The perturbation is added to the corresponding feature data of the corresponding point trace of the weighted first drone sample, and the formula is as follows: Y j (i)=X3 j (i)+δ j (i); Among them, δ j (i) represents the perturbation corresponding to the jth characteristic data of the i-th point trace, ∈ j (i) represents the sampling result of the jth feature data of the i-th point, μ j (i) represents the sample mean of the jth feature data of the i-th point, X3 j (i) represents the value of the jth feature data of the i-th point on the weighted first drone sample, J represents the number of feature data types, J=6, k j Indicates the preset parameters, Y j (i) represents the value of the j-th feature data of the i-th point trace of the second drone sample.

8. The method for training a drone identification model according to claim 1, characterized in that: The drone identification model is obtained by training a neural network model. The last layer of the neural network model is a fully connected classification layer. The classification layer is used to output the probability of belonging to the drone using a Sigmoid function. The loss function of the neural network model is a cross entropy loss, a weighted cross entropy loss or a FocalLoss.

9. A method for identifying a drone, characterized in that: The method comprises: Acquire radar data of the track of the target to be identified in real time, and divide it into windows according to the preset track length to obtain multiple window samples to be identified; Input each window sample into the drone recognition model in turn, determine the recognition result corresponding to each window sample, and judge whether the ratio of the number of window samples identified as drones in the recognition result to the total number of window samples is greater than a first preset ratio. If so, determine that the target to be identified is a drone; otherwise, determine that the target to be identified is not a drone. The drone recognition model is obtained by training using the drone recognition model training method according to any one of claims 1 to 8.

10. The drone identification method according to claim 9, characterized in that: The method further comprises: After determining that the target to be identified is a drone, continue to obtain window samples corresponding to the target to be identified, and determine whether the proportion of the number of window samples identified as drones in the identification result to the total number of window samples is less than a second preset proportion. If so, the target to be identified is re-determined as a non-drone. Otherwise, the determination result remains unchanged, and the second preset proportion is less than the first preset proportion.

Images