Unmanned aerial vehicle detection method and apparatus

Abstract

Embodiments of the present application provide a kind of unmanned aerial vehicle detection method and device, wherein method is realized based on radar, comprising: obtaining the reflection signal of target to be detected;Doppler spectrum set is determined based on the reflection signal;Based on the Doppler spectrum set, the tracking path of target to be detected is determined;Based on the tracking path, the Doppler spectrum set is extracted, and the Doppler-time graph corresponding to the target to be detected is obtained;The Doppler-time graph is input into unmanned aerial vehicle detection model, and the detection result output by the unmanned aerial vehicle detection model is obtained;Wherein, the unmanned aerial vehicle detection model is obtained after training based on Doppler time graph training sample and the label corresponding to Doppler time graph training sample, and the label is determined based on the periodic micro-motion feature in Doppler time graph training sample.This embodiment of the present application improves the accuracy of unmanned aerial vehicle detection.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and apparatus for detecting unmanned aerial vehicles (UAVs). Background Technology

[0002] With the increasing popularity of small drones, some drones do not comply with flight rules, which can affect social life, such as causing the leakage of others' privacy or flying in no-fly zones.

[0003] However, since drones are usually very small and difficult to detect by the human eye, drone detection has become an extremely important and urgent issue. Summary of the Invention

[0004] This invention provides a method and apparatus for detecting unmanned aerial vehicles (UAVs), which addresses the shortcomings of existing technologies in that UAVs are difficult to detect.

[0005] Firstly, a method for detecting unmanned aerial vehicles (UAVs), based on radar, includes:

[0006] Acquire the reflection signal of the target to be detected;

[0007] Determine the Doppler spectrum set based on the reflected signal;

[0008] The tracking path of the target to be detected is determined based on the Doppler spectrum set;

[0009] Based on the tracking path, the Doppler spectrum set is extracted to obtain the Doppler-time map corresponding to the target to be detected;

[0010] The Doppler-time map is input into the UAV detection model to obtain the detection results output by the UAV detection model;

[0011] The drone detection model is obtained by training on Doppler time map training samples and the corresponding labels of the Doppler time map training samples. The labels are determined based on the periodic micro-motion features in the Doppler time map training samples.

[0012] Optionally, determining the tracking path of the target to be detected based on the Doppler spectrum set includes:

[0013] Perform spectrum folding on each Doppler spectrum in the Doppler spectrum set to obtain the folding result for each Doppler spectrum;

[0014] The tracking path of the target to be detected is obtained based on the folding result.

[0015] Optionally, obtaining the tracking path of the target to be detected based on the folding result includes:

[0016] determine a maximum distance change value of the target to be detected in a time unit based on a maximum speed of the target to be detected, the time unit and a distance resolution of the radar;

[0017] determine a maximum cumulative folding result under a constraint of the maximum distance change value;

[0018] determine a maximum distance change value of the target to be detected in a time unit based on a maximum speed of the target to be detected, the time unit and a distance resolution of the radar;

[0019] Optionally, before the folding result is used to obtain the tracking path of the target to be detected, the method further comprises:

[0020] processing the folding result based on a spectrum subtraction method to obtain a folding result after noise reduction.

[0021] Optionally, the method further comprises:

[0022] correcting the tracking path based on a particle filter algorithm to obtain a corrected tracking path.

[0023] Optionally, the method further comprises:

[0024] removing direct current noise in the Doppler-time graph to obtain a Doppler-time graph after noise reduction.

[0025] Optionally, the method further comprises:

[0026] aligning a periodic micro-motion feature center in each Doppler spectrum with a center of the Doppler spectrum one by one in the Doppler-time graph.

[0027] Optionally, the method further comprises:

[0028] segmenting the Doppler-time graph to obtain a plurality of Doppler-time sub-graphs;

[0029] determining a maximum folding result corresponding to each Doppler-time sub-graph, the maximum folding result being a maximum value in the folding result corresponding to the Doppler spectrum in the Doppler-time sub-graph;

[0030] comparing the maximum folding result with a first threshold value one by one, and retaining a Doppler-time sub-graph corresponding to a maximum folding result not less than the first threshold value;

[0031] the inputting the Doppler-time graph into the unmanned aerial vehicle detection model comprises:

[0032] inputting the retained Doppler-time sub-graph into the unmanned aerial vehicle detection model.

[0033] In a second aspect, the present application provides a UAV detection system based on radar, comprising:

[0034] an acquisition unit configured to acquire a reflection signal of a target to be detected;

[0035] a feature extraction unit configured to determine a Doppler spectrum set based on the reflection signal;

[0036] a tracking unit configured to determine a tracking path of the target to be detected based on the Doppler spectrum set;

[0037] an extraction unit configured to perform spectrum extraction on the Doppler spectrum set based on the tracking path to obtain a Doppler-time graph corresponding to the target to be detected;

[0038] an identification unit configured to input the Doppler-time graph into a UAV detection model to obtain a detection result output by the UAV detection model;

[0039] wherein the UAV detection model is obtained by training based on Doppler-time graph training samples and labels corresponding to the Doppler-time graph training samples, and the labels are determined based on periodic micro-motion features in the Doppler-time graph training samples.

[0040] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the UAV detection method of the first aspect when executing the program.

[0041] In a fourth aspect, the present application provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the UAV detection method of the first aspect.

[0042] In a fifth aspect, the present application provides a computer program product comprising a computer program, wherein the computer program is executable on a processor to implement the UAV detection method of the first aspect.

[0043] The unmanned aerial vehicle detection method and device provided by the embodiment of the present application use a radar to receive a reflection signal of a to-be-detected target, thereby reducing the cost of unmanned aerial vehicle detection; whether the to-be-detected target is an unmanned aerial vehicle is detected by whether there is a periodic micro-motion feature in the Doppler spectrum of the to-be-detected target, the periodic micro-motion feature does not change with time and can stably exist between different types of unmanned aerial vehicles, thereby improving the accuracy of unmanned aerial vehicle detection; and the periodic micro-motion feature is caused by periodic micro-motion (such as rotation of a propeller) of the unmanned aerial vehicle, so that the unmanned aerial vehicle does not need to actively provide a signal to the radar, the radar can actively receive the reflection signal of the to-be-detected target, thereby realizing passive detection of the unmanned aerial vehicle; and based on the characteristic that the radar can receive high-altitude signals, the unmanned aerial vehicle detection method provided by the embodiment of the present application can detect high-altitude unmanned aerial vehicles. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0045] Figure 1 is a schematic diagram of the periodic micro-motion feature perception model of the unmanned aerial vehicle provided by the embodiment of the present application;

[0046] Figure 2 is a distance-Doppler spectrum schematic diagram of a six-rotor unmanned aerial vehicle at different altitudes provided by the embodiment of the present application;

[0047] Figure 3 is one of the flow schematic diagrams of the unmanned aerial vehicle detection method provided by the embodiment of the present application;

[0048] Figure 4 is a schematic diagram of the folding process of spectrum folding provided by the embodiment of the present application;

[0049] Figure 5 is the second flow schematic diagram of the unmanned aerial vehicle detection method provided by the embodiment of the present application;

[0050] Figure 6 is the flow schematic diagram of the unmanned aerial vehicle identification provided by the embodiment of the present application;

[0051] Figure 7 is the structural schematic diagram of the unmanned aerial vehicle detection system provided by the embodiment of the present application;

[0052] Figure 8 is the structural schematic diagram of the electronic device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0054] The technical terminology and background of this invention are introduced below:

[0055] This invention reveals significant limitations in the detection methods of related technologies for drone detection. Specifically, sound-based drone detection methods are susceptible to interference from complex environmental noise. Furthermore, drone sound attenuates rapidly in the air, resulting in a limited detection range for sound-based methods. When drones employ noise reduction techniques, detection accuracy further decreases. Vision-based drone detection methods can operate when drones are within sight, but their accuracy and reliability are easily affected by lighting conditions and visual background. Low detection accuracy persists even under poor line-of-sight, lighting, and visual background conditions. While thermal imaging and infrared cameras are possible options for vision-based drone detection, they are expensive and have limited coverage. Radio frequency signal-based drone detection technologies require specialized instruments to capture and analyze the communication of non-cooperative drones. Traditional radar is very expensive and power-intensive. In short, existing methods cannot balance accuracy and low cost, making them unsuitable for drone detection in everyday scenarios.

[0056] Doppler spectrum of drones:

[0057] Figure 1 This is a schematic diagram of the UAV periodic micro-motion feature perception model provided in an embodiment of the present invention, as shown below. Figure 1 As shown, Figure 1 The diagram illustrates the deployment of radar when identifying drones, such as... Figure 1 As shown, the radar is deployed on the ground with its antenna array pointing upwards; the radar can be a millimeter-wave radar. When a drone enters the radar's sensing area, the radar emits a transmission signal S. TX (t) is reflected by the drone, and the radar receives the received signal S. RX(t), it should be understood that the signal reflected by the UAV includes the reflected signal from the rotor and the reflected signal from the body, and the received signal received by the radar can also include the reflected signal from other targets in addition to the reflected signal from the UAV. Specifically, the millimeter wave radar transmits a frequency-modulated continuous wave (FMCW) signal for distance estimation and speed measurement. According to the difference in signal propagation time, the radar can extract the reflected signal corresponding to the UAV from the received signal.

[0058] Taking the UAV as an example of a target reflection, in order to facilitate understanding of the present application, the following will be introduced taking the case where the influence of the micro-motion of the UAV is ignored and the signal received by the radar only includes the reflected signal from the body of the UAV as an example.

[0059] The distance R(t) between the radar and the UAV can be obtained by calculating the beat frequency signal s(t) obtained by mixing the transmitted signal and the received signal:

[0060]

[0061] where s(t) represents the beat frequency signal calculated by the radar, represents the transmitted signal of the radar, S RX (t) represents the received signal of the radar, f C represents the starting frequency of the transmitted signal S RX (t), K represents the frequency variation rate of the transmitted signal S RX (t), a is the propagation attenuation coefficient, and c represents the speed of light.

[0062] The distance Fourier operation on the beat frequency signal can obtain the reflected signal S(t) of the UAV:

[0063] S(t) = aexp[j4πf c R(t) / c]

[0064] The velocity Fourier operation on the reflected signal S(t) of the UAV can obtain the Doppler spectrum S(f) of the reflected signal of the UAV:

[0065] S(f) = FFT(S(t)) = aδ(f-2vf c / c)

[0066] where FFT represents the velocity Fourier, v is the radial velocity of the UAV relative to the radar, and δ() represents the unit impulse function. It can be seen from the formula that the peak value in the Doppler spectrum corresponds to the speed of the target.

[0067] When considering a flying drone as the target and its periodic micro-motions, assuming the drone has Q rotors and each rotor blade contains P scattering points, the drone's reflected signal S(t) can be viewed as the superposition of the body's reflected signal and the reflected signals from these scattering points:

[0068]

[0069] Where, γ pqm The m-th frequency attenuation coefficient represents a single scattering point.

[0070] Periodic micro-motions of drones:

[0071] When a drone is in flight, it constantly undergoes periodic micro-movements, such as rotor rotation or body vibration.

[0072] Characteristics of periodic micro-motions:

[0073] This invention has revealed that the periodic micro-motions of a drone cause stable and continuous periodic characteristics in its reflected signals. Specifically, these micro-motions periodically modulate the frequency of the reflected signal, resulting in a series of peaks in the frequency of the reflected signal. This signal characteristic can be called Periodic Micro-Motion (PMM). PMM features are distinct from environmental noise and can help radar track and identify drones.

[0074] The characteristics of periodic micro-motions satisfy the following formula:

[0075]

[0076] Where PeakPos represents the peak position of the periodic peak in the PPM feature, and PeakValue represents the peak value of the periodic peak in the PPM feature. For example, as shown... Figure 1 As shown, the Doppler spectrum of the UAV exhibits periodic micro-motion characteristics that conform to the above formula.

[0077] Figure 2 This is a distance-Doppler spectrum diagram of a six-rotor UAV at different altitudes provided in an embodiment of the present invention, as shown below. Figure 2 As shown, a millimeter-wave radar is horizontally deployed on the ground to detect a hexacopter drone. The drone is remotely controlled to ascend horizontally to 48 meters at a speed of 1.5 meters per second, and then hovers at high altitude. In this embodiment of the invention, two reflected signal segments are randomly selected from the drone's ascent and hovering states, respectively, and their range-Doppler spectra are as follows. Figure 2As shown in the figure: the horizontal line part represents the Doppler spectrum corresponding to the distance of the UAV, and the PMM feature map is obtained after amplification. It can be seen that the PMM feature is stable at different altitudes and motion postures. When the UAV is far away from the radar, the peak value of the PMM feature will decay, but the periodicity is still stable.

[0078] The embodiment of the application finds that the PMM feature is stable during the flight of the UAV. The structure and position of the UAV only affect the peak value in the PMM feature, and do not affect the interval between the peaks. Such a feature can be used for the identification of multiple types of UAVs.

[0079] Therefore, the embodiment of the application provides a radar-based long-distance UAV detection method, which collects and detects the periodic micro-motion feature of the UAV in flight through the radar, and realizes the technology of long-distance UAV tracking and identification.

[0080] The UAV detection method provided by the embodiment of the application will be described below. Figures 3-6 The UAV detection method provided by the embodiment of the application will be described below.

[0081] Figure 3 is one of the flowcharts of the UAV detection method provided by the embodiment of the application, as Figure 3 shown, the UAV detection method provided by the embodiment of the application is realized based on a radar, and includes the following steps:

[0082] Step 310: acquiring a reflection signal of a to-be-detected target;

[0083] Specifically, the to-be-detected target refers to a target object that needs to be judged whether it is a UAV, and the reflection signal refers to a signal reflected by the to-be-detected target and received by the radar. The signal can be a signal in a period of time, and the radar can be a millimeter wave radar. The reflection signal of the to-be-detected target can be extracted from the received signal according to related technologies, which will not be described here.

[0084] Step 320: determining a Doppler spectrum set based on the reflection signal;

[0085] Specifically, the Doppler spectrum set can include multiple positioning-Doppler spectrums, and the positioning-Doppler spectrum is a range-Doppler spectrum or an angle-Doppler spectrum. Each positioning-Doppler spectrum corresponds to a time unit, and each positioning-Doppler spectrum includes multiple Doppler spectrums. The multiple Doppler spectrums included in one positioning-Doppler spectrum correspond to different positioning groups, and the positioning group is used to represent the relative position of the to-be-detected target and the radar.

[0086] Specifically, the time unit can refer to a time unit of the reflection signal sample, and the time unit can be a time point. The positioning group can be a distance group or an angle group, which can be divided into R distance groups according to the distance from the radar or R angle groups according to the angle relative to the radar. A plurality of Doppler spectra corresponding to the same time unit can constitute a distance-Doppler spectrum or an angle-Doppler spectrum. For the distance-Doppler spectrum, it means that each Doppler spectrum corresponds to a distance group; for the angle-Doppler spectrum, it means that each Doppler spectrum corresponds to an angle group.

[0087] For the distance group, in one embodiment, the radar has a detectable distance of 100 meters, and the positioning group can be divided into 10 distance groups, the first distance group represents a distance of 0-10 meters from the radar, the second distance group represents a distance of 10-20 meters from the radar, and the like in 10-meter intervals.

[0088] For the angle group, in one embodiment, the angle is 360 degrees, and the positioning group can be divided into 10 angle groups, the first angle group represents an angle of 0-36 degrees from the radar, the second angle group represents an angle of 36-72 degrees from the radar, and the like in 36-degree intervals.

[0089] It should be understood that, in the case of a distance-Doppler spectrum, the positioning group is a distance group; in the case of an angle-Doppler spectrum, the positioning group is an angle group.

[0090] For determining the distance-Doppler spectrum based on the reflection signal of the target to be detected, a distance-Doppler FFT operation can be performed on the reflection signal to obtain the distance-Doppler spectrum, which can be referred to in the above description and will not be repeated here. For determining the angle-Doppler spectrum based on the reflection signal of the target to be detected, a beam forming (BF) algorithm can be applied to the reflection signal of all antennas corresponding to the target to be detected to obtain the signal of each angle. Then, a Doppler FFT operation is performed on the signal of each angle to obtain the angle-Doppler spectrum.

[0091] In order to facilitate understanding of the embodiments of the present application, the distance-Doppler spectrum or the angle-Doppler spectrum will be used as an example in the present application, and the UAV detection method suitable for the distance-Doppler spectrum can also be suitable for the angle-Doppler spectrum, which will not be repeated hereinafter.

[0092] In order to facilitate understanding of the distance-Doppler spectrum in the embodiments of the present application, taking the determination of the distance-Doppler spectrum based on the reflection signal of the target to be detected obtained in one time point as an example, the distance-Doppler spectrum is represented as {D1, D2,..., D R} represents the distance-Doppler spectrum, D irepresents the Doppler spectrum corresponding to the i th distance group, and R represents the number of distance groups. In the embodiment of the present application, the Doppler spectrum in the reflection signal of the to-be-detected target in each time unit can be extracted continuously, and thus each time unit can correspond to a distance-Doppler spectrum. For example, the distance-Doppler spectrum corresponding to the first moment is: {D 1，1 , D 2，1 ,..., D R，1}, the distance-Doppler spectrum corresponding to the T th moment is: {D 1，T , D 2，T ,..., D R，T}, and D i，t represents the Doppler spectrum corresponding to the i th distance group and the t th moment.

[0093] In step 330, the tracking path of the to-be-detected target is determined based on the set of Doppler spectra.

[0094] Specifically, the tracking path of the to-be-detected target can represent the position of the to-be-detected target relative to the radar in each time unit. For example, the unmanned aerial vehicle is located at a direction of 4 degrees relative to the radar at the first moment and is located at a direction of 16 degrees relative to the radar at the second moment.

[0095] In one possible implementation, as described above, the periodic micro-motion feature can help the radar to track and identify the unmanned aerial vehicle. Therefore, the tracking path of the to-be-detected target can be determined by analyzing the maximum PMM feature path in the set of Doppler spectra of the to-be-detected target, that is, when the to-be-detected target has multiple path possibilities, the path with the maximum PMM feature is taken as the tracking path.

[0096] For example, the distance-Doppler spectrum corresponding to the first moment is: {D 1，1 , D 2，1 ,..., D R，1}, the distance-Doppler spectrum corresponding to the second moment is: {D 1，2 , D 2，2 ,..., D R，2}, and the distance-Doppler spectrum corresponding to the third moment is: {D 1，3 , D 2，3 ,..., D R，3}. In D 2，1 , D 2，2 , and D R，3 , the PMM feature can be present, and thus the tracking path of the to-be-detected target is determined to be: the to-be-detected target is located in the second distance group at the first moment, is located in the second distance group at the second moment, and is located in the R th distance group at the third moment.

[0097] Step 340, extracting the Doppler spectrum set based on the tracking path to obtain the Doppler-time graph corresponding to the to-be-detected target;

[0098] Specifically, the Doppler spectrum corresponding to the tracking path in each time unit is extracted, and the Doppler-time graph corresponding to the to-be-detected target can be obtained.

[0099] Taking the example of the embodiment in step 330, the Doppler spectrum D 2，1 corresponding to the first time, the second distance group, the Doppler spectrum D 2，2 corresponding to the second time, the second distance group, and the Doppler spectrum D R，3 corresponding to the third time, the Rth distance group are extracted, and the Doppler-time graph can be obtained, wherein each time unit in the Doppler-time graph corresponds to a Doppler spectrum graph.

[0100] Step 350, inputting the Doppler-time graph into the UAV detection model to obtain the detection result output by the UAV detection model;

[0101] The UAV detection model is obtained by training based on the Doppler-time graph training sample and the label corresponding to the Doppler-time graph training sample, and the label is determined based on the periodic micro-motion feature in the Doppler-time graph training sample.

[0102] Specifically, the UAV detection model can be a long short-term memory (LSTM). The number of Doppler-time graph training samples can be multiple, and the Doppler-time graph training sample can correspond to multiple time units, each time unit corresponding to a Doppler spectrum. When the Doppler-time graph training sample contains continuous PMM features, the corresponding label is a UAV, and when the Doppler-time graph training sample does not contain continuous PMM features, the corresponding label is a non-UAV.

[0103] When the to-be-detected target is a UAV, the Doppler-time graph corresponding to the to-be-detected target will contain unique and continuous PMM features, and the UAV detection model can distinguish whether the to-be-detected target is a UAV or other objects according to the PMM features in the Doppler-time graph.

[0104] In a possible implementation, since identifying the UAV from other objects is a binary classification problem, the UAV detection model in the embodiment of the present application can solve the identification problem of the UAV based on an LSTM-based detector. The LSTM network is suitable for processing data sequences, has excellent performance in identification tasks, and considering that each time unit in the Doppler-time graph contains L Doppler spectrum groups, the input dimension of the UAV detection model (LSTM network) in the embodiment of the present application is set to L. The UAV detection model contains two LSTM layers, and the hidden state size is set to 128; a fully connected layer is used to map the hidden state to the detection result, that is, the UAV or other objects.

[0105] Optionally, the Doppler-time graph can be divided into multiple segments, which can be referred to as Doppler-time sub-graphs, and the multiple segments are input into the UAV detection model.

[0106] Optionally, a cross-entropy loss function can be selected to train the UAV detection model.

[0107] Optionally, the cross-entropy loss function is:

[0108]

[0109] where N represents the number of Doppler-time graph training samples, y i represents the label corresponding to the Doppler-time graph training sample i, and the label corresponding to the UAV can be 1, and the label corresponding to the non-UAV can be 0. i represents the probability that the Doppler-time graph training sample i is predicted as the UAV.

[0110] It should be understood that the specific values of L and the hidden state size can be determined according to actual applications (such as computer processing capacity) and the like.

[0111] The UAV detection method provided by the embodiment of the present application uses the radar to receive the reflection signal of the to-be-detected target, thereby reducing the cost of UAV detection; whether the to-be-detected target is a UAV is detected by whether there is a periodic micro-motion feature in the Doppler spectrum of the to-be-detected target, and the periodic micro-motion feature does not change with time and can stably exist between different types of UAVs, thereby improving the accuracy of UAV detection; and the periodic micro-motion feature is caused by the periodic micro-motion (such as the rotation of the propeller) of the UAV, so that the to-be-detected target does not need to actively provide a signal to the radar, and the radar can actively receive the reflection signal of the to-be-detected target, thereby realizing passive detection of the UAV; and based on the characteristic that the radar can receive high-altitude signals, the UAV detection method provided by the embodiment of the present application can detect high-altitude UAVs.

[0112] The following further describes possible implementations of the above steps in specific embodiments.

[0113] Optionally, step 330, determining a tracking path of the target to be detected based on the set of Doppler spectra, comprises:

[0114] Step 331, performing spectrum folding on each Doppler spectrum in the set of Doppler spectra to obtain a folding result corresponding to each Doppler spectrum.

[0115] Considering the unpredictability of the propeller rotation speed, the peak periods of the PMM feature are uncertain and variable. Due to the low signal-to-noise ratio of the reflected signal, these periodic peaks can even be masked by the environmental noise. In the embodiments of the present application, spectrum folding can be used to amplify the difference between the PMM feature and the environmental noise.

[0116] Specifically, the PMM feature is extracted and enhanced by performing spectrum folding on the Doppler spectrum D i，t . Figure 4 is a schematic diagram of the folding process of the spectrum folding provided by the embodiments of the present application, and the folding process is as shown in Figure 4 , a Doppler spectrum D i，t may also be divided into L Doppler groups, Figure 4 and the values in the Doppler groups are represented by boxes. The black boxes represent the periodic peaks in the PMM feature, and the other box types represent noise. Figure 4 In the above example, the interval λ between adjacent peaks in the Doppler spectrum is 5, and the number of Doppler frequency groups L is 20. If the Doppler spectrum is folded accurately into a matrix with j = λ columns, the periodic peaks will be aligned in a column, and a significantly enhanced peak will appear in the average result of each column. In the embodiments of the present application, the maximum value in the column average result is referred to as the folding value, and the number of folding columns and the folding size are respectively j and j. When the folding size j is not equal to λ, the folding value will decrease rapidly due to the misalignment of the periodic peaks.

[0117] The embodiments of the present application find that the interval between adjacent peaks in the PMM feature is determined by the propeller rotation speed, and therefore has unpredictable and time-varying characteristics. Optionally, to find the correct folding size, all integers in an empirical range [j min , j max ] can be traversed, and the corresponding folding values are calculated, where j min and j max represent the minimum and maximum folding sizes respectively. According to empirical experiments, j min may be set to 2, and j max may be set to 20. When the folding size is j, the folding value of the Doppler spectrum D i，t can be calculated by the following formula:

[0118]

[0119] wherein k and m represent column index and row index in the folding matrix respectively, the number of rows in the folding matrix is represented by M, and the folding matrix can be calculated by .

[0120] The maximum folding value is selected as the folding result, and the folding result can be understood as representing the average value of all peaks of PMM. The folding result of Doppler spectrum D i，t can be obtained by the following formula:

[0121]

[0122] In this way, the folding result of each Doppler spectrum can be calculated by the embodiment of the application. Since the PMM feature of the UAV has stable periodicity, the folding result is much larger than the folding result of random environmental noise. Therefore, the folding result can be further used for UAV tracking and detection.

[0123] In step 332, a tracking path of the target to be detected is obtained based on the folding result. Specifically, since the folding result of the Doppler spectrum at each time is continuously extracted, and the folding result can be used to represent the PMM feature, a plurality of folding results can form a range-time-PMM graph (hereinafter referred to as R-PMM). For the convenience of understanding the embodiment of the application, an example in the form of a matrix is given. The first matrix corresponding to an R-PMM is:

[0124] P 1，1 P 1，2 P 1，3

[0125] P 2，1 P 2，2 P 2，3

[0126] P 3，1 P 3，2 P 3，3

[0127] wherein each row represents the folding result corresponding to the same distance group, each column represents the folding result corresponding to a time unit, and each folding result corresponds to a Doppler spectrum.

[0128] Since the Doppler spectrum corresponding to the UAV has a larger folding result relative to noise at each time unit, the UAV can be tracked by searching for the maximum path in the folding result determined by the Doppler spectrum, wherein the maximum path refers to the cumulative maximum value of the folding result in the path. It should be understood that the maximum path is the tracking path of the target to be detected.

[0129] For the convenience of understanding the embodiments of the present application, the above matrix is valued as an example:

[0130] P 1，1 = 1 P 1，2 = 3 P 1，3 = 2

[0131] P 2，1 = 2 P 2，2 = 4 P 2，3 = 0

[0132] P 3，1 = 0 P 3，2 = 1 P 3，3 = 3

[0133] The maximum folding result in the first column (that is, in the first time unit) is P 2，1 = 2, the maximum folding result in the second column is P 2，2 = 4, and the maximum folding result in the third column is P 3，3 = 3, and the maximum path, that is, the tracking path, is P 2，1 , P 2，2 and P 3，3 , the corresponding Doppler spectrum is extracted according to the maximum path.

[0134] The unmanned aerial vehicle detection method provided by the embodiments of the present application realizes the extraction and enhancement of the PMM feature by performing spectrum folding on the Doppler spectrum, avoids that the periodic peak value in the PMM feature is covered by environmental noise, is beneficial to improving the identification accuracy of the unmanned aerial vehicle in the process of identifying the unmanned aerial vehicle according to the PMM feature, and determines the path in which the unmanned aerial vehicle is most likely to exist according to the folding result, without performing Doppler spectrum extraction and detection on each path, so that the unmanned aerial vehicle is efficiently detected.

[0135] Since the intensity of the Doppler spectrum is proportional to the intensity of the reflected signal, the Doppler spectrum in the closer distance group has stronger intensity, which means that the folding result thereof is often larger. This causes the R-PMM to have different static background noise intensities in different distance groups.

[0136] Optionally, in order to reduce the influence of such noise, before the step 332 obtains the tracking path of the to-be-detected target based on the folding result, the method further includes:

[0137] The folding result is processed based on the spectrum subtraction method to obtain a denoised folding result.

[0138] The embodiments of the present application pre-process the R-PMM by using the spectrum subtraction method. The main idea is to subtract the estimation of the average background noise spectrum from the noisy R-PMM.

[0139] Specifically, taking the range-Doppler spectrum as an example, the average background noise spectrum can be estimated by the following formula:

[0140]

[0141] Wherein, N(r) represents the average background noise spectrum; N(r, t) represents the measured R-PMM in the background noise; T represents T range-Doppler spectra, that is, the total number of time units.

[0142] The influence of the background noise can be calculated as the normalized projection of the measured R-PMM on the background noise spectrum:

[0143]

[0144] Wherein, G(t) represents the influence of the background noise, that is, the normalized projection of the R-PMM on the background noise spectrum; R represents the total number of distance groups, r represents the rth distance group, N(r) represents the average background noise spectrum, is the Euclidean norm of the average background noise spectrum N(r), and S(r, t) is the R-PMM corresponding to the detected target measured, that is, the R-PMM to be denoised.

[0145] Finally, the background noise can be removed from the measured R-PMM:

[0146] S'(r, t) = S(r, t) - G(t)N(r)

[0147] Wherein, S'(r, t) represents the denoised R-PMM, which contains the denoised folding result, S(r, t) represents the R-PMM to be denoised, G(t) represents the influence of the average background noise, and N(r) represents the average background noise spectrum.

[0148] It should be understood that after the folding result is processed based on the spectrum subtraction method, the folding result in step 322 is the denoised folding result.

[0149] The unmanned aerial vehicle detection method provided by the embodiment of the application removes the average background noise in the initial R-PMM through the spectrum subtraction method, obtains the denoised folding result, determines the maximum path corresponding to the detected target through the denoised folding result, and can further improve the accuracy of the maximum path, thereby improving the accuracy of the unmanned aerial vehicle detection.

[0150] The folding result can represent the PMM feature of the to-be-detected target, which can be a UAV. Due to the unpredictability of UAV movement, the folding result can change rapidly. On the other hand, due to the complexity of the environment and the imperfection of the hardware, there can still be local dynamic noise near the UAV trajectory in the R-PMM. Both of these two cases can cause the tracking accuracy to decrease.

[0151] Therefore, the example of the present application determines the tracking path based on the continuity of the UAV running trajectory and the constraint of the movement distance within one time unit.

[0152] Optionally, in step 332, the tracking path of the to-be-detected target is obtained based on the folding result, comprising:

[0153] In step 3321, the maximum distance change value of the to-be-detected target within one time unit is determined based on the maximum speed of the to-be-detected target, the time unit and the distance resolution of the radar.

[0154] Specifically, due to the continuity of the UAV running trajectory, the running range of the UAV within a period of time changes limitedly, so the maximum distance change value of the to-be-detected target within one time unit can be determined according to the following formula:

[0155]

[0156] Wherein, V max represents the maximum speed of the to-be-detected target, T d represents the duration of one time unit, R res represents the distance resolution of the radar, and K represents the maximum distance change value within one time unit.

[0157] In step 3322, the maximum cumulative folding result is determined under the constraint of the maximum distance change value.

[0158] Specifically, the path corresponding to the UAV (Unmanned Aerial Vehicle, UAV) trajectory in the R-PMM always satisfies this constraint. Therefore, the UAV tracking path problem is converted into a constrained maximum path problem in the R-PMM, in which the change of the UAV path in adjacent columns (i.e. within one time unit) will not exceed the maximum distance change value K.

[0159] This constrained maximum path finding problem can be described as:

[0160]

[0161] s.t. |g(t)-g(t-1)|≤K

[0162] Wherein denotes a path in R-PMM.g * denotes the maximum path under the constraint, T denotes the distance-Doppler spectrum number in R-PMM, that is, the number of time units, which can also be understood as the duration of R-PMM, S' denotes the R-PMM after noise reduction, and t denotes the t moment.

[0163] It should be understood that g refers to a path formed by selecting one distance group from each column in R-PMM to the last column, and there can be multiple g in one R-PMM. Taking the first matrix in the above text as an example, there can be 3x3x3=27 paths g. Among the 27 paths, the distance change between each adjacent moment satisfies the constraint of the maximum distance change value K, and the path with the maximum sum of the cumulative folding results is the tracking path g * .

[0164] Optionally, under the constraint of the maximum distance change value, the maximum cumulative folding result corresponding to each time unit and each positioning group is calculated.

[0165] Specifically, the maximum cumulative folding result can be determined by dynamic programming, and the score at (r, t) is defined as the maximum cumulative folding result under the constraint, which can be expressed as:

[0166]

[0167] wherein θ(r, t) denotes the maximum cumulative folding result under the constraint at r distance and t moment, K denotes the maximum distance change value, k∈[-K, K] denotes the traversal value within the constraint range [-K, K], denotes the maximum value of θ(r+k, t-1) within the range k∈[-K, K], θ(r+k, t-1) denotes the maximum cumulative folding result under the constraint at r+k distance and t-1 moment, and S'(r, t) denotes the folding result after noise reduction at r distance and t moment.

[0168] Since the flight trajectory of the unmanned aerial vehicle has the characteristic of continuity, the maximum cumulative folding result at t moment is determined by the maximum cumulative folding result at t-1 moment and the possible folding result at t moment in the θ(r, t) scoring formula. When the maximum cumulative folding result corresponding to a positioning group and a time unit is calculated, it is not necessary to calculate the cumulative folding results of all paths at this position, but only the maximum cumulative folding result at the previous moment and the possible folding result at the current moment are needed to determine the maximum cumulative folding result at the current moment, thereby improving the calculation efficiency.

[0169] Step 3323, taking the path corresponding to the maximum cumulative folding result as the tracking path.

[0170] Specifically, the tracking path is from the first time unit to the last time unit, and a positioning group is selected in each time unit, and each combination of the positioning group and the time unit (r, t) corresponds to a maximum score θ(r, t) max Therefore, when the positioning group with the maximum score of the last time unit in the plurality of positioning groups is found, the selection of the positioning group of the first time unit can be inversely deduced from T-1, T-2, and so on.

[0171] For the convenience of understanding the embodiments of the present application, the distance group is taken as an example. In order to obtain the tracking path, the embodiments of the present application first find the distance group with the maximum score in the last time unit, and then obtain the remaining part of the path by the following calculation method:

[0172]

[0173]

[0174] wherein g * (t) represents the distance group selected at the t th time in the tracking path, θ(g * (t+1)+k, t) represents the maximum cumulative folding result of the (t+1)+k distance groups at the t th time, g * (t+1) represents the distance group selected at the t+1 th time in the tracking path, and T represents the number of distance-Doppler spectrum, that is, the total number of time units.

[0175] This backtracking process provides a constrained maximum path g * In this way, the embodiments of the present application can obtain the constrained maximum path in the R-PMM, that is, the tracking path of the unmanned aerial vehicle, that is, the distance group selected in each time unit from the first time unit to the last time unit.

[0176] The unmanned aerial vehicle detection method provided by the embodiments of the present application can more accurately and quickly find the tracking path by constraining the maximum distance change value in one time unit in the determination process of the tracking path, improve the accuracy of the tracking path, and thus improve the accuracy and efficiency of the unmanned aerial vehicle detection.

[0177] Optionally, the method further comprises:

[0178] Step 333, correcting the tracking path based on the particle filtering algorithm to obtain a corrected tracking path.

[0179] Specifically, a particle filter (also referred to as a particle filtering) algorithm can be applied to the tracking result to reduce tracking errors. The particle filter algorithm can estimate a target state by combining observation and prediction. Specifically, a state variable in the particle filter algorithm includes a distance and a speed of a target to be detected, and the observation result is the tracking path obtained before step 333.

[0180] Optionally, the embodiment of the present application initializes 5000 uniformly distributed particles, and uses a polynomial resampling algorithm as a resampling method of the particles.

[0181] The unmanned aerial vehicle detection method provided by the embodiment of the present application reduces errors in the tracking path through particle filtering, facilitates subsequent extraction of a Doppler-time graph with higher accuracy, and thus improves the accuracy of unmanned aerial vehicle detection.

[0182] Optionally, the method further comprises:

[0183] Step 334: removing direct current noise in the Doppler-time graph to obtain a Doppler-time graph after noise reduction.

[0184] Specifically, the embodiment of the present application can remove direct current noise from the Doppler-time graph. A signal with a frequency of zero in a Doppler spectrum is a direct current component, and interference caused by the direct current component is direct current noise.

[0185] Considering that the main speed of the unmanned aerial vehicle is close to 0 when the unmanned aerial vehicle hovers, the corresponding direct current component contains a peak corresponding to the main speed. It should be understood that the main speed can be determined based on the maximum peak in the Doppler spectrum. In order to retain the peak corresponding to the main speed, the embodiment of the present application averages the direct current components in the Doppler spectrum when the main speed peak is not close to the direct current component. Then, the average value is subtracted from all direct current components to obtain a Doppler-time graph after noise reduction.

[0186] It should be understood that, for the main speed peak not close to the direct current component, it means that the main speed peak is not close to the direct current component when the unmanned aerial vehicle moves. Only when the unmanned aerial vehicle hovers, the direct current component is close to the direct current component, that is, when the unmanned aerial vehicle body speed is close to 0.

[0187] It should be understood that, after removing the direct current noise in the Doppler-time graph, the Doppler-time graph input into the unmanned aerial vehicle detection model is the Doppler-time graph after noise reduction.

[0188] The unmanned aerial vehicle detection method provided by the embodiment of the present application reduces errors in the Doppler-time graph by removing the interference of the direct current component, that is, the direct current component, and improves the accuracy of unmanned aerial vehicle detection.

[0189] Since the main velocity of the UAV determines the center of the PMM feature, but the PMM feature changes with the movement of the UAV. Therefore, the change of the PMM feature will hinder the use of the periodicity of the PMM feature by the embodiments of the present application.

[0190] Optionally, the method further comprises:

[0191] Step 335, in the Doppler-time graph, aligning the center of each periodic micro-motion feature with the center of the Doppler spectrum.

[0192] Specifically, therefore, the embodiments of the present application use a feature alignment algorithm on the Doppler-time graph to align each PMM feature center with the Doppler spectrum center. It should be understood that the center of the Doppler spectrum is the center position of the spectrum, for example, when the spectrum range is [-10, 10], the center of the Doppler spectrum is 0; the PMM feature center is the frequency position corresponding to the main velocity. When the main velocity peak in the Doppler spectrum is not near the direct current component, the embodiments of the present application shift the entire Doppler spectrum in the direction of the main velocity peak to the direct current component, and supplement by linear interpolation.

[0193] The UAV detection method provided by the embodiments of the present application avoids the interference of the change of the periodic micro-motion feature (such as the change of the peak position) on the recognition of the periodic micro-motion feature, and further avoids the interference of the change of the periodic micro-motion feature on the UAV detection, thereby improving the accuracy of the UAV detection and recognition.

[0194] Since there may be errors in the tracking path, especially when the UAV is at high altitude. Directly using the obtained Doppler-time graph is not always reliable.

[0195] Optionally, the method further comprises:

[0196] Step 336, segmenting the Doppler-time graph to obtain a plurality of Doppler-time partial graphs;

[0197] Specifically, in the case where the Doppler-time graph contains a PMM feature, in order to reduce the influence of the tracking path error, the embodiments of the present application divide the Doppler-time graph into fixed-length segments, i.e. Doppler-time partial graphs. It should be understood that one time unit in the Doppler-time graph corresponds to one Doppler spectrum, and the Doppler-time partial graph is a segment of the Doppler-time graph, so in the Doppler-time partial graph, one time unit also corresponds to one Doppler spectrum, i.e. there can be at least one Doppler spectrum in the Doppler-time partial graph.

[0198] Step 337, determine the maximum folding result corresponding to each Doppler-time part graph, the maximum folding result being the maximum value in the folding results corresponding to the Doppler spectrum in the Doppler-time part graph;

[0199] Specifically, at least one Doppler spectrum is contained in one Doppler-time part graph, each Doppler spectrum corresponding to a folding result, and the maximum value in the folding results corresponding to the Doppler-time part graph is taken as the maximum folding result of the Doppler-time part graph.

[0200] Exemplarily, one Doppler-time part graph P1 contains three time units of t3, t4 and t5, t3 corresponding to Doppler spectrum D3, t4 corresponding to Doppler spectrum D4, and t5 corresponding to Doppler spectrum D5, the folding result of Doppler spectrum D3 being 56, the folding result of Doppler spectrum D4 being 688, and the folding result of Doppler spectrum D5 being 24, then the maximum folding result of Doppler-time part graph P1 is 688.

[0201] Step 338, compare the maximum folding result with the first threshold value one by one, and retain the Doppler-time part graph corresponding to the maximum folding result not less than the first threshold value;

[0202] The inputting the Doppler-time graph into the UAV detection model comprises:

[0203] The retained Doppler-time part graph is inputted into the UAV detection model.

[0204] The maximum folding result of each Doppler-time part graph is compared with the first threshold value one by one, when the maximum folding result is less than the first threshold value, it is considered that the Doppler-time part graph does not contain PMM characteristics and is discarded.

[0205] The retained Doppler-time part graph is inputted into the UAV detection model, it should be understood that the form of the Doppler-time graph training sample can also be a Doppler-time graph segment, therefore the UAV detection model can identify the Doppler-time part graph.

[0206] Optionally, the first threshold value can be set to 30000.

[0207] Exemplarily, taking Doppler-time part graph P1 as an example, the maximum folding result of Doppler-time part graph P1 is 688, which is less than the first threshold value, and Doppler-time part graph P1 is discarded, and Doppler-time part graph P1 will not be inputted into the UAV detection model.

[0208] The unmanned aerial vehicle detection method provided by the embodiment avoids the interference of tracking path errors on unmanned aerial vehicle detection, and improves the accuracy of unmanned aerial vehicle detection and identification.

[0209] It should be understood that the above steps 333, 334, 335, and 336 to 338 can be executed alone or in combination.

[0210] The unmanned aerial vehicle detection method provided by the embodiment will be described below in combination with an embodiment.

[0211] The unmanned aerial vehicle detection method provided by the embodiment can realize long-distance unmanned aerial vehicle detection and identification.

[0212] Figure 5 is a flowchart of the unmanned aerial vehicle detection method provided by the embodiment, as shown in the figure, Figure 5 The unmanned aerial vehicle detection method provided by the embodiment includes feature extraction, unmanned aerial vehicle tracking, and unmanned aerial vehicle identification.

[0213] Feature extraction: the distance-Doppler FFT is used to continuously extract the distance-Doppler spectrum corresponding to the reflection signal of the target to be detected.

[0214] Unmanned aerial vehicle tracking: through feature extraction, the folding results in each distance-Doppler spectrum can be continuously extracted, and these results jointly form a distance-time-PMM (R-PMM) graph.

[0215] Correspondingly, for angle grouping, through feature extraction, the folding results in each angle-Doppler spectrum can be continuously extracted, and these results jointly form an angle-time-PMM (A-PMM) graph.

[0216] Unmanned aerial vehicle recognition: after obtaining the tracking path of the to-be-detected target, a Doppler spectrum is extracted at the position of the to-be-detected target according to the tracking path, and a Doppler-time graph is formed. After pre-processing the Doppler-time graph through direct current component removal and feature alignment, an embodiment of the present application inputs a Doppler-time picture segment containing a PMM feature into an unmanned aerial vehicle detection model based on a long short-term memory artificial neural network, and identifies whether the to-be-detected target is an unmanned aerial vehicle through the unmanned aerial vehicle detection model.

[0217] The following will take the case of the to-be-detected target being an unmanned aerial vehicle as an example to introduce each part in detail:

[0218] (I) Feature extraction:

[0219] The spectrum folding technique is used to amplify the difference between the PMM feature and the environmental noise. As described above, an embodiment of the present application can obtain a range-Doppler spectrum by performing a range-Doppler FFT on the reflected signal of the millimeter wave. Specifically, an embodiment of the present application uses {D1, D2,..., D R} to represent the Doppler spectrum in the range-Doppler spectrum. D i and R represent the number of Doppler spectra and range groups corresponding to the i-th range group, respectively. The number of Doppler frequency groups in each Doppler spectrum is L. When the unmanned aerial vehicle appears in the i-th range group, the PMM feature will appear in the corresponding Doppler spectrum D i . However, considering the unpredictability of the propeller rotation speed, the peak period of the PMM feature is uncertain and variable. Due to the low signal-to-noise ratio of the reflected signal, these periodic peaks can even be masked by the environmental noise.

[0220] An embodiment of the present application uses the spectrum folding technique to extract and enhance the PMM feature. The folding process is shown in Figure 4 , where the values within the Doppler group are represented by boxes. The black box represents the periodic peak in the PMM feature, and the other box types represent noise. The interval λ between adjacent peaks in the Doppler spectrum is 5, and the number of Doppler frequency groups L is 20. If an embodiment of the present application accurately folds the Doppler spectrum into a matrix with j = λ columns, the periodic peaks will be aligned in a column, and a significantly enhanced peak will appear in the average result of each column. For the sake of clarity, an embodiment of the present application refers to the maximum value in the column average result and the folding size as the folding value and the folding size, respectively. When the folding size j is not equal to λ, the folding value will decrease rapidly due to the misalignment of the periodic peaks.

[0221] It is worth noting that the interval between adjacent peak values in the PMM feature is determined by the propeller rotation speed, and therefore has the characteristics of being unpredictable and possibly changing at any time. In order to find the correct folding size, an embodiment of the present application traverses an empirical range [j min , j maxall integers in (1) and compute the corresponding folded values, where j min and j max represent the minimum and maximum folded size, respectively. According to extensive experimental experience of embodiments of the present application, they are set to 2 and 20. The folded value of the Doppler spectrum D i when the folded size is j can be computed by the following equation:

[0222]

[0223] where k and m represent the column index and row index in the folded matrix, respectively. The number of rows in the folded matrix is denoted by M, which can be computed by Embodiments of the present application select the maximum folded value as the folded result. The folded result of the Doppler spectrum D i can be obtained by the following equation:

[0224]

[0225] In this way, the folded result of each Doppler spectrum can be computed. Since the PMM feature of the UAV has stable periodicity, its folded result is much larger than that of random environmental noise. These folded results can be further used for UAV tracking and detection.

[0226] (ii) UAV tracking

[0227] First, the folded results are pre-processed by the spectrum subtraction method to reduce the influence of static background noise. Then, accurate UAV tracking is achieved through dynamic programming and particle filtering.

[0228] In order to estimate the UAV trajectory, the folded results at each time are extracted continuously, which form a range-time-PMM map, denoted as R-PMM. Assuming that there are T range-Doppler spectra {RD1, RD2,..., RDT} and each range-Doppler spectrum has R range bins, the value of R-PMM(r, t) represents the folded result of the rth range bin in the Tth range-Doppler spectrum. Since the Doppler spectrum corresponding to the UAV has a large folded result at each time, embodiments of the present application can track the UAV by searching for the maximum path in the R-PMM.

[0229] However, since the intensity of Doppler spectrum is proportional to the intensity of reflected signal, Doppler spectrum in closer distance bins has stronger intensity, which means their fold results tend to be larger. This leads to different static background noise intensity in R-PMM in different distance bins. To mitigate the impact of this noise, embodiments of the present application employ a spectral subtraction method to preprocess R-PMM. The main idea is to subtract the estimate of average background noise spectrum from the noisy R-PMM. Specifically, the average background noise spectrum can be estimated by the following formula:

[0230]

[0231] where N(r, t) is the measured R-PMM in background noise. The impact of background noise can be calculated as the normalized projection of the measured R-PMM on the background noise spectrum:

[0232]

[0233] where is the Euclidean norm of the noise spectrum, S(r, t) is the measured R-PMM in the tracking phase, and finally, the background noise can be removed from the measured R-PMM:

[0234] S'(r, t) = S(r, t) - G(t)N(r)

[0235] By this method, the static background noise related to distance can be removed, and the preprocessed R-PMM can be used to find the maximum path corresponding to the tracking result of the UAV.

[0236] Since the fold result represents the PMM feature of the UAV, due to the unpredictability of UAV movement, the fold result can change rapidly. On the other hand, due to the complexity of the environment and the imperfection of the hardware, there may still be local dynamic noise near the UAV trajectory in the R-PMM. Both of these situations will lead to a decrease in tracking accuracy.

[0237] Considering the continuity of the UAV running trajectory, its running range changes limitedly in a period of time. Assuming the maximum speed of the UAV is V max , the duration of the distance-velocity spectrum is T d , and the distance resolution of the radar is R res , then the maximum distance unit change of the UAV in adjacent columns in R-PMM is:

[0238]

[0239] The path corresponding to the UAV trajectory in the R-PMM always satisfies this constraint. Therefore, embodiments of the present application propose to transform the UAV tracking problem into a constrained maximum path problem in the R-PMM, where the change of the UAV path between adjacent columns will not exceed K. This constrained maximum path finding problem can be formulated as:

[0240]

[0241] s.t. |g(t) - g(t - 1)| < K

[0242] where g(r, t) denotes a path in the R-PMM. This problem can be solved by dynamic programming. Embodiments of the present application define the score at (r, t) as the constrained maximum cumulative fold result, which can be denoted as:

[0243] Since θ(r, t) considers both the continuity of the trajectory and the cumulative fold result of the previous trajectory, the process of calculating the score is able to obtain the constrained optimal trajectory through the (r, t) position. In order to obtain the complete constrained maximum path, embodiments of the present application first find the distance-time cell with the maximum score in the last column, and then obtain the rest of the path by the following calculation method:

[0244]

[0245]

[0246] This backtracking process provides the constrained maximum path g * In this way, embodiments of the present application are able to obtain the constrained maximum path in the R-PMM, that is, the tracking path of the UAV.

[0247] Further, the particle filter algorithm is applied to the tracking path to reduce the tracking error. The particle filter algorithm can estimate the target state by combining observation and prediction. Specifically, the state variable in the particle filter algorithm includes the distance and speed of the UAV, and the observation result is the tracking path obtained by embodiments of the present application before. Embodiments of the present application initialize 5000 uniformly distributed particles, and use the polynomial resampling algorithm as the resampling method of the particles.

[0248]

[0249] ​Optionally, the embodiments of the present application provide the user with the UAV angle estimation result. Due to the limited number of millimeter wave radar antennas, the angle resolution is limited, and the angle estimation result fluctuates greatly. In order to obtain the angle estimation result, the embodiments of the present application first apply a beamforming algorithm to the reflection signals of all antennas corresponding to the UAV to obtain the signal at each angle. Then, they are subjected to Doppler FFT operation to obtain the angle-Doppler spectrum. Finally, the same operation as applied to the range-Doppler spectrum is performed on the angle-Doppler spectrum at each angle.

[0250] (III) UAV identification

[0251] After obtaining the tracking result of the UAV, the embodiments of the present application further identify the UAV by using the PMM feature. The embodiments of the present application first extract the Doppler spectrum belonging to the target to be detected from a series of range-Doppler spectra (Doppler spectrum set) according to the tracking path. Then, the Doppler spectrum containing the PMM feature is input into a UAV detection model based on a long short-term memory artificial neural network, i.e., an LSTM detector, for UAV identification. Figure 6 is a flowchart of the UAV identification provided by the embodiments of the present application, and the UAV identification process is as shown in Figure 6 .

[0252] According to the tracking result and the range-Doppler spectrum, the embodiments of the present application can extract a series of Doppler spectra from the position of the target to form a Doppler-time graph. When the target is a UAV, the graph will contain unique and continuous PMM features, which can be used to distinguish UAVs from other objects.

[0253] The embodiments of the present application first remove the direct current noise from the Doppler-time graph. Considering that the main speed of the UAV is close to 0 when hovering, the corresponding direct current component contains the peak corresponding to the main speed. In order to preserve the peak corresponding to the main speed, the embodiments of the present application average the direct current components in the Doppler spectrum when the main speed peak is not close to the direct current component. Then, the average value is subtracted from all direct current components.

[0254] In addition, since the main speed of the UAV determines the center of the PMM feature, this feature will change with the movement of the UAV. Therefore, it will hinder the use of the periodicity of the PMM feature by the embodiments of the present application. Therefore, the embodiments of the present application design a feature alignment algorithm on the Doppler-time graph to align the center of each PMM feature to the center of the Doppler spectrum. When the main speed peak in the Doppler spectrum is not near the D direct current component, the embodiments of the present application shift the entire Doppler spectrum in the direction of the direct current component along the main speed peak, and supplement it by linear interpolation.

[0255] The pre-processed Doppler-time map contains PMM features. However, there can be errors in the tracking results, especially when the UAV is at high altitude. Directly using the obtained Doppler-time map is not always reliable. To reduce the impact of tracking errors, embodiments of the present application first segment the Doppler-time map into fixed-length segments (Doppler-time sub-maps). Then, embodiments of the present application compare the maximum fold result of each segment with a first threshold. When the maximum fold result is less than the first threshold, embodiments of the present application consider that the Doppler-time sub-map does not contain PMM features and discard it. According to extensive empirical experiments, embodiments of the present application set the first threshold to 30000.

[0256] Identifying a UAV from other objects is a binary classification problem. Embodiments of the present application design a LSTM-based UAV detection model to solve this problem. LSTM networks are suitable for processing data sequences and have excellent performance in recognition tasks. Considering that each time slot in the Doppler-time map contains L Doppler bins, the input dimension of the UAV detection model of embodiments of the present application is set to L. The UAV detection model contains two LSTM layers, with a hidden state size of 128. Finally, a fully connected layer is used to map the hidden state to the recognition result, i.e., UAV or non-UAV. Embodiments of the present application select the cross-entropy loss function to train the network of embodiments of the present application.

[0257] In one embodiment, it is implemented based on the commercial millimeter wave radar Texas Instruments IWR6843IS KODS. There are 3 Tx antennas and 4 Rx antennas on the radar board. Embodiments of the present application make the three TX antennas transmit FMCW signals with a frequency of 60.25 GHz and a bandwidth of 1.92 GHz in turn, and all Rx antennas receive reflected signals. The duration of a single signal is 900us, and each frame includes 100 signals, with a tracking interval of 90ms. The frequency slope of the FMCW signal is 9.994MHz / us, and the ADC sampling rate is 6250kHz. Therefore, the maximum sensing range of the radar can reach The angle of the radar field of view (FoV) is about 120°. The raw data of the radar is collected by the TI DCA1000EVM data acquisition board and transmitted to a computer with an Intel Core i9-11900H 2.5GHz CPU for processing. To verify the stability of the technology, embodiments of the present application have carried out experimental verification under various configurations of UAV flight altitude, flight trajectory and flight speed.

[0258] The UAV detection system provided by the present application is described below, and the UAV detection system described below can be mutually corresponding with the UAV detection method described above.

[0259] Figure 7is a structural schematic diagram of a UAV detection system provided by an embodiment of the present application, as shown in the figure, the present application provides a UAV detection system, comprising: Figure 7

[0260] The acquisition unit 710 is configured to acquire a reflection signal of a target to be detected.

[0261] The feature extraction unit 720 is configured to determine a Doppler spectrum set based on the reflection signal.

[0262] The tracking unit 730 is configured to determine a tracking path of the target to be detected based on the Doppler spectrum set.

[0263] The extraction unit 740 is configured to perform spectrum extraction on the Doppler spectrum set based on the tracking path, and obtain a Doppler-time graph corresponding to the target to be detected.

[0264] The identification unit 750 is configured to input the Doppler-time graph into a UAV detection model, and obtain a detection result output by the UAV detection model.

[0265] The UAV detection model is obtained based on Doppler-time graph training samples and labels corresponding to the Doppler-time graph training samples, and the labels are determined based on periodic micro-motion features in the Doppler-time graph training samples.

[0266] Optionally, the tracking unit 730 is configured to perform spectrum folding on each Doppler spectrum in the Doppler spectrum set, and obtain a folding result corresponding to each Doppler spectrum.

[0267] The tracking unit 730 is configured to obtain the tracking path of the target to be detected based on the folding result.

[0268] Optionally, the tracking unit 730 is configured to determine a maximum distance change value of the target to be detected within a time unit based on a maximum speed of the target to be detected, the time unit, and a distance resolution of the radar.

[0269] The tracking unit 730 is configured to determine a maximum cumulative folding result under the constraint of the maximum distance change value.

[0270] The tracking unit 730 is configured to take a path corresponding to the maximum cumulative folding result as the tracking path.

[0271] Optionally, the system further comprises a noise reduction unit.

[0272] The noise reduction unit is configured to process the folding result based on a spectrum subtraction method, and obtain a noise-reduced folding result.

[0273] ​Optionally, the system further comprises a correction unit.

[0274] The correction unit is configured to correct the tracking path based on a particle filter algorithm to obtain a corrected tracking path.

[0275] Optionally, the noise reduction unit is further configured to remove direct current noise in the Doppler-time graph to obtain a noise-reduced Doppler-time graph.

[0276] Optionally, the system further comprises an alignment unit.

[0277] The alignment unit is configured to align, in the Doppler-time graph, a periodic micro-motion feature center in each Doppler spectrum with a center of the Doppler spectrum.

[0278] Optionally, the system further comprises a segmentation unit.

[0279] The segmentation unit is configured to segment the Doppler-time graph to obtain a plurality of Doppler-time sub-graphs.

[0280] The segmentation unit is further configured to determine a maximum folding result corresponding to each Doppler-time sub-graph, the maximum folding result being a maximum value in folding results corresponding to Doppler spectra in the Doppler-time sub-graph.

[0281] The segmentation unit is further configured to compare the maximum folding result with a first threshold value one by one, and retain a Doppler-time sub-graph corresponding to a maximum folding result that is not less than the first threshold value.

[0282] The identification unit is configured to input the retained Doppler-time sub-graph into a UAV detection model.

[0283] It should be noted that the above system provided by the embodiments of the present application can realize all the method steps realized by the above method embodiments, and can achieve the same technical effects. Therefore, the same parts and beneficial effects of the method embodiments will not be described in detail.

[0284] Figure 8 An example of an electronic device is shown in the physical structure diagram as shown in Figure 8As shown, the electronic device can include a processor 810, a communications interface 820, a memory 830, and a communications bus 840, wherein the processor 810, the communications interface 820, and the memory 830 complete mutual communication through the communications bus 840. The processor 810 can invoke a logic instruction in the memory 830 to execute a kind of unmanned aerial vehicle detection method based on radar implementation, including: obtaining the reflection signal of the target to be detected;Based on the reflection signal, a plurality of Doppler spectrum sets are determined;Based on the plurality of Doppler spectrum sets, the tracking path of the target to be detected is determined;Based on the tracking path, the plurality of Doppler spectrum sets are extracted, and the Doppler-time graph corresponding to the target to be detected is obtained;The Doppler-time graph is input into the unmanned aerial vehicle detection model, and the detection result output by the unmanned aerial vehicle detection model is obtained;Wherein, the unmanned aerial vehicle detection model is obtained after training based on the Doppler-time graph training sample and the label corresponding to the Doppler-time graph training sample, and the label is determined based on the periodic micro-motion feature in the Doppler-time graph training sample.

[0285] In addition, the logic instruction in the memory 830 described above can be realized in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0286] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program being stored in a non-transitory computer-readable storage medium, and the computer program being executable by a processor to enable a computer to perform a UAV detection method provided by any of the above methods based on radar, comprising: obtaining a reflection signal of a to-be-detected target; determining a Doppler spectrum set based on the reflection signal; determining a tracking path of the to-be-detected target based on the Doppler spectrum set; performing spectrum extraction on the Doppler spectrum set based on the tracking path to obtain a Doppler-time graph corresponding to the to-be-detected target; inputting the Doppler-time graph into a UAV detection model to obtain a detection result output by the UAV detection model; wherein the UAV detection model is obtained by training based on Doppler-time graph training samples and labels corresponding to the Doppler-time graph training samples, and the labels are determined based on periodic micro-motion features in the Doppler-time graph training samples.

[0287] In another aspect, the present application also provides a non-transitory computer-readable storage medium, which stores a computer program, and the computer program is executable by a processor to implement a UAV detection method provided by any of the above methods based on radar, comprising: obtaining a reflection signal of a to-be-detected target; determining a Doppler spectrum set based on the reflection signal; determining a tracking path of the to-be-detected target based on the Doppler spectrum set; performing spectrum extraction on the Doppler spectrum set based on the tracking path to obtain a Doppler-time graph corresponding to the to-be-detected target; inputting the Doppler-time graph into a UAV detection model to obtain a detection result output by the UAV detection model; wherein the UAV detection model is obtained by training based on Doppler-time graph training samples and labels corresponding to the Doppler-time graph training samples, and the labels are determined based on periodic micro-motion features in the Doppler-time graph training samples.

[0288] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0289] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0290] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting unmanned aerial vehicles (UAVs), characterized in that, Radar-based implementation, including: Acquire the reflection signal of the target to be detected; Determine the Doppler spectrum set based on the reflected signal; The tracking path of the target to be detected is determined based on the Doppler spectrum set; Based on the tracking path, the Doppler spectrum set is extracted to obtain the Doppler-time map corresponding to the target to be detected; The Doppler-time map is input into the UAV detection model to obtain the detection results output by the UAV detection model; The drone detection model is obtained by training on Doppler time map training samples and the corresponding labels of the Doppler time map training samples. The labels are determined based on the periodic micro-motion features in the Doppler time map training samples. The step of determining the tracking path of the target to be detected based on the Doppler spectrum set includes: Perform spectrum folding on each Doppler spectrum in the Doppler spectrum set to obtain the folding result for each Doppler spectrum; The tracking path of the target to be detected is obtained based on the folding result.

2. The UAV detection method according to claim 1, characterized in that, The step of obtaining the tracking path of the target to be detected based on the folding result includes: Based on the maximum velocity of the target to be detected, the time unit, and the range resolution of the radar, determine the maximum range change value of the target to be detected within one time unit; Under the constraint of the maximum distance change value, determine the maximum cumulative folding result; The path corresponding to the maximum cumulative folding result is taken as the tracking path.

3. The UAV detection method according to claim 1, characterized in that, Before obtaining the tracking path of the target to be detected based on the folding result, the method further includes: The folding result is processed using the spectral subtraction method to obtain a denoised folding result.

4. The UAV detection method according to any one of claims 1-3, characterized in that, The method further includes: The tracking path is corrected based on the particle filter algorithm to obtain the corrected tracking path.

5. The UAV detection method according to any one of claims 1-3, characterized in that, The method further includes: The DC noise in the Doppler-time plot is removed to obtain the denoised Doppler-time plot.

6. The UAV detection method according to any one of claims 1-3, characterized in that, The method further includes: In the Doppler-time diagram, the center of the periodic micro-motion feature in each Doppler spectrum is aligned with the center of the Doppler spectrum.

7. The UAV detection method according to any one of claims 1-3, characterized in that, The method further includes: The Doppler-time map is segmented to obtain multiple Doppler-time partial maps; Determine the maximum folding result corresponding to each Doppler-time partial plot, wherein the maximum folding result is the maximum value among the folding results corresponding to the Doppler spectrum in the Doppler-time partial plot; The maximum folding result is compared with the first threshold one by one, and the Doppler-time portion of the image corresponding to the maximum folding result that is not less than the first threshold is retained. The step of inputting the Doppler-time map into the UAV detection model includes: The preserved Doppler-time partial map is input into the UAV detection model.

8. A drone detection system, characterized in that, Radar-based implementation, including: Acquisition unit, used to acquire the reflection signal of the target to be detected; The feature extraction unit is used to determine the Doppler spectrum set based on the reflected signal; A tracking unit is used to determine the tracking path of the target to be detected based on the Doppler spectrum set; The extraction unit is used to extract the spectrum from the Doppler spectrum set based on the tracking path to obtain the Doppler-time map corresponding to the target to be detected; The identification unit is used to input the Doppler-time map into the UAV detection model to obtain the detection result output by the UAV detection model; The drone detection model is obtained by training on Doppler time map training samples and the corresponding labels of the Doppler time map training samples. The labels are determined based on the periodic micro-motion features in the Doppler time map training samples. The step of determining the tracking path of the target to be detected based on the Doppler spectrum set includes: Perform spectrum folding on each Doppler spectrum in the Doppler spectrum set to obtain the folding result for each Doppler spectrum; The tracking path of the target to be detected is obtained based on the folding result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the UAV detection method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the UAV detection method as described in any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the UAV detection method as described in any one of claims 1 to 7.

Images