Lightweight and efficient unmanned aerial vehicle 3D tracking and identification method based on millimeter wave radar

Abstract

The invention discloses a lightweight and efficient unmanned aerial vehicle 3D tracking and identification method based on a millimeter wave radar, and the method comprises the steps: carrying out the feature extraction of a millimeter wave signal related to the motion of an unmanned aerial vehicle, and enhancing the difference between the periodic features and noise of the unmanned aerial vehicle through a spectrum folding technology; a Doppler spectrum sequence is obtained after signal processing, the distance of the unmanned aerial vehicle is measured and tracked in real time, and fluctuation of Doppler spectrum intensity and static background noise level in different ranges is reduced; according to the measured distance of the unmanned aerial vehicle, angle position information of the unmanned aerial vehicle is measured and tracked based on an arrival direction algorithm of the subspace; and constructing a lightweight unmanned aerial vehicle identification network model based on a spiking neural network SNN architecture, and obtaining an unmanned aerial vehicle 3D identification result. According to the scheme, the improved angle estimation algorithm is combined with dynamic programming and the particle filter, real-time 3D unmanned aerial vehicle tracking can be achieved, meanwhile, the calculation complexity is reduced, and the tracking calculation time efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a lightweight and efficient unmanned aerial vehicle 3D tracking and identification method based on millimeter wave radar. Background Art

[0002] The widespread deployment of small unmanned aerial vehicles (UAVs) has raised a number of associated threats, such as invasion of personal privacy, unauthorized transportation of contraband, concerns about public safety, and the possibility of physical harm to individuals. Figure 1 As shown, uninvited intrusion into personal space is of particular concern, posing a significant risk to an individual’s privacy and safety. Without strict regulations, drones can easily illegally intrude into personal space at high altitudes to engage in activities such as voyeurism or theft. Given the small size of these drones, it is difficult to detect them with the naked eye, so it becomes particularly urgent to develop effective detection mechanisms. The development of drone detection systems is critical to achieving several major goals. First, the detection method should preferably be passive and not require the cooperation of the drone during the detection process. Second, the system must be able to detect drones at different altitudes. The detection system should also be cost-effective and easy to deploy, such as running directly on an FPGA to enable real-time and fast detection on a mobile terminal. Finally, the system should have the general ability to detect a variety of drones under different lighting and noise conditions.

[0003] Unfortunately, existing drone detection methods have limitations that hinder their effectiveness. For example, sound-based drone detection methods are easily interfered by complex environmental noise. At the same time, the sound of drones decays quickly in the air, limiting the detection range of sound-based methods.

[0004] Vision-based drone detection techniques have shown effectiveness in scenarios where drones are perceptible; however, their accuracy and reliability are susceptible to changes in lighting conditions and visual environments. While thermal imaging and infrared radiometric imaging cameras offer viable alternatives, they are limited by high cost and limited surveillance range. Some studies advocate detecting drones by analyzing radio frequency signals, which requires specialized equipment to intercept and interpret transmissions from uncooperative drones. Conventional radar systems are characterized by high cost and high power requirements, which make them unsuitable for routine drone detection operations. Therefore, none of the current methods are ideally suited for drone detection in common operating environments. Non-cooperative drone detection faces many challenges, including weak signal reflections due to the small size of drones, susceptibility to environmental noise, the need to cover a large detection range to detect high-altitude drones, the lack of critical information about non-cooperative drones, and the need to accurately detect drones under varying lighting and noise conditions in complex everyday scenarios. mmHawkeye represents a pioneer in millimeter-wave-based long-range drone detection methods. Exploiting the signal characteristics triggered by drones’ periodic micromotions, mmHawkeye can track the distance information of drones through dynamic programming and particle filtering, and then use long short-term memory networks for recognition. It has demonstrated the ability to detect and classify non-cooperative UAVs even in weak signal conditions. However, the ongoing challenge lies in achieving high-resolution target angle estimation while maintaining cost-effective signal processing. In addition, the rapid maneuverability and wide operating range of UAVs require widespread installation of surveillance systems that are both energy-efficient and compact. These demands pose further challenges to existing approaches that require large computational resources and consume a lot of energy. Summary of the invention

[0005] Based on the above technical problems, the present invention proposes a lightweight and efficient UAV 3D tracking and identification method based on millimeter wave radar. The specific technical solution is as follows.

[0006] A lightweight and efficient UAV 3D tracking and identification method based on millimeter wave radar includes the following steps:

[0007] Step S1: extracting features of millimeter wave signals related to the UAV motion, and using spectrum folding technology to enhance the distinction between the UAV periodic features and noise;

[0008] Step S2: After signal processing, the Doppler spectrum sequence D is obtained S , measure and track the distance of UAVs in real time, and mitigate the fluctuations of Doppler spectrum intensity and static background noise level at different ranges;

[0009] Step S3: According to the distance of the UAV measured in step S2, the angular position information of the UAV is measured and tracked based on the subspace-based arrival direction algorithm;

[0010] Step S4: Build a lightweight drone recognition network model based on the pulse neural network (SNN) architecture to obtain the drone 3D recognition results.

[0011] Specifically, the step S1 includes the following sub-steps:

[0012] Step S11: After the range Doppler FFT processing, the reflected millimeter wave signal generates a range Doppler spectrum, which is expressed as D i represents the Doppler spectrum in the i-th range bin, and R represents the total range Doppler spectrum;

[0013] Step S12: Fold the Doppler spectrum into a matrix to amplify the periodicity, iterating the specified range [j min ,j max ] determine the ideal folding size; where j min and j max are the lower and upper limits of the expected folded size derived from the range of propeller rotational speeds;

[0014] Step S13: Each Doppler spectrum D with a specified folding size j i The folding value calculation formula is:

[0015]

[0016] Where k and m represent the folding column index and row index respectively; the folding result of each Doppler spectrum is expressed as:

[0017]

[0018] Specifically, the Doppler spectrum sequence D S Contains T-range Doppler spectra, where each spectrogram contains R-range intervals, value D S (r, t) represents the folding result of the rth distance interval in the tth distance Doppler spectrum. The trajectory of the UAV is expressed as a sequence D S A series of distance intervals with higher values.

[0019] Specifically, the step S2 of measuring and tracking the distance of drones in real time also includes: using trajectory continuity to reduce tracking differences, and imposing constraints on the maximum distance box deviations of adjacent columns of drones to maintain tracking accuracy.

[0020] Specifically, the step S2 of alleviating the fluctuation of Doppler spectrum intensity and static background noise level in different ranges specifically includes:

[0021] Preprocessing is performed using a spectrum subtraction algorithm to subtract an estimated average background noise spectrum calculated by averaging the background noise;

[0022] The maximum cumulative folding result is calculated using a dynamic programming algorithm and passed through D S Backtracking in the space identifies the best UAV trajectory, backtracking from the last column to the first column, and finally produces the constrained maximum path g, which represents the best tracking solution;

[0023] A particle filter is used to optimize the initial tracking estimate and mitigate potential errors. A weight is assigned to each particle according to its consistency with the folded result, focusing the particle filter on the most likely UAV trajectory.

[0024] Specifically, the step S3 includes the following sub-steps:

[0025] A uniform sampling matrix Ω is randomly generated, each row of which has a single random non-zero entry, and an orthogonal basis of the covariance matrix of the millimeter wave signal is calculated by orthogonal normalization of the sampling matrix;

[0026] Using the unitary invariance of the spectrum, the Frobenius norm and the low-rank decomposition, we can obtain an approximate signal subspace And generate a pseudo spectrum, the calculation formula is:

[0027]

[0028] in:

[0029] a(θ)=[a0(θ),a1(θ),...,a M-1 (θ)] H ;

[0030]

[0031] I M represents the M×M identity matrix, Represents the approximate signal subspace, and the peak position in the pseudo spectrum is the direction information of the target, specifically:

[0032]

[0033] peak is the peak value judgment function.

[0034] Specifically, step S4 includes the following sub-steps:

[0035] The feature alignment algorithm is used to align the center frequency with the center of the Doppler spectrum, and linear interpolation is used to fill the gap;

[0036] To mitigate tracking discrepancies, a series of fixed-length segments are generated, and the maximum folding results of each segment below a preset empirical threshold are excluded;

[0037] The segmented Doppler spectrum is input into the UAV identification network model Spiking Neural Network (SNN) architecture for UAV classification.

[0038] Specifically, the UAV identification network model is a pulse neural network SNN architecture of L-128-128-2, where L is the number of Doppler bins in each time slot; the basic unit of the pulse neural network SNN architecture is constructed using an intermediate frequency spike neuron model, including membrane potential accumulation, spike generation and reset mechanisms; the floating-point synaptic weights are quantized to 1-bit representation, the information entropy of the binary weights is increased during the forward propagation process, the binary peak s[t] and the binary weight B are integrated, and the convolution operation is performed using bitwise operations.

[0039] Beneficial effects of the invention: The invention proposes a lightweight and efficient 3D UAV tracking and identification method based on millimeter wave radar, which can effectively handle low signal-to-noise ratio and uncertain signals by capturing the micro-Doppler characteristics of UAV motion. The improved angle estimation algorithm combined with dynamic programming and particle filter can achieve real-time 3D UAV tracking while reducing computational complexity and significantly improving tracking calculation time efficiency. In addition, using the position of the UAV and the corresponding Doppler information, a simple UAV identification model based on the spiking neural network (SNN) architecture is developed, and its size is reduced to one-tenth of the original size, but the system performance remains close to the original level. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of a drone invasion threat scenario in the prior art solution;

[0041] Figure 2 It is a flow chart of a lightweight and efficient UAV 3D tracking and identification method based on millimeter wave radar of the present invention;

[0042] Figure 3 It is an overall design architecture diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described with reference to the accompanying drawings.

[0044] The present invention proposes a lightweight and efficient UAV 3D tracking and identification method based on millimeter wave radar, and its architecture is as follows: Figure 2 , 3 As shown, the following steps are included:

[0045] Step S1: extracting features of millimeter wave signals related to the UAV motion, and using spectrum folding technology to enhance the distinction between the UAV periodic features and noise;

[0046] Step S2: After signal processing, the Doppler spectrum sequence D is obtainedS , measure and track the distance of UAVs in real time, and mitigate the fluctuations of Doppler spectrum intensity and static background noise level at different ranges;

[0047] Step S3: According to the distance of the UAV measured in step S2, the angular position information of the UAV is measured and tracked based on the subspace-based arrival direction algorithm;

[0048] Step S4: Build a lightweight drone recognition network model based on the pulse neural network (SNN) architecture to obtain the drone 3D recognition results.

[0049] In this embodiment, step S1 specifically includes the following sub-steps:

[0050] Step S11: After the range Doppler FFT processing, the reflected millimeter wave signal generates a range Doppler spectrum, which is expressed as D i represents the Doppler spectrum in the i-th range bin, and R represents the total range Doppler spectrum;

[0051] Step S12: Fold the Doppler spectrum into a matrix to amplify the periodicity, iterating the specified range [j min ,j max ] determine the ideal folding size; where j min and j max are the lower and upper limits of the expected folded size derived from the range of propeller rotational speeds;

[0052] Step S13: Each Doppler spectrum D with a specified folding size j i The folding value calculation formula is:

[0053]

[0054] Where k and m represent the folding column index and row index respectively; the folding result of each Doppler spectrum is expressed as:

[0055]

[0056] In this embodiment, the Doppler spectrum sequence D S Contains T-range Doppler spectra, where each spectrogram contains R-range intervals, value D S (r, t) represents the folding result of the rth distance interval in the tth distance Doppler spectrum. The trajectory of the UAV is expressed as a sequence D S A series of distance intervals with higher values.

[0057] In this embodiment, step S2 of measuring and tracking the distance of drones in real time also includes: using trajectory continuity to reduce tracking differences, and imposing constraints on the maximum distance box deviations of adjacent columns of drones to maintain tracking accuracy.

[0058] In this embodiment, step S2 of alleviating the fluctuation of Doppler spectrum intensity and static background noise level in different ranges specifically includes:

[0059] Preprocessing is performed using a spectrum subtraction algorithm to subtract an estimated average background noise spectrum calculated by averaging the background noise;

[0060] The maximum cumulative folding result is calculated using a dynamic programming algorithm and passed through D S Backtracking in the space identifies the best UAV trajectory, backtracking from the last column to the first column, and finally produces the constrained maximum path g, which represents the best tracking solution;

[0061] A particle filter is used to optimize the initial tracking estimate and mitigate potential errors. A weight is assigned to each particle according to its consistency with the folded result, focusing the particle filter on the most likely UAV trajectory.

[0062] In this embodiment, step S3 specifically includes the following sub-steps:

[0063] A uniform sampling matrix Ω is randomly generated, each row of which has a single random non-zero entry, and an orthogonal basis of the covariance matrix of the millimeter wave signal is calculated by orthogonal normalization of the sampling matrix;

[0064] Using the unitary invariance of the spectrum, the Frobenius norm and the low-rank decomposition, we can obtain an approximate signal subspace And generate a pseudo spectrum, the calculation formula is:

[0065]

[0066] in:

[0067] a(θ)=[a0(θ),a1(θ),...,a M-1 (θ)] H ;

[0068]

[0069] I M represents the M×M identity matrix, Represents the approximate signal subspace, and the peak position in the pseudo spectrum is the direction information of the target, specifically:

[0070]

[0071] peak is the peak value judgment function.

[0072] In this embodiment, step S4 includes the following sub-steps:

[0073] The feature alignment algorithm is used to align the center frequency with the center of the Doppler spectrum, and linear interpolation is used to fill the gap;

[0074] To mitigate tracking discrepancies, a series of fixed-length segments are generated, and the maximum folding results of each segment below a preset empirical threshold are excluded;

[0075] The segmented Doppler spectrum is input into the UAV identification network model Spiking Neural Network (SNN) architecture for UAV classification.

[0076] In this embodiment, the drone identification network model is a pulse neural network SNN architecture of L-128-128-2, where L is the number of Doppler bins in each time slot; a medium frequency spike neuron model is used to construct the basic unit of the pulse neural network SNN architecture, including membrane potential accumulation, spike generation and reset mechanism; the floating point synaptic weight is quantized to 1 bit representation, the information entropy of the binary weight is increased during the forward propagation process, the binary peak s[t] and the binary weight B are integrated, and the convolution operation is performed using bitwise operation.

[0077] In one embodiment, a lightweight and efficient 3D tracking and identification technology for drones based on millimeter wave radar includes the following steps:

[0078] S1. Extraction of millimeter wave signal features related to UAV motion. After range Doppler FFT processing, the reflected millimeter wave signal generates a range Doppler spectrum, which is represented by D i represents the Doppler spectrum in the i-th range bin, and R represents the total range Doppler spectrum. If the drone is in the i-th range bin, then the corresponding i The spectrum shows periodic peaks. However, due to the unpredictable propeller speed, these periodic peaks show variable periods and may be masked by noise due to low signal-to-noise ratio (SNR). We use spectrum folding technology to enhance the distinction between the UAV periodic features and noise. The core concept of spectrum folding is to amplify the periodic pattern by folding the Doppler spectrum into a matrix. To determine the ideal folding size, it is necessary to iteratively specify the range [j min ,j max ], where j min and j max The Doppler spectrum resolution is determined by the lower and upper bounds of the expected folding size derived from the propeller rotation speed range. This predefined traversal range is intended to maintain computationally feasible complexity. The folding value for each Doppler spectrum D with a specified folding size j is calculated according to the following equation:

[0079]

[0080] Here, k and m represent the folded column index and row index, respectively. iThe final folding result of PMM shows the prominence of the PMM features and is obtained by selecting the maximum folding value of all traversed sizes.

[0081]

[0082] By calculating the folding result (P) for each Doppler spectrum, the periodic characteristics of drones and background noise can be effectively distinguished. The inherent periodicity ensures significantly higher folding results compared to the random fluctuation characteristics of noise. This significant difference in folding results is a strong indicator of drone detection within a specific range.

[0083] S2, real-time drone distance measurement and tracking. After signal processing in S1, the Doppler spectrum sequence D can be obtained. S , the sequence contains T-range Doppler spectra, where each spectrogram contains R-range intervals. The value DS(r, t) represents the folding result of the r-th range interval within the t-th range Doppler spectrum. The trajectory of the drone is represented by the sequence D S A series of distance intervals with higher values ​​can be tracked by identifying the paths with the greatest cumulative collapse results over time.

[0084] Considering the rapid changes in the folding results caused by the UAV motion and the impact of local dynamic noise on tracking accuracy, trajectory continuity is used to mitigate tracking differences. Constraints are imposed on the maximum range bin deviations of adjacent columns of UAVs to maintain tracking accuracy. In order to mitigate the fluctuations in Doppler spectrum intensity and static background noise levels at different ranges, a spectrum subtraction algorithm is first used for preprocessing, which subtracts the estimated average background noise spectrum calculated by averaging the background noise. A dynamic programming algorithm is then used to calculate the maximum cumulative folding result and the D S Backtracking in the space to identify the best drone trajectory. This process requires backtracking from the last column to the first column, ultimately producing a constrained maximum path (denoted as g) that represents the best tracking solution. At the same time, a particle filter is used to further optimize this initial tracking estimate and mitigate potential errors. By assigning a weight to each particle based on its consistency with the folded result, the particle filter effectively focuses on the most likely drone trajectory.

[0085] S3, fast and efficient UAV angle measurement and tracking. Based on the UAV distance result estimated by S2, the subspace-based arrival direction algorithm is applied to achieve higher resolution, and the angular position information (azimuth and pitch angle) of the UAV has been calculated. In the process of generating pseudo-spectrum to estimate the target direction, the conventional MUSIC algorithm requires a lot of computing requirements due to the detailed grid search involved, resulting in high computational complexity and time complexity, especially the calculation of the signal subspace Us. In order to reduce the computational complexity of obtaining the subspace, the present invention first randomly generates a uniform sampling matrix Ω with a single random non-zero entry in each row. Its random nature ensures the linear independence of the sampling columns, which helps to distinguish between noise and signal subspaces. The orthogonal basis of the millimeter wave signal covariance matrix is ​​then calculated by orthogonal normalization of the sampling matrix. Secondly, by utilizing the unitary invariance and Frobenius norm of the spectrum and low-rank decomposition, an approximate signal subspace can be obtained. And generate a pseudo spectrum:

[0086]

[0087] Finally, the peak position in the pseudo spectrum is the direction information of the target:

[0088]

[0089] Once the reflected signal from each angle is obtained, a Doppler FFT operation can be applied to derive the angular Doppler spectrum. This feature can be extracted from the angle bin corresponding to the position of the UAV using the method detailed in 1. These folded results from the angle bins can form an angle-time-dependent feature space. The same procedures as for the range-time-dependent feature space DS are then applied, including spectral subtraction, dynamic programming-based tracking, and particle filtering.

[0090] S4. Lightweight drone identification network. Since the unpredictable speed of drones can affect the position of the center frequency, a feature alignment algorithm is first used to align their centers with the center of the Doppler spectrum, and linear interpolation is used to fill the gap. Next, to mitigate tracking discrepancies, segmentation is applied to generate a series of fixed-length sections, and the maximum folding results of each segment below an empirical threshold are excluded to improve accuracy. In this work, the threshold is 3000. Finally, the segmented Doppler spectra are input into the SNN for drone classification. We choose the IF spiking neuron model as the basic unit for building the drone identification SNN. The behavior of the IF neuron model involves three parts: membrane potential accumulation, spike generation, and reset mechanism. To further exploit the energy efficiency advantage of our model, we quantize the floating-point synaptic weights to a 1-bit representation. To mitigate the accuracy degradation caused by the quantization operation, we retain more information during the forward propagation process by increasing the information entropy of the binary weights. Therefore, by integrating the binary peaks s[t] and the binary weights B, we can perform convolution operations using cost-effective bitwise operations, thereby improving computational efficiency. For the architecture design of our model, considering that each time slot in the Doppler time map contains L Doppler bins, we formulate an L-128-128-2 SNN structure to obtain the recognition results.

[0091] In another embodiment, if Figure 3 As shown in the figure, first, the millimeter wave radar Tx antenna transmits FMCW signals, which are reflected by the drone and received by the radar Rx antenna, and TF signals are extracted for subsequent drone tracking and detection. After the system obtains the TF signal, the signal preprocessing module starts to continuously extract the range Doppler spectrum using the range Doppler FFT method. Next, the spectrum characteristics of the drone are amplified by the spectrum folding technology to distinguish the environmental noise. Subsequently, the distance estimation module converts the distance tracking of the target into a longest path search problem and solves it through dynamic programming and particle filtering. Using the estimated drone range, the angle estimation module uses a complex 2D MUSIC algorithm to derive the angle Doppler spectrum corresponding to a specific range, which helps to estimate the azimuth and elevation angles similar to the distance estimation process. Finally, the system obtains a series of 3D coordinates (including distance, azimuth and elevation) of the flying drone trajectory. After obtaining the drone's trajectory, the system undergoes a noise reduction, alignment and segmentation process, and the Doppler segments are input into a specially designed spike neural network for recognition.

[0092] The specific execution steps are as follows.

[0093] S1. Signal preprocessing. Given the received TF signal, the CFAR technique is first used to remove the interference of environmental noise on the signal, and then the Range-Doppler-FFT is used to obtain the range-Doppler spectrum. Subsequently, the spectrum expansion technique is used to convert the spectrum into a matrix form to amplify the periodic characteristics caused by the UAV movement and further improve the distinction between the UAV periodic characteristics and noise.

[0094] S2, distance tracking. The drone tracking process first preprocesses the series of folded results DS obtained in S1 by spectral subtraction to reduce the influence of static background noise. Then, the drone tracking mechanism integrates dynamic programming and particle filtering technology to traverse the search sequence D S The maximum distance interval with a series of higher values ​​in the range is the path corresponding to the optimal UAV motion trajectory. Each range bin of the UAV is obtained by backtracking the distance, which has been used for UAV angle estimation. In the implementation, the filter uses 5000 particles and multiple resampling algorithms to iteratively refine the state estimation and improve the tracking accuracy.

[0095] S3, angle tracking. Given a series of range bin values ​​for the drone, the system uses the proposed arrival angle estimation algorithm to extract the azimuth and elevation information of the candidate drone location point from the FMCW reflection signal corresponding to the range bin. The angle-Doppler spectrum diagram can be obtained by using dopper-FFT at the signal of the candidate location point, and the angle information corresponding to the drone in each range bin can be screened out by combining the method in S1. Finally, dynamic programming and particle filtering technology are used to extract the final azimuth and elevation information, and combined with the distance information in S2 to form a series of drone 3D location information.

[0096] S4, drone identification. The system can extract the Doppler time map related to the drone based on the real-time 3D position information of the drone, and use the feature alignment algorithm and linear interpolation algorithm to eliminate the influence of tracking errors. Subsequently, the Doppler time map is segmented, and the spectrum map with the maximum folding result below the threshold is deleted. In the present invention, the threshold is 3000. Finally, the remaining segmented Doppler spectra are input into the SNN for drone classification and identification.

[0097] The present invention has shown and described the basic principles and main features of the present invention and the advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A lightweight and efficient UAV 3D tracking and identification method based on millimeter wave radar, characterized in that: The following steps are involved: Step S1: extracting features of millimeter wave signals related to the UAV motion, and using spectrum folding technology to enhance the distinction between the UAV periodic features and noise; Step S2: After signal processing, the Doppler spectrum sequence D is obtained S , measure and track the distance of UAVs in real time, and mitigate the fluctuations of Doppler spectrum intensity and static background noise level at different ranges; Step S3: According to the distance of the UAV measured in step S2, the angular position information of the UAV is measured and tracked based on the subspace-based arrival direction algorithm; Step S4: Build a lightweight drone recognition network model based on the pulse neural network (SNN) architecture to obtain the drone 3D recognition results.

2. According to claim 1, a lightweight and efficient UAV 3D tracking and identification method based on millimeter wave radar is characterized in that: The step S1 specifically includes the following sub-steps: Step S11: After the range Doppler FFT processing, the reflected millimeter wave signal generates a range Doppler spectrum, which is expressed as D i represents the Doppler spectrum in the i-th range bin, and R represents the total range Doppler spectrum; Step S12: Fold the Doppler spectrum into a matrix to amplify the periodicity, iterating the specified range [j min ,j max ] determine the ideal folding size; where j min and j max are the lower and upper limits of the expected folded size derived from the range of propeller rotational speeds; Step S13: Each Doppler spectrum D with a specified folding size j i The folding value calculation formula is: Where k and m represent the folding column index and row index respectively; the folding result of each Doppler spectrum is expressed as:

3. According to claim 1, a lightweight and efficient UAV 3D tracking and identification method based on millimeter wave radar is characterized in that: The Doppler spectrum sequence D S Contains T-range Doppler spectra, where each spectrogram contains R-range intervals, value D S (r, t) represents the folding result of the rth distance interval in the tth distance Doppler spectrum. The trajectory of the UAV is expressed as a sequence D S A series of distance intervals with higher values.

4. The lightweight and efficient 3D tracking and identification method of UAV based on millimeter wave radar according to claim 3 is characterized in that: The step S2 of measuring and tracking the distance of the drones in real time also includes: using trajectory continuity to reduce tracking differences and imposing constraints on the maximum distance box deviations of adjacent columns of drones to maintain tracking accuracy.

5. The lightweight and efficient 3D tracking and identification method of UAV based on millimeter wave radar according to claim 3 is characterized in that: The step S2 of alleviating the fluctuation of Doppler spectrum intensity and static background noise level in different ranges specifically includes: Preprocessing is performed using a spectrum subtraction algorithm to subtract an estimated average background noise spectrum calculated by averaging the background noise; The maximum cumulative folding result is calculated using a dynamic programming algorithm and passed through D S Backtracking in the space identifies the best UAV trajectory, backtracking from the last column to the first column, and finally produces the constrained maximum path g, which represents the best tracking solution; A particle filter is used to optimize the initial tracking estimate and mitigate potential errors. A weight is assigned to each particle according to its consistency with the folded result, focusing the particle filter on the most likely UAV trajectory.

6. The lightweight and efficient 3D tracking and identification method of UAV based on millimeter wave radar according to claim 1 is characterized in that: The step S3 specifically includes the following sub-steps: A uniform sampling matrix Ω is randomly generated, each row of which has a single random non-zero entry, and an orthogonal basis of the covariance matrix of the millimeter wave signal is calculated by orthogonal normalization of the sampling matrix; Using the unitary invariance of the spectrum, the Frobenius norm and the low-rank decomposition, we can obtain an approximate signal subspace And generate a pseudo spectrum, the calculation formula is: in: a(θ)=[a0(θ),a1(θ),...,a M-1 (i)] H ; I M represents the M×M identity matrix, Represents the approximate signal subspace, and the peak position in the pseudo spectrum is the direction information of the target, specifically: peak is the peak value judgment function.

7. The lightweight and efficient 3D tracking and identification method of UAV based on millimeter wave radar according to claim 1 is characterized in that: The step S4 comprises the following sub-steps: The feature alignment algorithm is used to align the center frequency with the center of the Doppler spectrum, and linear interpolation is used to fill the gap; To mitigate tracking discrepancies, a series of fixed-length segments are generated, and the maximum folding results of each segment below a preset empirical threshold are excluded; The segmented Doppler spectrum is input into the UAV identification network model Spiking Neural Network (SNN) architecture for UAV classification.

8. The lightweight and efficient 3D tracking and identification method of UAV based on millimeter wave radar according to claim 7 is characterized in that: The UAV identification network model is a pulse neural network SNN architecture of L-128-128-2, where L is the number of Doppler bins in each time slot; a medium frequency spike neuron model is used to construct the basic unit of the pulse neural network SNN architecture, including membrane potential accumulation, spike generation and reset mechanism; the floating point synaptic weight is quantized to 1 bit representation, the information entropy of the binary weight is increased during the forward propagation process, the binary peak value s[t] and the binary weight B are integrated, and the convolution operation is performed using bitwise operation.

Images