High-speed maneuvering target detection method based on deep learning and generalized radon-fourier transform

Abstract

The application discloses a high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transformation, and the method comprises the following steps: obtaining three-dimensional data through echo signal preprocessing; inputting the three-dimensional data into a target / noise classification network to obtain the probability of existence of a target in each distance unit; inputting the data of the distance unit in which the target may exist into a velocity acceleration classification network to output the search interval of the velocity and acceleration of the target; accumulating energy in the search interval by using generalized Radon-Fourier transformation; and finally obtaining a detection result through constant false alarm detection. The application combines the neural network and the generalized Radon-Fourier transformation, and can effectively balance the detection performance and the calculation amount. The application has a high accuracy in estimating the distance unit where the target is located and the motion parameters of the target; in addition, unnecessary search operations can be reduced according to the rough estimation result of the network, so that the calculation amount is greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to a high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform. Background Technology

[0002] With the rapid development of aerospace technology, a large number of high-speed, highly maneuverable aircraft have emerged. These targets typically possess characteristics such as long range, high maneuverability, and small radar cross-section (RCS). Without changing system parameters, long-term coherent accumulation is a good method to improve detection probability. However, high-speed maneuvering targets exhibit range cell migration (RCM) and doppler frequency migration (DFM) phenomena during long-term coherent accumulation, rendering traditional coherent accumulation methods inapplicable.

[0003] In related technologies, long-term coherent accumulation methods for high-speed maneuvering targets are generally divided into three categories. The first category is Radon-based methods, which effectively accumulate echo energy through parameter search, theoretically achieving optimal detection performance. However, due to their large search range, their computational cost is enormous. The second category is Keystone-based methods, which offer high detection performance and low computational cost, but cannot eliminate the effects of DFM (Discrete Function Mechanism). The third category is Correlation-based methods, which have low computational cost but poor detection performance and are sensitive to noise. It is evident that in related technologies, there is a contradiction between detection performance and computational cost in long-term coherent accumulation for high-speed maneuvering targets. Methods with good detection performance have excessively high computational costs, while algorithms with low computational cost have poor detection performance. Achieving high-performance radar detection of high-speed maneuvering targets with low computational cost remains a challenge. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention proposes a high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform.

[0005] The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] A high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform includes:

[0007] Step 1: Obtain the echo signal of the pulse sequence from the radar system;

[0008] Step 2: Perform pulse compression, amplitude normalization, and separation of real and imaginary parts on the echo signal in sequence to obtain three-dimensional data;

[0009] Step 3: Input the 3D data into the trained target / noise classification network to obtain the probability of the presence of a target in each distance unit, and determine the distance units where the target may exist based on the probability.

[0010] Step 4: Input the distance data of the potential target into the trained velocity and acceleration classification network to determine the search interval where the target's velocity and acceleration are located;

[0011] Step 5: Accumulate the energy of the echo signal using the generalized Radon-Fourier transform within the search interval;

[0012] Step 6: The accumulated results are then subjected to constant false alarm rate (CFAR) testing to obtain the detection result.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] This invention provides a high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform. A target / noise classification network pre-emptively eliminates a large number of pure noise samples. Then, the range cell data indicating the possible presence of a target is input into a velocity-acceleration classification network to estimate the target motion parameter range. Within the estimated search range, the generalized Radon-Fourier transform is used to accumulate energy. On the one hand, due to the network's powerful expressive ability and the effective accumulation of target energy by the generalized Radon-Fourier transform, the method exhibits excellent detection performance. On the other hand, by introducing the network, a large number of unnecessary search operations in the generalized Radon-Fourier transform are reduced, and the increased computational cost of the deep separable convolutional network is minimal, resulting in a very low computational cost.

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of a high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform provided in an embodiment of the present invention;

[0017] Figure 2 This is a schematic diagram of a target / noise classification network and velocity acceleration classification network model provided in an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram of the detection probability curves under different signal-to-noise ratios in an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram illustrating the computational complexity analysis of an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0021] Figure 1 This is a flowchart of a high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform provided in an embodiment of the present invention. Please refer to [link / reference]. Figure 1 This invention provides a high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform, comprising:

[0022] Step 1: Obtain the echo signal of the pulse sequence from the radar system;

[0023] Step 2: Perform pulse compression, amplitude normalization, and separation of real and imaginary parts on the echo signal in sequence to obtain three-dimensional data;

[0024] Step 3: Input the 3D data into the trained target / noise classification network to obtain the probability of the presence of a target in each distance unit, and determine the distance units where the target may exist based on the probability.

[0025] Step 4: Input the distance data of the potential target into the trained velocity-acceleration classification network to estimate the search interval where the target's velocity-acceleration is located;

[0026] Step 5: Accumulate the energy of the echo signal using the generalized Radon-Fourier transform within the search interval;

[0027] Step 6: The accumulated results are then subjected to constant false alarm rate (CFAR) testing to obtain the detection result.

[0028] In this embodiment, the radar system first acquires the transmitted signal and the echo signal, and then performs high-speed maneuvering target detection based on the pulse-compressed signal.

[0029] Optionally, step 1, the step of acquiring the echo signal of the pulse sequence from the radar system, includes:

[0030] The echo signal of the pulse sequence is obtained and represented as:

[0031]

[0032] The radar transmitted waveform is a linear frequency modulated signal with a pulse duration of T. p t represents fast time, t m =mT rm = 1, ..., M, where M is the number of pulses, T r It is the pulse repetition period, f c Here, μ is the carrier frequency of the signal, and μ is the modulation slope of the signal. Assuming there are K targets, the radial distance from the k-th target to the radar can be expressed as: v k ,a k These are the target's radial velocity and acceleration, A. k n(t,t) represents the echo amplitude of the k-th target. m The noise is an additively stationary, zero-mean complex Gaussian distribution with noise power σ. 2 ;

[0033] The echo signal is sampled and discretized.

[0034] In this embodiment, the acquired echo signal is sampled according to a preset sampling time interval and a preset number of sampling points. The discretized echo signal data can be represented as follows:

[0035] x m =s r (t=nΔt,t m ), n=1,2,…,N L ;

[0036] Where Δt is the fast time sampling interval, and N L This represents the number of sampling points.

[0037] Optionally, step 2, which involves sequentially performing pulse compression, amplitude normalization, and separation of the real and imaginary parts of the echo signal to obtain three-dimensional data, includes:

[0038] Step 21: Perform a Fast Fourier Transform on the discretized echo signal data;

[0039] Step 22: Perform frequency domain pulse compression on the echo signal after Fast Fourier Transform, and obtain the pulse-compressed echo signal through Inverse Fourier Transform.

[0040] Step 23: Perform amplitude normalization on the pulse-compressed echo data to obtain the echo signal vector.

[0041] Where, x pc =s pc (t=nΔt,t m ), n=1,2,…,N L Post-pulse compression echo signal s pc (t,t m This can be represented as:

[0042]

[0043] in n pc (t,t m A represents the noise signal after pulse compression. k '=A k BT p Let be the amplitude after the k-th target pulse compression. Without considering pulse compression loss, the pulse compression gain is BT. p ;

[0044] Step 24: Extract the real and imaginary parts of the normalized echo data and concatenate them to obtain the three-dimensional data.

[0045]

[0046] Figure 2 This is a schematic diagram of a target / noise classification network and velocity / acceleration classification network model provided in an embodiment of the present invention. Please refer to... Figure 2 In this embodiment, the target / noise classification network may include two parts in terms of structure: a feature extraction layer and a target / noise classification layer.

[0047] Specifically, the input to the feature extraction layer is multi-pulse echo data after pulse compression. In the feature extraction layer, layers 1 to 10 all employ depthwise separable convolutions. The kernel size for layers 1 to 3 is 10*8, for layers 4 to 9 it is 10*7, and for layer 10 it is 8*5. The number of kernels in layers 1 to 10 are 64, 68, 72, 76, 80, 84, 88, 92, 96, and 100, respectively. The activation function is LeakyReLU. After passing through the feature extraction layer, a feature map of size 100*3*3 is obtained, which is input to the target / noise classification network and the velocity acceleration classification network. It is worth noting that a batch normalization layer is added after each convolutional layer to accelerate the convergence speed of the network.

[0048] Furthermore, the input data of the target / noise classification layer is the output of the feature extraction layer, namely the feature map (100*3*3). The target / noise classification layer contains two depthwise separable convolutional layers. The first convolutional layer has 100 convolutional kernels with a kernel size of 3*3 and uses LeakyReLU as the activation function, followed by a batch normalization layer. The second convolutional layer has 2 convolutional kernels with a kernel size of 1*1. Then, the output result is subjected to softmax to obtain the probability O of the presence of the target in each distance unit.

[0049] Optionally, in step 3, the 3D data is input into the trained target / noise classification network to obtain the probability O of the presence of a target in each distance unit.

[0050] In this embodiment, after obtaining the probability O of the presence of a target at each distance cell output by the target / noise classification network model, the probability O is compared with the detection threshold T:

[0051] If (O) l ≥ T, it indicates that there is a target at the l-th distance cell.

[0052] If (O) l < T, it indicates that there is no target at the l-th distance cell.

[0053] Perform subsequent processing on the input data corresponding to the distance cells with targets, including inputting the corresponding feature map data into the speed and acceleration classification layer to output the speed and acceleration estimation interval, and accumulating energy using the generalized Radon-Fourier transform on the estimation interval; no subsequent processing is performed on the distance cells without targets.

[0054] The setting of the threshold value affects the performance of the method. If the threshold value is set too high, the probability of missed detection will be too large, that is, a large number of distance cells with targets will be misjudged as having no targets and no subsequent processing will be performed, resulting in a decrease in detection performance; if the threshold value is set too low, the false alarm rate will be too high, that is, a large number of distance cells without targets will be misjudged as having targets and subsequent processing will be performed on them, bringing a lot of unnecessary computational workload. The threshold value can be set flexibly according to the requirements of the task for detection performance and computational workload. The model provided in the embodiment of the present invention uses a threshold value with a 20% false alarm rate to achieve a compromise between detection performance and computational workload.

[0055] Figure 2 It is a schematic structural diagram of the target / noise classification network and the speed and acceleration classification network model provided in the embodiment of the present invention. Please refer to Figure 2 In this embodiment, the speed and acceleration classification network can include two parts in structure, namely the feature extraction layer and the speed and acceleration classification layer. In order to reduce the computational workload, the speed and acceleration classification network model and the target / noise classification network model share the same feature extraction layer.

[0056] The input data of the speed and acceleration classification layer is the output of the feature extraction layer, that is, the feature map (100*3*3). The speed and acceleration classification layer includes two depthwise separable convolutional layers. The number of convolutional kernels of the first convolutional layer is 100, the size of the convolutional kernel is 3*3, the activation function is LeakyReLU, followed by a batch normalization layer; the number of convolutional kernels of the second convolutional layer is the number of speed and acceleration sub-intervals. In the embodiment of the present invention, the number of speed and acceleration sub-intervals is set to 20, the size of the convolutional kernel is 1*1, and softmax is performed on the output result to estimate the sub-interval in which the target motion speed and acceleration are located.

[0057] Optionally, in step 4, the data of the distance units where the target may exist are input into the trained velocity acceleration classification network to determine the search interval where the target's velocity acceleration is located.

[0058] Optionally, the sub-intervals of the target's velocity and acceleration are divided as follows, assuming the velocity and acceleration ranges of the target to be detected are [-v... max ,v max ] and [-a max ,a max Then the motion parameter space is:

[0059]

[0060] The velocity dimension and acceleration dimension are each uniformly divided into K... v and K a If there are multiple sub-segments, then we can obtain K. v *K a Sub-intervals:

[0061]

[0062] Any target can be assigned a sub-interval based on its motion parameters, and for the sub-interval S kv,ka Assign labels, using S k Let K represent the k-th subinterval, where k = 1, 2, ..., K, and K = K. v *K a is the number of subintervals.

[0063] Optionally, the target / noise classification network model and the velocity / acceleration classification network model are trained according to the following steps:

[0064] Step a: Obtain a training sample set, which is divided into positive and negative samples based on the presence or absence of a target echo signal. Positive samples are those containing a target echo signal and are labeled with the interval where the target velocity and acceleration are located; negative samples are not labeled with the interval where the velocity and acceleration are located. The target motion velocity dimension and acceleration dimension are each uniformly divided into K... v and K a Each sub-segment yields K. v *K a Sub-intervals, denoted by S k Let k represent the k-th sub-interval, and let k be the corresponding velocity-acceleration interval label. Then the sample label can be represented as:

[0065]

[0066] Among them, y label This is a label indicating whether a target exists. If a target exists, then y... label=1; if there is no target, then y label =0. y para This is the label indicating the sub-interval to which the target velocity and acceleration belong; only positive samples have y. para Tag. S k This represents the k-th velocity-acceleration sub-interval.

[0067] Step b: Randomly initialize the network parameters of the neural network to be trained; wherein the neural network to be trained is a deep separable convolutional neural network, including a target / noise classification network and a velocity acceleration classification network;

[0068] Step c: Input training samples into the neural network to be trained for iterative training, and calculate the target / noise classification loss and velocity acceleration classification loss for positive samples, and calculate the target / noise classification loss for negative samples;

[0069] In this embodiment, the loss function can be expressed as:

[0070]

[0071] Among them, y label The true label for sample / noise classification Predict labels for sample / noise classification, y para Classify speed / acceleration using real labels. Predict labels for velocity / acceleration classification, II(y) para =1) is the indicator function, which calculates the velocity acceleration label loss when the sample is a positive sample, and λ is the weighting factor;

[0072] Step d: Determine the current loss value based on the output of the current iteration and the preset loss function;

[0073] Step e: Detect whether the current loss value is less than or equal to the preset threshold; if not, update the network parameters of the neural network to be trained and return to step c;

[0074] In this step, if the current loss value is greater than the preset threshold, the network parameters are updated according to the following formula:

[0075]

[0076] Where θ represents the updated network parameters, α represents the learning rate, which is a positive real number close to 0, and B represents the number of samples in each group.

[0077] Further, after updating the network parameters, the process returns the input training samples to the neural network to be trained, and the next round of training begins.

[0078] Step f: If yes, then training ends, and the target / noise classification network model and velocity acceleration classification network model are obtained.

[0079] Of course, during the training process of the neural network to be trained, the number of iterations can be used to determine whether the training has ended, but this application does not limit this.

[0080] Optionally, in step 4, the search interval can be obtained by following these steps:

[0081] Step 41: Input the distance unit data of possible targets into the trained velocity acceleration network to obtain the velocity acceleration interval label where the predicted target motion parameters are located;

[0082] Step 42: Determine the search range for the target's velocity and acceleration based on the predicted label.

[0083] Optionally, in step 5, the search interval can be divided into sub-intervals according to a preset method; within the sub-intervals, the energy of the echo signal is accumulated using the generalized Radon-Fourier transform.

[0084] The energy of the echo signal is accumulated within the sub-interval using the generalized Radon-Fourier transform.

[0085]

[0086] Where l' is the distance unit that the target / noise classification network determines to be the presence of a target, and (v',a')∈S k S k Let k be the k-th velocity-acceleration sub-interval, and This represents the search trajectory determined by the initial search distance r', initial search velocity v', and initial search acceleration a'.

[0087] Optionally, in step 6, the accumulated results are subjected to constant false alarm rate (CFAR) detection to obtain the detection result. The threshold value of CFAR detection is related to the false alarm probability.

[0088] The following simulation experiments demonstrate the effectiveness of the high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform.

[0089] I. Experimental conditions:

[0090] The main parameters of the radar used in the simulation experiment are shown in Table 1. In the simulation experiment, the velocities of all targets were within the range of [-500 m / s, 500 m / s], and the accelerations were within the range of [-150 m / s²]. 2 150m / s 2 Within the interval [ ]. The training set samples are selected from 10 [ ]. 6radar echo signals of the target and 10 6 For a pure noise signal, the signal-to-noise ratio of the compressed radar echo pulse follows a uniform distribution from -6dB to 3dB.

[0091] Hardware platform: AMD Ryzen 7 4800H with Radeon Graphics @ 2.90GHz, 16.0GB RAM; Software platform: PyCharm, Matlab. See Table 1:

[0092] Table 1 Simulation Parameters

[0093] carrier frequency 2GHz Transmit waveform bandwidth 40MHz Sampling frequency 80MHz Pulse repetition frequency 200Hz Coherent accumulation pulse number 64 Pulse duty cycle 10％

[0094] II. Experimental Content and Results:

[0095] The proposed algorithm uses a depthwise separable convolutional neural network designed based on radar system parameters. Figure 2 This is the neural network structure used in the simulation experiment. The network input receptive field size is 91*64, the target / noise classification network outputs a vector of length 2, the target velocity acceleration interval is divided into 20 sub-intervals, and the velocity acceleration network outputs a vector of length 20.

[0096] Figure 3 These are the detection probability curves (false alarm rate P) of embodiments of the present invention under different signal-to-noise ratios. fa =10 -6 The graph shows the detection probability curves as follows: blue curves represent the ideal detection probability curve; orange curves represent the detection probability curve using the Generalized Radon-Fourier Transform (GRFT); red curves represent the detection probability curve of the proposed FSGRFT algorithm (100% false alarm rate for the target / noise classification network), with the target / noise network output threshold set to 0, meaning all distance units are fed into the velocity / acceleration classification network for further processing; and purple curves represent the detection probability curve of the proposed FSGRFT algorithm (20% false alarm rate for the target / noise classification network), with the target / noise network output threshold set to 0.006538 (20% false alarm rate for the target / noise classification network). As can be seen from the graph, the GRFT algorithm achieves near-ideal detection performance; the signal-to-noise ratio loss of FSGRFT (100% false alarm rate for the target / noise classification network) compared to the ideal curve is less than 1 dB; and the signal-to-noise ratio loss of FSGRFT (20% false alarm rate for the target / noise classification network) compared to the ideal curve is less than 1.2 dB.

[0097] Figure 4This is a computational analysis of an embodiment of the present invention. The blue curve represents the computational cost required by the GRFT algorithm; the orange curve represents the computational cost required by the proposed FSGRFT algorithm (target / noise classification network with 100% false alarms); and the green curve represents the computational cost required by the proposed FSGRFT algorithm (target / noise classification network with 20% false alarms). It can be seen that the FSGRFT algorithm (target / noise classification network with 100% false alarms) has a signal-to-noise ratio loss of less than 0.7 dB compared to the GRFT algorithm, and its computational cost is reduced to 1 / 17.91 of the original. The FSGRFT algorithm (target / noise classification network with 100% false alarms) has a signal-to-noise ratio loss of less than 0.9 dB compared to the GRFT algorithm, and its computational cost is reduced to 1 / 63.15 of the original.

[0098] This invention combines a deep classifiable convolutional neural network with the GRFT algorithm, which can significantly reduce the computational load of the algorithm while ensuring detection performance, thus achieving a trade-off between detection performance and computational load.

[0099] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0100] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0101] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0102] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform, characterized in that, Includes the following steps: Step 1: Obtain the echo signal of the pulse sequence from the radar system; Step 2: Perform pulse compression, amplitude normalization, and separation of real and imaginary parts on the echo signal in sequence to obtain three-dimensional data; Step 3: Input the 3D data into the trained target / noise classification network to obtain the probability of the presence of a target in each distance unit, and determine the distance units where the target may exist based on the probability. Step 4: Input the distance data of the possible target into the trained velocity and acceleration classification network to estimate the search interval where the target's velocity and acceleration are located; Step 5: Accumulate the energy of the echo signal using the generalized Radon-Fourier transform within the search interval; Step 6: The accumulated results are then subjected to constant false alarm rate (CFAR) testing to obtain the detection result; The target / noise classification network consists of two parts: a feature extraction layer and a target / noise classification layer. The velocity acceleration classification network also consists of two parts: a feature extraction layer and a velocity acceleration classification layer. The target / noise classification network model and the velocity acceleration classification network model share the same feature extraction layer. The input data for the target / noise classification layer is the output of the feature extraction layer. The target / noise classification layer contains two depthwise separable convolutional layers. The first convolutional layer has 100 kernels, each 3*3 in size, and uses LeakyReLU activation function, followed by a batch normalization layer. The second convolutional layer has 2 kernels, each 1*1 in size, and performs softmax on the output to determine the probability of a target being present at that distance unit. The input data for the velocity-acceleration classification layer is the output of the feature extraction layer. The velocity-acceleration classification layer contains two depthwise separable convolutional layers. The first convolutional layer has 100 convolutional kernels, each 3*3 in size, and uses LeakyReLU activation function, followed by a batch normalization layer. The second convolutional layer has the number of convolutional kernels equal to the number of velocity-acceleration sub-intervals, each 1*1 in size. Softmax is applied to the output to estimate the target's velocity and acceleration interval. The target / noise classification network model and velocity acceleration classification network model mentioned in step 3 are trained according to the following steps: Step a: Obtain a training sample set, which is divided into positive and negative samples based on the presence or absence of a target echo signal. Positive samples are those containing a target echo signal and are labeled with the interval where the target velocity and acceleration are located; negative samples are not labeled with the interval where the velocity and acceleration are located. The target motion velocity dimension and acceleration dimension are then evenly divided into... and Each sub-segment yields... Sub-intervals, using Indicates the first The number of sub-intervals, located at the... The labels for each sample in the motion parameter range are set as follows: Then the label of the sample can be represented as: ， Step b: Randomly initialize the network parameters of the neural network to be trained; wherein the neural network to be trained includes a target / noise classification network and a velocity acceleration classification network, and the neural network to be trained is a deep separable convolutional neural network; Step c: Input training samples into the neural network to be trained for iterative training, and calculate the target / noise classification loss and velocity acceleration classification loss for positive samples, and calculate the target / noise classification loss for negative samples; The loss function can be expressed as: ， in, The true label for classifying the sample / noise. Predict labels for sample / noise classification. Classify speed and acceleration using real labels. Predict labels for velocity acceleration classification. The indicator function is used to calculate the velocity acceleration label loss when the sample is a positive sample. Step d: Determine the current loss value based on the output of the current iteration and the preset loss function; Step e: Detect whether the current loss value is less than or equal to the preset threshold; if not, update the network parameters of the neural network to be trained and return to step c; Step f: If yes, then training ends, and the target / noise classification network model and velocity acceleration classification network model are obtained; If the target's velocity and acceleration ranges are respectively and Then the motion parameter space is: The velocity dimension and acceleration dimension are each uniformly divided into and Each sub-segment yields... Sub-intervals: ； For any target, assign it a sub-interval based on its motion parameters, and then assign the sub-interval... Assigning labels, using Indicates the first Sub-intervals, , This represents the number of sub-intervals to search.

2. The high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform according to claim 1, characterized in that, Step 1 includes: Step 11, obtain the echo signal of the pulse sequence, represented as: ； The radar transmitted waveform is a linear frequency modulated signal with a pulse duration of [missing information]. , It represents fast time. , The number of pulses. It is the pulse repetition period. It is the carrier frequency of the signal. It is the modulation slope of the signal, assuming there is The first goal, the... The radial distance from a target to the radar can be expressed as: , These are the target's radial velocity and acceleration, respectively. Representing the The echo amplitude of each target, It is an additively stationary, zero-mean, complex Gaussian distribution with noise power of ; Step 12, for the echo signal Discretization is performed in the fast time dimension according to a preset sampling time interval and a preset number of sampling points. The discretized echo signal can be represented as: ； in, For fast sampling intervals, This represents the number of sampling points.

3. The high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform according to claim 2, characterized in that, Step 2 includes: Step 21: Perform a Fast Fourier Transform on the discretized echo signal data; Step 22: Perform frequency domain pulse compression on the echo signal after the fast Fourier transform, and obtain the pulse-compressed echo signal through inverse Fourier transform. Step 23: Perform amplitude normalization on the pulse-compressed echo data to obtain the echo signal vector; Step 24: Extract the real and imaginary parts of the normalized echo data and stitch them together to obtain three-dimensional data.

4. The high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform according to claim 3, characterized in that, The echo signal after pulse compression in step 22 is represented as follows: ; The echo signal vector in step 23 is represented as follows: ; The three-dimensional data obtained in step 24 is represented as follows .

5. The high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform according to claim 1, characterized in that, Step 3 includes: Step 31: Input the 3D data into the trained target / noise classification network to obtain the probability of the presence of a target in each distance unit; Step 32: Compare the probability of the target existing in the distance cell with a preset detection threshold; Step 33: If the probability of a target existing in a distance cell is greater than a preset detection threshold, then it is considered that a target may exist in that distance cell; Step 34: If the probability of the target existing in the distance cell is less than or equal to the preset detection threshold, then it is considered that there is no target in the distance cell.

6. The high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform according to claim 1, characterized in that, Step 4 includes: Step 41: Input the distance unit data of possible targets into the trained velocity acceleration network to obtain the velocity acceleration interval label where the predicted target motion parameters are located; Step 42: Determine the search range for the target's velocity and acceleration based on the predicted label.

7. The high-speed maneuvering target detection method based on deep learning and generalized Radon-Fourier transform according to claim 1, characterized in that, In step 5, the energy of the echo signal is accumulated within the search interval using the generalized Radon-Fourier transform. The accumulated energy can be expressed as: in, The distance cell that the target / noise classification network identifies as containing a target. , For the first A velocity-acceleration sub-interval, and Indicates the initial search distance Initial search speed and initial acceleration of the search The determined search trajectory.

Images