Flying bird and unmanned aerial vehicle identification method based on compensated multi-frame target slice and flight path fusion

Abstract

The invention discloses a bird and unmanned aerial vehicle identification method based on compensated multi-frame target slice and flight path fusion, and the method specifically comprises the steps: firstly, obtaining a multi-frame R-D domain slice data set of a rotor unmanned aerial vehicle and a bird through the staring distance compensation and Doppler compensation operation; secondly, a bidirectional GRU network model fusing track information is adopted, multi-frame target R-D domain slice data are input, and target feature extraction and integration are completed through double-layer convolution and a full connection layer; then, target track features are added into the integrated target feature set, and the target track features and the integrated target feature set are input into a bidirectional GRU network together; and finally, completing the output of a target identification result label through a full connection layer, and completing an identification task. According to the method, the track information and the multi-frame target information are combined, the track features and the motion features of the target are fully utilized, the target recognition accuracy is improved, and the method is simple, high in recognition speed and high in real-time performance.

Description

Technical Field

[0001] This invention relates to the field of target recognition technology, and in particular to a method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths. Background Technology

[0002] With the development of technology, low-altitude, slow-speed, small drones have been applied in various military and civilian fields, including logistics delivery, pesticide spraying, low-altitude reconnaissance, and long-range attacks. As the demand for drone applications increases, drone defense has gradually become a prominent issue.

[0003] Radar, as an important drone detection device, is currently the mainstream drone detection equipment. However, because bird targets are quite similar to drones in size, reflection characteristics, and movement patterns, the difficulty of distinguishing between drones and birds is greatly increased.

[0004] The existing technologies for identifying drones and birds mainly have the following problems: (1) lack of drone datasets with different postures; (2) lack of intuitive and effective methods for identifying drone and bird data; and (3) lack of real-time classification technology that can be applied to actual radar systems.

[0005] Therefore, there is an urgent need to invent a highly accurate and real-time bird and drone identification method to solve the problem of accurate and real-time target identification of drones and birds. Summary of the Invention

[0006] The purpose of this invention is to provide a highly accurate and real-time method for bird and UAV identification based on the fusion of compensated multi-frame target slices and flight paths.

[0007] The technical solution to achieve the purpose of this invention is: a method for bird and UAV identification based on the fusion of compensated multi-frame target slices and flight paths, comprising the following steps:

[0008] Step 1: Obtain multi-frame RD domain slice datasets of rotary-wing UAVs and birds by performing gaze distance compensation and Doppler compensation calculations;

[0009] Step 2: Input the multi-frame RD domain slice dataset of rotary-wing UAVs and birds into a two-layer convolutional network to extract the spatial features of the targets;

[0010] Step 3: Input the target spatial features extracted in Step 2 into the fully connected layer to generate the target feature vector;

[0011] Step 4: Concatenate and fuse the distance, speed, bearing, and elevation trajectory information with the target feature vector generated in Step 3;

[0012] Step 5: Input the concatenated and fused target feature vector into the bidirectional GRU network to obtain a more comprehensive contextual representation;

[0013] Step 6: Input the processing result from Step 5 into the fully connected layer for target recognition;

[0014] Step 7: Determine if the number of processing steps has reached the number of samples. If the number of samples has not been reached, perform a forward sliding window and then return to step 2 to continue the loop processing flow of steps 2 to 6. If the number of samples has been reached, output the recognition result.

[0015] Furthermore, the acquisition of the multi-frame RD domain slice dataset of rotary-wing UAVs and birds described in step 1 is as follows:

[0016] The RD domain slice dataset is a slice of data containing target feature information extracted from the corresponding RD domain dataset based on the trajectory information obtained by radar tracking during the actual flight of the target, including distance, azimuth, speed and time information.

[0017] Furthermore, the RD domain dataset is obtained in the following way:

[0018] Raw I / Q data of various types of rotorcraft UAVs and birds are acquired by a high data rate staring radar over a large airspace. Staring distance compensation and Doppler compensation calculations are performed on the data to finally obtain the RD domain dataset of rotorcraft UAVs and birds.

[0019] The high data rate staring radar with a large airspace is a radar device with 360-degree omnidirectional coverage and uses simultaneous multi-digital beam receiving technology.

[0020] Furthermore, the specific steps for the gaze distance compensation and Doppler compensation calculations are as follows:

[0021] Step 1-1: Arrange the radar echo IQ data according to three dimensions: azimuth, pulse, and range. After pulse compression processing, obtain... Where b represents the azimuth wave position; n represents the pulse number, corresponding to the slow time domain; and r represents the range unit, corresponding to the fast time domain.

[0022] Steps 1-2: For each azimuth position b, use Keystone transform to compensate for range migration, and obtain range-compensated data samples. ;

[0023] Steps 1-3: For each azimuth position b and range cell r, use the Doppler compensation method with joint constraints of computational complexity and detection probability to... Processing is performed to achieve energy focusing of the target in the range-Doppler domain.

[0024] Furthermore, in steps 1-3, for each azimuth position b and range unit r, the Doppler compensation method with joint constraints of computational complexity and detection probability is used to... Processing is performed to achieve target energy focusing in the range-Doppler domain, specifically including:

[0025] Step 1-3-1: Optimize the number of first-stage acceleration channels based on the criterion of minimizing computational complexity :

[0026]

[0027]

[0028] Step 1-3-2: Optimize the number of acceleration channels based on detection probability constraints:

[0029] ①Increase the number of first-stage acceleration channels Initialize to ;

[0030] ②For the acceleration range Uniform sampling Individual acceleration values:

[0031]

[0032] In the formula, This is the index for the first-level acceleration channel;

[0033] ③ Extract echo IQ data from the target location's position and range cells. For each Doppler demodulation processing is performed on radar echoes that have had their range migration eliminated in the slow time domain:

[0034]

[0035] In the formula, This is a slow-time domain sample of the target echo after distance compensation. For slow time, Wavelength;

[0036] ④ Sample data after Doppler demodulation Perform coherent accumulation and search for the acceleration channel index number corresponding to the coarse estimate of the target radial acceleration. :

[0037]

[0038] ⑤ Detection probability constraint decision: If satisfied If the result is positive, the detection is successful; otherwise, the detection fails. Repeat the experiment multiple times, count the number of successful detections, and estimate the detection probability. ;

[0039] ⑥ If satisfied The optimal number of first-order acceleration channels under the joint constraints of output computational complexity and detection probability. Perform second-stage acceleration channel demodulation processing; otherwise, let Turn ②;

[0040] Step 1-3-3: Second-stage acceleration channel demodulation:

[0041] ① Based on the optimal number of first-level acceleration channels Calculate the number of second-stage acceleration channels. :

[0042]

[0043] In the formula, For floor operations, The value of should be such that the remaining acceleration component after the second-stage acceleration compensation is within one Doppler resolution cell;

[0044] ②Based on the first-level acceleration channel index number Reposition the acceleration range to Further analysis of the acceleration range Refined uniform sampling One acceleration value;

[0045]

[0046] ③ For each Fine Doppler demodulation of radar echoes in the slow time domain is performed, and coherent accumulation of sample data after fine Doppler demodulation is performed.

[0047]

[0048] In the formula, k is the discrete Doppler frequency;

[0049] ④ Search for the acceleration channel index number corresponding to the precise estimate of the target's radial acceleration. ;

[0050]

[0051] ⑤ Output the long-term coherent accumulation result .

[0052] Furthermore, the RD domain slice dataset in step 1 has a slice size of 32*32 pixels, and each frame slice contains complete target features in the range-Doppler domain; the dataset is a manually labeled dataset, including samples of various types of rotary-wing UAVs and birds, with no less than 1,000 samples in each category, covering various flight attitudes.

[0053] Furthermore, step 2 is specifically as follows:

[0054] Input a multi-frame RD domain slice dataset into a two-layer convolutional network to extract the spatial features of the target.

[0055] The two-layer convolutional network includes convolutional layer 1, pooling layer 1, convolutional layer 2, and pooling layer 2. The convolutional layer 1 has a kernel size of 3×3 and 32 output channels. Pooling layer 1 uses 2×2 max pooling. The convolutional layer 2 has a kernel size of 3×3 and 64 output channels. Pooling layer 2 uses 2×2 max pooling.

[0056] Furthermore, in step 3, the fully connected layer uses the ReLU activation function, and the generated target feature vector has a dimension of 256.

[0057] Furthermore, step 4 specifically includes:

[0058] The distance, speed, bearing, and elevation angle trajectory information are weighted and fused with the target feature vector. The weighted fusion method includes: normalizing the distance, speed, bearing, and elevation angle trajectory information respectively, and then weighting and summing them with the target feature vector according to preset weights to obtain the fused feature vector.

[0059] The weights for distance track information are 0.3, speed track information is 0.4, azimuth track information is 0.15, elevation track information is 0.15, and the dimension of the fused feature vector is 260.

[0060] Furthermore, step 5 is detailed below:

[0061] The concatenated and fused feature vectors are input into a bidirectional GRU network to obtain a more comprehensive contextual representation.

[0062] The bidirectional GRU network contains two layers of GRU units, each layer containing 128 units. The input sequence is N consecutive frames of data, where N ranges from 4 to 8 frames.

[0063] The input sequence of the bidirectional GRU network is N consecutive frames of data. The value of N is dynamically adjusted according to the target's movement speed. The faster the movement speed, the smaller the value of N.

[0064] Compared with the prior art, the present invention has the following significant advantages: (1) By utilizing the compensated RD domain target slice information, the signal-to-noise ratio of the target can be improved and more accurate target spatial information can be obtained; (2) By combining track information with target information of multiple frames, the trajectory features and motion features of the target are fully utilized, which can effectively improve the target recognition accuracy; (3) By normalizing and weighting the different track information, the utilization of track information is effectively improved, and the method is simple, fast in recognition, and has strong real-time performance. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating a method for bird and drone identification based on the fusion of compensated multi-frame target slices and flight paths according to the present invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0067] like Figure 1 As shown, the present invention provides a method for bird and UAV identification based on the fusion of compensated multi-frame target slices and flight paths, comprising the following steps:

[0068] Step 1: Obtain multi-frame RD domain slice datasets of rotary-wing UAVs and birds by performing gaze distance compensation and Doppler compensation calculations;

[0069] Step 2: Input the multi-frame RD domain slice dataset of rotary-wing UAVs and birds into a two-layer convolutional network to extract the spatial features of the targets;

[0070] Step 3: Input the target spatial features extracted in Step 2 into the fully connected layer to generate the target feature vector;

[0071] Step 4: Concatenate and fuse the distance, speed, bearing, and elevation trajectory information with the target feature vector generated in Step 3;

[0072] Step 5: Input the concatenated and fused target feature vector into the bidirectional GRU network to obtain a more comprehensive contextual representation;

[0073] Step 6: Input the processing result from Step 5 into the fully connected layer for target recognition;

[0074] Step 7: Determine if the number of processing steps has reached the number of samples. If the number of samples has not been reached, perform a forward sliding window and then return to step 2 to continue the loop processing flow of steps 2 to 6. If the number of samples has been reached, output the recognition result.

[0075] As a specific example, the acquisition of multi-frame RD domain slice datasets of rotary-wing UAVs and birds described in step 1 is as follows:

[0076] The RD domain slice dataset is a slice of data containing target feature information extracted from the corresponding RD domain dataset based on the trajectory information obtained by radar tracking during the actual flight of the target, including distance, azimuth, speed and time information.

[0077] As a specific example, the RD domain dataset is obtained in the following way:

[0078] Raw I / Q data of various types of rotorcraft UAVs and birds are acquired by a high data rate staring radar over a large airspace. Staring distance compensation and Doppler compensation calculations are performed on the data to finally obtain the RD domain dataset of rotorcraft UAVs and birds.

[0079] The high data rate staring radar with a large airspace is a radar device with 360-degree omnidirectional coverage and uses simultaneous multi-digital beam receiving technology.

[0080] As a specific example, the gaze distance compensation and Doppler compensation calculations are performed using the following steps:

[0081] Step 1-1: Arrange the radar echo IQ data according to three dimensions: azimuth, pulse, and range. After pulse compression processing, obtain... Where b represents the azimuth wave position; n represents the pulse number, corresponding to the slow time domain; and r represents the range unit, corresponding to the fast time domain.

[0082] Steps 1-2: For each azimuth position b, use Keystone transform to compensate for range migration, and obtain range-compensated data samples. ;

[0083] Steps 1-3: For each azimuth position b and range cell r, use the Doppler compensation method with joint constraints of computational complexity and detection probability to... Processing is performed to achieve energy focusing of the target in the range-Doppler domain.

[0084] As a specific example, steps 1-3 describe the application of a Doppler compensation method with computational complexity and detection probability joint constraints for each azimuth position b and range unit r. Processing is performed to achieve target energy focusing in the range-Doppler domain, specifically including:

[0085] Step 1-3-1: Optimize the number of first-stage acceleration channels based on the criterion of minimizing computational complexity :

[0086]

[0087]

[0088] Step 1-3-2: Optimize the number of acceleration channels based on detection probability constraints:

[0089] ①Increase the number of first-stage acceleration channels Initialize to ;

[0090] ②For the acceleration range Uniform sampling Individual acceleration values:

[0091]

[0092] In the formula, This is the index for the first-level acceleration channel;

[0093] ③ Extract echo IQ data from the target location's position and range cells. For each Doppler demodulation processing is performed on radar echoes that have had their range migration eliminated in the slow time domain:

[0094]

[0095] In the formula, This is a slow-time domain sample of the target echo after distance compensation. For slow time, Wavelength;

[0096] ④ Sample data after Doppler demodulation Perform coherent accumulation and search for the acceleration channel index number corresponding to the coarse estimate of the target radial acceleration. :

[0097]

[0098] ⑤ Detection probability constraint decision: If satisfied If the result is positive, the detection is successful; otherwise, the detection fails. Repeat the experiment multiple times, count the number of successful detections, and estimate the detection probability. ;

[0099] ⑥ If satisfied The optimal number of first-order acceleration channels under the joint constraints of output computational complexity and detection probability. Perform second-stage acceleration channel demodulation processing; otherwise, let Turn ②;

[0100] Step 1-3-3: Second-stage acceleration channel demodulation:

[0101] ① Based on the optimal number of first-level acceleration channels Calculate the number of second-stage acceleration channels. :

[0102]

[0103] In the formula, For floor operations, The value of should be such that the remaining acceleration component after the second-stage acceleration compensation is within one Doppler resolution cell;

[0104] ②Based on the first-level acceleration channel index number Reposition the acceleration range to Further analysis of the acceleration range Refined uniform sampling One acceleration value;

[0105]

[0106] ③ For each Fine Doppler demodulation of radar echoes in the slow time domain is performed, and coherent accumulation of sample data after fine Doppler demodulation is performed.

[0107]

[0108] In the formula, k is the discrete Doppler frequency;

[0109] ④ Search for the acceleration channel index number corresponding to the precise estimate of the target's radial acceleration. ;

[0110]

[0111] ⑤ Output the long-term coherent accumulation result .

[0112] As a specific example, the RD domain slice dataset in step 1 has a slice size of 32*32 pixels, and each frame slice contains complete target features in the range-Doppler domain; the dataset is a manually labeled dataset, including samples of various types of rotary-wing UAVs and birds, with no less than 1,000 samples in each category, covering a variety of flight attitudes.

[0113] As a specific example, step 2 is as follows:

[0114] Input a multi-frame RD domain slice dataset into a two-layer convolutional network to extract the spatial features of the target.

[0115] The two-layer convolutional network includes convolutional layer 1, pooling layer 1, convolutional layer 2, and pooling layer 2. The convolutional layer 1 has a kernel size of 3×3 and 32 output channels. Pooling layer 1 uses 2×2 max pooling. The convolutional layer 2 has a kernel size of 3×3 and 64 output channels. Pooling layer 2 uses 2×2 max pooling.

[0116] As a specific example, in step 3, the fully connected layer uses the ReLU activation function, and the resulting target feature vector has a dimension of 256.

[0117] As a specific example, step 4 is as follows:

[0118] The distance, speed, bearing, and elevation angle trajectory information are weighted and fused with the target feature vector. The weighted fusion method includes: normalizing the distance, speed, bearing, and elevation angle trajectory information respectively, and then weighting and summing them with the target feature vector according to preset weights to obtain the fused feature vector.

[0119] The weights for distance track information are 0.3, speed track information is 0.4, azimuth track information is 0.15, elevation track information is 0.15, and the dimension of the fused feature vector is 260.

[0120] As a specific example, step 5 is as follows:

[0121] The concatenated and fused feature vectors are input into a bidirectional GRU network to obtain a more comprehensive contextual representation.

[0122] The bidirectional GRU network contains two layers of GRU units, each layer containing 128 units. The input sequence is N consecutive frames of data, where N ranges from 4 to 8 frames.

[0123] The input sequence of the bidirectional GRU network is N consecutive frames of data. The value of N is dynamically adjusted according to the target's movement speed. The faster the movement speed, the smaller the value of N.

[0124] Example 1

[0125] This embodiment employs a dual-layer convolutional-GRU network model that fuses trajectory information. It takes a compensated multi-frame target slice RD domain dataset as input, extracts target features through dual-layer convolution and fully connected layers, adds normalized target trajectory features to the target feature set, inputs it into the bidirectional GRU network, and finally outputs the target recognition result label through a fully connected layer, thus completing the recognition task. The specific steps are as follows:

[0126] Step 1: Acquire raw I / Q data of various types of rotorcraft UAVs / birds using a high data rate staring radar over a large airspace. Perform staring distance compensation and Doppler compensation calculations on this data to finally obtain the RD domain dataset of the rotorcraft UAV / bird. The RD domain slice dataset is obtained by extracting the track information (including distance, azimuth, speed, and time information) obtained by radar tracking during the actual flight of the target.

[0127] Step 2: Input the multi-frame RD domain slice dataset into a two-layer convolutional network to extract the spatial features of the target. The two-layer convolutional network includes convolutional layer 1, pooling layer 1, convolutional layer 2, and pooling layer 2. The kernel size of convolutional layer 1 is 3×3, and the number of output channels is 32. Pooling layer 1 uses 2×2 max pooling. The kernel size of convolutional layer 2 is 3×3, and the number of output channels is 64. Pooling layer 2 uses 2×2 max pooling.

[0128] Step 3: Input the target space features extracted in Step 2 into the fully connected layer to generate the target feature vector, with a feature vector dimension of 256;

[0129] Step 4: Weighted fusion of distance, speed, azimuth, and elevation trajectory information with the target feature vector, where the weight of distance trajectory information is 0.3, the weight of speed trajectory information is 0.4, the weight of azimuth trajectory information is 0.15, and the weight of elevation trajectory information is 0.15. The dimension of the fused feature vector is 260.

[0130] Step 5: Input the fused feature vector into the bidirectional GRU network. The bidirectional GRU network contains two layers of GRU units, each layer containing 128 units. The input sequence is N consecutive frames of data, and the value of N ranges from 4 to 8 frames.

[0131] Step 6: Determine the target type by inputting the output and input of the bidirectional GRU network to the fully connected layer;

[0132] Step 7: Determine if the number of processing iterations has reached the required number of samples. If not, perform a forward sliding window, then proceed to Step 2 and continue the loop processing from Step 2 to Step 6. If the required number of samples has been reached, output the recognition result.

[0133] This invention provides a method for bird and UAV identification by fusing compensated multi-frame target slices and flight paths. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths, characterized in that, Includes the following steps: Step 1: Obtain multi-frame RD domain slice datasets of rotary-wing UAVs and birds by performing gaze distance compensation and Doppler compensation calculations; Step 2: Input the multi-frame RD domain slice dataset of the rotary-wing UAV and flying bird into a two-layer convolutional network to extract the spatial features of the target; Step 3: Input the target spatial features extracted in Step 2 into the fully connected layer to generate the target feature vector; Step 4: Concatenate and fuse the distance, speed, bearing, and elevation trajectory information with the target feature vector generated in Step 3; Step 5: Input the concatenated and fused target feature vector into the bidirectional GRU network to obtain a more comprehensive contextual representation; Step 6: Input the processing result from Step 5 into the fully connected layer for target recognition; Step 7: Determine if the number of processing steps has reached the number of samples. If the number of samples has not been reached, perform a forward sliding window and then return to step 2 to continue the loop processing flow of steps 2 to 6. If the number of samples has been reached, output the recognition result.

2. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 1, characterized in that, Step 1, which involves obtaining a multi-frame RD domain slice dataset of rotary-wing UAVs and birds, is detailed below: The RD domain slice dataset is a slice of data containing target feature information extracted from the corresponding RD domain dataset based on the trajectory information obtained by radar tracking during the actual flight of the target, including distance, azimuth, speed and time information.

3. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 2, characterized in that, The RD domain dataset was obtained in the following way: Raw I / Q data of various types of rotorcraft UAVs and birds are acquired by a high data rate staring radar over a large airspace. Staring distance compensation and Doppler compensation calculations are performed on the data to finally obtain the RD domain dataset of rotorcraft UAVs and birds. The high data rate staring radar with a large airspace is a radar device with 360-degree omnidirectional coverage and uses simultaneous multi-digital beam receiving technology.

4. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 3, characterized in that, The specific steps for the gaze distance compensation and Doppler compensation calculations are as follows: Step 1-1: Arrange the radar echo IQ data according to three dimensions: azimuth, pulse, and range. After pulse compression processing, obtain... , where b represents the azimuth wave position; n represents the pulse number, corresponding to the slow time domain; r represents the distance unit, corresponding to the fast time domain; Steps 1-2: For each azimuth position b, use Keystone transform to compensate for range migration, and obtain range-compensated data samples. ; Steps 1-3: For each azimuth position b and range cell r, use the Doppler compensation method with joint constraints of computational complexity and detection probability to... Processing is performed to achieve energy focusing of the target in the range-Doppler domain.

5. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 4, characterized in that, Steps 1-3 describe the application of the Doppler compensation method, which combines computational complexity and detection probability constraints, to each azimuth position b and range unit r. Processing is performed to achieve target energy focusing in the range-Doppler domain, specifically including: Step 1-3-1: Optimize the number of first-stage acceleration channels based on the criterion of minimizing computational complexity : ； ； Step 1-3-2: Optimize the number of acceleration channels based on detection probability constraints: ①Increase the number of first-stage acceleration channels Initialize to ; ②For the acceleration range Uniform sampling Individual acceleration values: ； In the formula, This is the index for the first-level acceleration channel; ③ Extract echo IQ data from the target location's position and range cells. For each Doppler demodulation processing is performed on radar echoes that have had their range migration eliminated in the slow time domain: ； In the formula, This is a slow-time domain sample of the target echo after distance compensation. For slow time, Wavelength; ④ Sample data after Doppler demodulation Perform coherent accumulation and search for the acceleration channel index number corresponding to the coarse estimate of the target radial acceleration. : ； ⑤ Detection probability constraint decision: If satisfied If the result is positive, the detection is successful; otherwise, the detection fails. Repeat the experiment multiple times, count the number of successful detections, and estimate the detection probability. ; ⑥ If satisfied The optimal number of first-order acceleration channels under the joint constraints of output computational complexity and detection probability. Perform second-stage acceleration channel demodulation processing; otherwise, let Turn ②; Step 1-3-3: Second-stage acceleration channel demodulation: ① Based on the optimal number of first-level acceleration channels Calculate the number of second-stage acceleration channels. : ； In the formula, For floor operations, The value of should be such that the remaining acceleration component after the second-stage acceleration compensation is within one Doppler resolution cell; ②Based on the first-level acceleration channel index number Reposition the acceleration range to Further analysis of the acceleration range Refined uniform sampling One acceleration value; ； ③ For each Fine Doppler demodulation of radar echoes in the slow time domain is performed, and coherent accumulation of sample data after fine Doppler demodulation is performed. ； In the formula, k is the discrete Doppler frequency; ④ Search for the acceleration channel index number corresponding to the precise estimate of the target's radial acceleration. ; ； ⑤ Output the long-term coherent accumulation result .

6. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 5, characterized in that, The RD domain slice dataset in step 1 has a slice size of 32*32 pixels, and each frame slice contains complete target features in the distance-Doppler domain. The dataset is a manually labeled dataset that includes samples of various types of rotary-wing drones and birds, with no fewer than 1,000 samples in each category, covering a variety of flight attitudes.

7. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 6, characterized in that, Step 2 is described in detail below: Input a multi-frame RD domain slice dataset into a two-layer convolutional network to extract the spatial features of the target. The two-layer convolutional network includes convolutional layer 1, pooling layer 1, convolutional layer 2, and pooling layer 2. The convolutional layer 1 has a kernel size of 3×3 and 32 output channels. Pooling layer 1 uses 2×2 max pooling. The convolutional layer 2 has a kernel size of 3×3 and 64 output channels. Pooling layer 2 uses 2×2 max pooling.

8. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 7, characterized in that, In step 3, the fully connected layer uses the ReLU activation function, and the generated target feature vector has a dimension of 256.

9. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 8, characterized in that, Step 4 specifically involves: The distance, speed, bearing, and elevation angle trajectory information are weighted and fused with the target feature vector. The weighted fusion method includes: normalizing the distance, speed, bearing, and elevation angle trajectory information respectively, and then weighting and summing them with the target feature vector according to preset weights to obtain the fused feature vector. The weights for distance track information are 0.3, speed track information is 0.4, azimuth track information is 0.15, elevation track information is 0.15, and the dimension of the fused feature vector is 260.

10. The method for bird and UAV recognition based on the fusion of compensated multi-frame target slices and flight paths according to claim 9, step 5 is as follows: The concatenated and fused feature vectors are input into a bidirectional GRU network to obtain a more comprehensive contextual representation. The bidirectional GRU network contains two layers of GRU units, each layer containing 128 units. The input sequence is N consecutive frames of data, where N ranges from 4 to 8 frames. The input sequence of the bidirectional GRU network is N consecutive frames of data. The value of N is dynamically adjusted according to the target's movement speed. The faster the movement speed, the smaller the value of N.

Images