A method, device, equipment, medium and product for cross-angle recognition of a UAV

Abstract

The application belongs to the field of unmanned aerial vehicle identification, and provides an unmanned aerial vehicle cross-angle identification method, device, equipment, medium and product, which are used for improving the identification accuracy of unmanned aerial vehicle model types. Based on at least one sample time-frequency graph, a training sample set is obtained. Based on the training sample set including the sample time-frequency graph of the added target category label and the observation angle label, a preset neural network is trained to obtain a target identification model used for identifying the unmanned aerial vehicle model. Then, the micro-Doppler time-frequency graph of the unmanned aerial vehicle to be identified is input into the target identification model to obtain the category identification result output by the target identification model, which is used for indicating the unmanned aerial vehicle model to be identified. Thus, the misjudgment under the unseen angle is avoided, and the accuracy of the unmanned aerial vehicle model category identification is improved.

Description

Technical Field

[0001] This invention relates to the field of drone identification, and more specifically, to a method, apparatus, equipment, medium, and product for cross-angle drone identification. Background Technology

[0002] With the continuous expansion of low-altitude application scenarios, reliable identification of drone targets has become an important issue in low-altitude supervision and security. Existing methods typically rely on FMCW radar to acquire drone echoes and utilize micro-Doppler features generated by rotor motion to identify drone types.

[0003] However, in actual detection, changes in the observation angle of the UAV by the FMCW radar affect the energy distribution and texture morphology of the time-frequency map, causing the same target to exhibit different characteristics at different angles. Existing identification methods mostly use the time-frequency map directly for classification, lacking the distinction between the target's inherent characteristics and the characteristics affected by the angle. This leads to misjudgments and low identification accuracy when the angle is not observed. Summary of the Invention

[0004] This application aims to provide a method for cross-angle identification of unmanned aerial vehicles (UAVs) to solve the problem that existing UAV model and type identification methods are prone to misjudgment and have low identification accuracy when the angle is not visible.

[0005] Firstly, this application provides a method for cross-angle recognition of unmanned aerial vehicles (UAVs), comprising: Obtain at least one sample time-frequency map; the sample time-frequency map is a micro-Doppler time-frequency map of the corresponding UAV model during flight; A training sample set is obtained based on at least one sample time-frequency map. The training sample set includes sample time-frequency maps with added target category labels and observation angle labels. The target category label is used to indicate the UAV model corresponding to the sample time-frequency map, and the observation angle label is used to indicate the observation angle corresponding to the sample time-frequency map. The observation angle is the detection angle when the FMCW radar detects the UAV. Based on the training sample set, a pre-set neural network is trained to obtain a target recognition model for identifying drone models; The micro-Doppler time-frequency image of the UAV to be identified is input into the target recognition model to obtain the category recognition result output by the target recognition model; the category recognition result is used to indicate the model of the UAV to be identified.

[0006] In summary, due to the adoption of the above technical solution, the beneficial effects of this invention are as follows: This application obtains a training sample set based on at least one sample time-frequency map. Based on the training sample set including sample time-frequency maps with added target category labels and observation angle labels, a preset neural network is trained to obtain a target recognition model for identifying drone models. Then, the micro-Doppler time-frequency map of the drone to be identified is input into the target recognition model to obtain the category recognition result output by the target recognition model, indicating the drone model to be identified. Thus, by combining the target's intrinsic features regarding the target category with the angular influence features regarding the observation angle to identify drone models, the distinction between the intrinsic features of the target and the angular influence features is expanded, avoiding misjudgments when the angle is not observed, thereby improving the accuracy of drone model category identification.

[0007] One possible implementation involves obtaining at least one sample time-frequency plot, including: The FMCW radar was used to collect drone target echoes from different models of drones at different observation angles. A two-dimensional Fourier transform is performed on the echoes of each UAV target to generate the corresponding sample time-frequency diagram.

[0008] One possible implementation is that the pre-defined neural network includes a shared feature extraction network, a base classifier, a weak angle decoupling unit, and an adaptive fusion unit; Among them, the shared feature extraction network is used to extract the overall texture, energy distribution and micro-Doppler modulation features of each sample's time-frequency map to obtain the corresponding shared features; The base classifier is used to identify drone categories by sharing features, and obtains effective discriminative information that retains the original features as the base classification features; The weak angle decoupling unit is used to extract UAV category information and observation angle information from the shared features to obtain decoupled classification features and fusion weights; The adaptive fusion unit is used to perform weighted fusion of basic classification features and decoupled classification features through fusion weights to obtain the corresponding classification and recognition results.

[0009] One possible implementation is that the weak angle decoupling unit includes a category feature mapping subunit, an angle feature mapping subunit, a decoupling classifier, and a fusion gating subunit; Among them, the category feature mapping subunit is used to extract discriminative information about the drone category from the shared features as target identity features; The angle feature mapping subunit is used to extract information about the time-frequency feature changes caused by the observation angle from the shared features as angle mapping features; The decoupling classifier is used to classify target identity features to obtain decoupling classification features; The fusion gating subunit generates fusion weights based on the target identity features and angle mapping features.

[0010] One possible implementation involves training a pre-defined neural network in three phases: a first training phase, a second training phase, and a third training phase. The training phase consists of three stages: the first stage is used to train the classification of UAV models to establish stable category discrimination boundaries; the second stage incorporates angle supervision and orthogonal constraints to gradually distinguish information about UAV categories from information about observation angles; and the third stage incorporates weak angle adversarial constraints to reduce residual information about observation angles in target identity features.

[0011] One possible implementation, where the total loss function for training a pre-defined neural network is shown in the following equation: ; in, Losses are categorized by drone type. To monitor losses from different angles, The first weighting coefficient corresponding to the angle supervision loss. The loss is an orthogonal constraint between identity features and angle features. This is the second weighting coefficient corresponding to the orthogonal constraint loss. To mitigate losses from weak angles, The third weighting coefficient corresponds to the loss against weak angles.

[0012] Secondly, this application provides a drone cross-angle recognition device, comprising: The acquisition module is used to acquire at least one sample time-frequency map; the sample time-frequency map is a micro-Doppler time-frequency map of the corresponding UAV model during flight; The training module is used to obtain a training sample set based on at least one sample time-frequency map. The training sample set includes sample time-frequency maps with added target category labels and observation angle labels. The target category label is used to indicate the UAV model corresponding to the sample time-frequency map, and the observation angle label is used to indicate the observation angle corresponding to the sample time-frequency map. The observation angle is the detection angle when the FMCW radar detects the UAV. The training module is also used to train a preset neural network based on a training sample set to obtain a target recognition model for identifying drone models. The identification module is used to input the micro-Doppler time-frequency image of the UAV to be identified into the target identification model and obtain the category identification result output by the target identification model; the category identification result is used to indicate the model of the UAV to be identified.

[0013] Thirdly, this application provides an electronic device comprising: a processor and a memory; the memory storing processor-executable instructions; when the processor is configured to execute the instructions, causing the electronic device to implement the method of the first aspect described above.

[0014] Fourthly, this application provides a computer-readable storage medium comprising: computer software instructions; which, when executed in an electronic device, cause the electronic device to implement the method described in the first aspect.

[0015] Fifthly, this application provides a computer program product comprising a computer program that, when run on an electronic device, causes the electronic device to perform the method described in the first aspect.

[0016] The beneficial effects of the second to fifth aspects mentioned above are described in the corresponding description of the first aspect and will not be repeated here. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a method for cross-angle recognition of unmanned aerial vehicles (UAVs) provided in an embodiment of this application; Figure 2 The UAV target and its micro-Doppler time-frequency image provided in the embodiments of this application; Figure 3 A schematic diagram of a network structure provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating a comparison of recognition accuracy results provided in an embodiment of this application. Figure 5 A schematic diagram of a classification confusion matrix provided in an embodiment of this application; Figure 6 This is a schematic diagram illustrating the composition of a drone cross-angle recognition device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] The terms “comprising” and “having”, and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0019] It should be noted that in the embodiments of this application, the words "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0020] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0021] The following is a detailed description of a cross-angle recognition method for unmanned aerial vehicles (UAVs) provided in this application, with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating a method for cross-angle recognition of unmanned aerial vehicles (UAVs) provided in an embodiment of this application. Figure 2 The image shows a drone target and its micro-Doppler time-frequency diagram provided in the embodiments of this application. Figure 3 This is a schematic diagram of a network structure provided in an embodiment of this application.

[0023] S101. Obtain at least one sample time-frequency diagram.

[0024] Among them, the sample time-frequency diagram is a micro-Doppler time-frequency diagram of the corresponding UAV model during flight.

[0025] In one possible implementation, FMCW (Frequency-Modulated Continuous-Wave) radar is used to collect UAV target echoes of different models at different observation angles, and a two-dimensional Fourier transform is performed on each UAV target echo to generate a corresponding sample time-frequency map.

[0026] S102. Obtain a training sample set based on at least one sample time-frequency map. The training sample set includes sample time-frequency maps with added target category labels and observation angle labels.

[0027] The target category label is used to indicate the UAV model corresponding to the sample time-frequency map, and the observation angle label is used to indicate the observation angle corresponding to the sample time-frequency map. The observation angle is the detection angle when the FMCW radar detects the UAV.

[0028] For example, such as Figure 2 As shown, for the first Each sample, i.e., the drone being detected, has its corresponding time-frequency graph denoted as... The target category label is denoted as This is used to monitor drone categories. The observation angle label is denoted as... This is used to describe the observation angle corresponding to the drone.

[0029] S103. Based on the training sample set, train the preset neural network to obtain a target recognition model for identifying drone models.

[0030] like Figure 3 As shown, the preset neural network includes a shared feature extraction network, a basic classifier, a weak angle decoupling unit, and an adaptive fusion unit; the trained target recognition model is an adaptive fusion angle decoupled feature network (ADFNet).

[0031] Among them, the shared feature extraction network adopts a convolutional neural network structure such as RestNet18 to extract the overall texture, energy distribution and micro-Doppler modulation features of each sample time-frequency map, that is, the corresponding micro-Doppler time-frequency map, to obtain the corresponding shared features. This process can be expressed as shown in the following equation (1): (1); in, For shared features, To share the feature extraction network, This is a time-frequency plot of the sample.

[0032] The base classifier is used to identify drone categories by sharing features, obtaining effective discriminative information that retains the original features as the base classification features. .

[0033] The weak angle decoupling unit is used to extract UAV category information and observation angle information from the shared features to obtain decoupled classification features. And fusion weights.

[0034] Specifically, the weak angle decoupling unit includes a category feature mapping subunit, an angle feature mapping subunit, a decoupling classifier, and a fusion gating subunit.

[0035] Among them, the category feature mapping subunit is used to extract the discrimination information about the UAV category from the shared features as the target identity feature; the angle feature mapping subunit is used to extract the information about the time-frequency feature change caused by the observation angle from the shared features as the angle mapping feature. This process can be expressed as shown in the following equation (2): (2); in, For target identity characteristics, For angle mapping features, For shared features, For category feature mapping subunit, This is the angle feature mapping sub-unit.

[0036] To reduce redundancy between target identity features and angle mapping features, orthogonal constraints are introduced to distribute identity features and angle features in different feature subspaces as much as possible.

[0037] The decoupled classifier is used to classify target identity features, and the resulting decoupled classification features are used to enhance cross-angle robustness.

[0038] The weak angle decoupling unit also includes an angle classifier and an adversarial angle classifier. The angle classifier predicts the angle through angle mapping features to ensure that the angle information is explicitly modeled. The target identity features are input into the adversarial angle classifier through the gradient inversion layer (GRL) to reduce the predictability of the angle in the identity features, thereby suppressing the interference of the observation angle on the category recognition.

[0039] The fusion gating subunit generates fusion weights based on the target identity features and angle mapping features. This process can be expressed as shown in the following equation (3): (3); in, To integrate weights, For the Sigmoid function, W The calculation process involves classifying target identity features using a decoupled classifier and predicting angles for angle mapping features using an angle classifier, where b is a preset fusion constant.

[0040] Specifically, based on the decoupled classification features and the angle prediction features output by the angle classifier, corresponding fusion weights are generated.

[0041] The adaptive fusion unit is used to perform weighted fusion of the basic classification features and the decoupled classification features through fusion weights to obtain the corresponding classification and recognition results. This process can be expressed as shown in the following equation (4): (4); in, To integrate weights, To decouple classification features, Basic classification features, The classification and identification result is the final classification result.

[0042] When the base classifier is more reliable in classifying the basic features of the current sample, the model tends to retain the original discriminative information; when the decoupled classification features output by the decoupled classifier are more reliable, the model increases the contribution of the target identity feature classification result, thereby improving the recognition stability under complex angle conditions.

[0043] Training a pre-defined neural network includes a first training phase, a second training phase, and a third training phase.

[0044] The training phase consists of three stages: the first stage is used to train the classification of UAV models to establish stable category discrimination boundaries; the second stage incorporates angle supervision and orthogonal constraints to gradually distinguish information about UAV categories from information about observation angles; and the third stage incorporates weak angle adversarial constraints to reduce residual information about observation angles in target identity features.

[0045] The total loss function for training the preset neural network is shown in equation (5) below: (5); in, Losses are categorized by drone type. To monitor losses from different angles, The first weighting coefficient corresponding to the angle supervision loss. The loss is an orthogonal constraint between identity features and angle features. This is the second weighting coefficient corresponding to the orthogonal constraint loss. To mitigate losses from weak angles, The third weighting coefficient corresponds to the loss against weak angles.

[0046] This step improves the generalization ability of the FMCW radar UAV identification model under unseen observation angles by explicitly modeling the influence of the observation angle and using an adaptive fusion mechanism to balance the original discriminative features and angle decoupling features.

[0047] S104. Input the micro-Doppler time-frequency image of the UAV to be identified into the target recognition model to obtain the category recognition result output by the target recognition model.

[0048] In this embodiment, a training sample set is obtained based on at least one sample time-frequency image. Based on this training sample set, which includes sample time-frequency images with added target category and observation angle labels, a preset neural network is trained to obtain a target recognition model for identifying drone models. Then, the micro-Doppler time-frequency image of the drone to be identified is input into the target recognition model, resulting in a category recognition result output by the target recognition model indicating the drone model's type. Thus, by combining the target's intrinsic features related to the target category with the angular influence features related to the observation angle, the target recognition model is trained to identify drone models. This expands the distinction between the target's intrinsic features and the angular influence features, avoiding misjudgments when the angle is not observed, thereby improving the accuracy of drone model category identification.

[0049] To verify the effectiveness of the UAV cross-angle recognition method provided in this application, Figure 4 This is a schematic diagram illustrating a comparison of recognition accuracy results provided in an embodiment of this application.

[0050] Combination Figure 4 As shown, Figure 4 The results show a comparison of the recognition accuracy of baseline networks such as ResNet18 and the target recognition model of the UAV cross-angle recognition method proposed in this application embodiment, namely the target recognition model using ADFNet, under the conditions of no observation angles at 60° and 120°. It can be seen that the UAV cross-angle recognition method has achieved higher recognition accuracy under both angle conditions.

[0051] Figure 5 A schematic diagram of a classification confusion matrix provided in an embodiment of this application, such as... Figure 5 The results show that, when the UAV cross-angle recognition method provided in this application is applied, the samples of various UAV models are mainly concentrated in the diagonal region, indicating that the UAV cross-angle recognition method provided in this application has good category discrimination ability and cross-angle recognition stability. From the above results, it can be seen that the present invention can effectively alleviate the problem of micro-Doppler feature distribution shift caused by changes in observation angle.

[0052] In some embodiments, this application also provides a drone cross-angle recognition device. This drone cross-angle recognition device may include one or more functional modules for implementing the drone cross-angle recognition method of the above method embodiments.

[0053] For example, Figure 6 This is a schematic diagram illustrating the composition of a drone cross-angle recognition device provided in an embodiment of this application. Figure 6 As shown, the UAV cross-angle recognition device includes: acquisition module 201, training module 202 and recognition module 203.

[0054] The acquisition module 201 is used to acquire at least one sample time-frequency map; the sample time-frequency map is a micro-Doppler time-frequency map of the corresponding UAV model during flight; Training module 202 is used to obtain a training sample set based on at least one sample time-frequency map. The training sample set includes sample time-frequency maps with added target category labels and observation angle labels. The target category label is used to indicate the UAV model corresponding to the sample time-frequency map, and the observation angle label is used to indicate the observation angle corresponding to the sample time-frequency map. The observation angle is the detection angle when the FMCW radar detects the UAV. The training module 202 is also used to train a preset neural network based on a training sample set to obtain a target recognition model for identifying drone models. The identification module 203 is used to input the micro-Doppler time-frequency image of the UAV to be identified into the target identification model to obtain the category identification result output by the target identification model; the category identification result is used to indicate the model of the UAV to be identified.

[0055] In some embodiments, the acquisition module 201 is specifically used to acquire UAV target echoes of different models of UAVs at different observation angles using FMCW radar; A two-dimensional Fourier transform is performed on the echoes of each UAV target to generate the corresponding sample time-frequency diagram.

[0056] In some embodiments, the preset neural network in the training module 202 includes a shared feature extraction network, a basic classifier, a weak angle decoupling unit, and an adaptive fusion unit; Among them, the shared feature extraction network is used to extract the overall texture, energy distribution and micro-Doppler modulation features of each sample's time-frequency map to obtain the corresponding shared features; The base classifier is used to identify drone categories by sharing features, and obtains effective discriminative information that retains the original features as the base classification features; The weak angle decoupling unit is used to extract UAV category information and observation angle information from the shared features to obtain decoupled classification features and fusion weights; The adaptive fusion unit is used to perform weighted fusion of basic classification features and decoupled classification features through fusion weights to obtain the corresponding classification and recognition results.

[0057] In some embodiments, the weak angle decoupling unit in the training module 202 includes a category feature mapping subunit, an angle feature mapping subunit, a decoupling classifier, and a fusion gating subunit; Among them, the category feature mapping subunit is used to extract discriminative information about the drone category from the shared features as target identity features; The angle feature mapping subunit is used to extract information about the time-frequency feature changes caused by the observation angle from the shared features as angle mapping features; The decoupling classifier is used to classify target identity features to obtain decoupling classification features; The fusion gating subunit generates fusion weights based on the target identity features and angle mapping features.

[0058] In some embodiments, the training module 202 is specifically used to train a preset neural network, including a first training phase, a second training phase, and a third training phase. The training phase consists of three stages: the first stage is used to train the classification of UAV models to establish stable category discrimination boundaries; the second stage incorporates angle supervision and orthogonal constraints to gradually distinguish information about UAV categories from information about observation angles; and the third stage incorporates weak angle adversarial constraints to reduce residual information about observation angles in target identity features.

[0059] In some embodiments, the total loss function for training a preset neural network in the training module 202 is as follows: ; in, Losses are categorized by drone type. To monitor losses from different angles, The first weighting coefficient corresponding to the angle supervision loss. The loss is an orthogonal constraint between identity features and angle features. This is the second weighting coefficient corresponding to the orthogonal constraint loss. To mitigate losses from weak angles, The third weighting coefficient corresponds to the loss against weak angles.

[0060] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device includes: a processor 702, a communication interface 703, and a bus 704. Optionally, the electronic device may also include a memory 701.

[0061] Processor 702 may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 702 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 702 may also be a combination of functions implementing computing capabilities, such as a combination including CPU0 and CPU1, a DSP, and a microprocessor.

[0062] The communication interface 703 includes a receiving unit and a transmitting unit, and is used to connect with other devices via a communication network. This communication network can be Ethernet, a wireless access network, a wireless local area network (WLAN), etc.

[0063] The memory 701 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), disk storage medium or other magnetic storage device, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0064] In one possible implementation, the memory 701 can exist independently of the processor 702. The memory 701 can be connected to the processor 702 via a bus 704 and is used to store instructions or program code. When the processor 702 calls the instructions or program code stored in the memory 701, it can implement the UAV cross-angle recognition method provided in this embodiment of the invention.

[0065] In another possible implementation, the memory 701 can also be integrated with the processor 702.

[0066] The 704 bus can be an extended industry standard architecture (EISA) bus, etc. The 704 bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0067] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the service calling device can be divided into different functional modules to complete all or part of the functions described above.

[0068] This application also provides a computer-readable storage medium. All or part of the processes in the above method embodiments can be executed by computer instructions instructing related hardware. The program can be stored in the aforementioned computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. The computer-readable storage medium can be any of the foregoing embodiments or memory. The aforementioned computer-readable storage medium can also be an external storage device of the aforementioned service invocation device, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the aforementioned service invocation device. Further, the aforementioned computer-readable storage medium can include both internal storage units of the aforementioned service invocation device and external storage devices. The aforementioned computer-readable storage medium is used to store the aforementioned computer program and other programs and data required by the aforementioned service invocation device. The aforementioned computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0069] This application also provides a computer program product, which includes a computer program that, when run on a computer, causes the computer to execute the UAV cross-angle recognition method provided in the above embodiments.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for cross-angle recognition of unmanned aerial vehicles (UAVs), characterized in that, The method includes: Obtain at least one sample time-frequency map; the sample time-frequency map is a micro-Doppler time-frequency map of the corresponding UAV model during flight; A training sample set is obtained based on the at least one sample time-frequency map. The training sample set includes sample time-frequency maps with added target category labels and observation angle labels. The target category label is used to indicate the UAV model corresponding to the sample time-frequency map, and the observation angle label is used to indicate the observation angle corresponding to the sample time-frequency map. The observation angle is the detection angle when the FMCW radar detects the UAV. Based on the training sample set, a preset neural network is trained to obtain a target recognition model for identifying drone models; The micro-Doppler time-frequency image of the drone to be identified is input into the target identification model to obtain the category identification result output by the target identification model; the category identification result is used to indicate the model of the drone to be identified.

2. The UAV cross-angle recognition method according to claim 1, characterized in that, The step of obtaining at least one sample time-frequency map includes: The FMCW radar was used to collect drone target echoes from different models of drones at different observation angles. A two-dimensional Fourier transform is performed on the echoes of each UAV target to generate the corresponding sample time-frequency diagram.

3. The UAV cross-angle recognition method according to claim 1, characterized in that, The preset neural network includes a shared feature extraction network, a basic classifier, a weak angle decoupling unit, and an adaptive fusion unit; The shared feature extraction network is used to extract the overall texture, energy distribution and micro-Doppler modulation features from the time-frequency maps of each sample to obtain the corresponding shared features. The basic classifier is used to identify the UAV category through the shared features, and obtains effective discriminative information that retains the original features as the basic classification features; The weak angle decoupling unit is used to extract UAV category information and observation angle information from the shared features to obtain decoupled classification features and fusion weights; The adaptive fusion unit is used to perform weighted fusion of the basic classification features and the decoupled classification features using the fusion weights to obtain the corresponding classification and recognition results.

4. The UAV cross-angle recognition method according to claim 3, characterized in that, The weak angle decoupling unit includes a category feature mapping subunit, an angle feature mapping subunit, a decoupling classifier, and a fusion gating subunit; The category feature mapping subunit is used to extract discrimination information about the drone category from the shared features as target identity features; The angle feature mapping subunit is used to extract information about the time-frequency feature changes caused by the observation angle as angle mapping features from the shared features; The decoupling classifier is used to classify the target identity features to obtain the decoupling classification features; The fusion gating subunit generates the fusion weights based on the target identity features and the angle mapping features.

5. The UAV cross-angle recognition method according to any one of claims 1 to 4, characterized in that, The training of the preset neural network includes a first training phase, a second training phase, and a third training phase. The first training phase is used to perform category classification training on UAV models to establish stable category discrimination boundaries. The second training phase incorporates angle supervision and orthogonal constraints to gradually distinguish information about UAV categories from information about observation angles. The third training phase incorporates weak angle adversarial constraints to reduce residual information about observation angles in target identity features.

6. The UAV cross-angle recognition method according to claim 5, characterized in that, The total loss function for training the preset neural network is shown in the following formula: ； in, Losses are categorized by drone type. To monitor losses from different angles, The first weighting coefficient corresponding to the angle supervision loss. The loss is an orthogonal constraint between identity features and angle features. This is the second weighting coefficient corresponding to the orthogonal constraint loss. To mitigate losses from weak angles, This is the third weighting coefficient corresponding to the loss against the weak angle.

7. A cross-angle recognition device for unmanned aerial vehicles (UAVs), characterized in that, include: The acquisition module is used to acquire at least one sample time-frequency map; the sample time-frequency map is a micro-Doppler time-frequency map of the corresponding UAV model during flight; The training module is used to obtain a training sample set based on the at least one sample time-frequency map. The training sample set includes sample time-frequency maps with added target category labels and observation angle labels. The target category label is used to indicate the UAV model corresponding to the sample time-frequency map, and the observation angle label is used to indicate the observation angle corresponding to the sample time-frequency map. The observation angle is the detection angle when the FMCW radar detects the UAV. The training module is also used to train a preset neural network based on the training sample set to obtain a target recognition model for identifying drone models. The identification module is used to input the micro-Doppler time-frequency image of the drone to be identified into the target identification model to obtain the category identification result output by the target identification model; the category identification result is used to indicate the model of the drone to be identified.

8. An electronic device, characterized in that, It includes a processor and a memory, the processor being coupled to the memory; the memory is used to store computer instructions, which are loaded and executed by the processor to enable the computer device to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes computer-executable instructions that, when executed on a computer, cause the computer to perform the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when run on an electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 6.

Images