Unmanned aerial vehicle recognition method and device based on small target detection and electronic device

Abstract

Embodiments of the present disclosure disclose a small target detection-based unmanned aerial vehicle identification method, device and electronic equipment. A specific implementation of the method comprises: collecting a target image sequence of a target flight restricted area; generating an unmanned aerial vehicle flight intention feature, an unmanned aerial vehicle heat source feature, an unmanned aerial vehicle shape feature and an unmanned aerial vehicle flight trajectory feature according to the target image sequence and a target detection model; collecting a first sound signal and a second sound signal; performing environmental noise stripping on the first sound signal and the second sound signal respectively to obtain an optimized first sound signal and an optimized second sound signal; generating an unmanned aerial vehicle voiceprint feature according to the optimized first sound signal and the optimized second sound signal; and generating unmanned aerial vehicle identity information according to the unmanned aerial vehicle flight intention feature, the unmanned aerial vehicle heat source feature, the unmanned aerial vehicle shape feature, the unmanned aerial vehicle flight trajectory feature and the unmanned aerial vehicle voiceprint feature. The implementation achieves effective identity recognition of unmanned aerial vehicles with a private control protocol.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the fields of computer technology and drone identification, specifically to drone identification methods, apparatus, and electronic devices based on small target detection. Background Technology

[0002] Currently, various technologies exist for identifying drones, such as using radio detection for drone identification and control. However, some drones employ proprietary control protocols, resulting in this approach being less effective for drone identification and control. Summary of the Invention

[0003] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0004] Some embodiments of this disclosure propose a method, apparatus, and electronic device for drone identification based on small target detection to address the technical problems mentioned in the background section above.

[0005] In a first aspect, some embodiments of this disclosure provide a drone identification method based on small target detection. The method includes: acquiring a sequence of target images for a target restricted airspace, wherein the target restricted airspace is a pre-defined airspace restricting drone flight, and the target images are fused images composed of multispectral images and visible light images; generating drone flight intention features, drone heat source features, drone morphological features, and drone flight trajectory features based on the target image sequence and a target detection model; acquiring a first sound signal and a second sound signal, wherein the first sound signal and the second sound signal are acquired along a first direction and a second direction, respectively, both the first direction and the second direction being towards the target restricted airspace, and the first direction being horizontal; performing environmental noise stripping on the first sound signal and the second sound signal respectively to obtain optimized first sound signal and optimized second sound signal; generating drone voiceprint features based on the optimized first sound signal and the optimized second sound signal; and generating drone identity information based on the drone flight intention features, the drone heat source features, the drone morphological features, the drone flight trajectory features, and the drone voiceprint features.

[0006] Secondly, some embodiments of this disclosure provide a drone identification device based on small target detection. The device includes: a first acquisition unit configured to acquire a sequence of target images for a target restricted airspace, wherein the target restricted airspace is a pre-defined airspace restricting drone flight, and the target image is a fused image composed of a multispectral image and a visible light image; a first generation unit configured to generate drone flight intention features, drone heat source features, drone morphological features, and drone flight trajectory features based on the target image sequence and a target detection model; and a second acquisition unit configured to acquire a first sound signal and a second sound signal, wherein the first sound signal and the second sound signal are... The system collects data along both the first and second directions, which are directed towards the restricted airspace of the target. The first direction is horizontal. An environmental noise stripping unit is configured to strip environmental noise from the first and second sound signals to obtain optimized first and second sound signals. A second generation unit is configured to generate UAV voiceprint features based on the optimized first and second sound signals. A third generation unit is configured to generate UAV identity information based on the UAV flight intention features, UAV heat source features, UAV morphological features, UAV flight trajectory features, and UAV voiceprint features.

[0007] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation of the first aspect above.

[0008] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0009] The above embodiments of this disclosure have the following beneficial effects: The UAV identification method based on small target detection in some embodiments of this disclosure achieves accurate identification of UAV identities. Specifically, the reason for inaccurate identification is that some UAVs use proprietary control protocols, resulting in poor identification and control performance of this solution. UAVs mainly possess two types of signals: control signals and image transmission signals. Because some UAVs use proprietary control protocols, the control signals cannot be effectively parsed, thus preventing effective identification of the UAV based on the detection of its corresponding radio signals. Therefore, the UAV identification method based on small target detection in some embodiments of this disclosure first acquires a target image sequence for a restricted flight airspace. The restricted flight airspace is a pre-defined airspace restricting UAV flight, and the target image is a fused image composed of multispectral and visible light images. For most drones, characterized by low-speed, low-altitude, and small size, this disclosure addresses this type of slow, small target by acquiring target images (fusion images composed of multispectral and visible light images) of the target in restricted airspace, thereby capturing drone features from multiple angles as much as possible from the image perspective. Secondly, based on the aforementioned target image sequence and target detection model, drone flight intention features, drone heat source features, drone morphological features, and drone flight trajectory features are generated. This is achieved through machine learning, combining only images to create a multi-angle drone profile depicting the drone's flight intention, heat source, morphology, and flight trajectory. Next, a first sound signal and a second sound signal are acquired, respectively, along a first direction and a second direction, both pointing towards the restricted airspace of the target. The first direction is horizontal. In practice, to further enrich the drone profile, considering that drones, especially common multi-rotor drones, exhibit rotational noise (mainly composed of thickness noise and load noise) during flight, and that different drones often have different rotational noise characteristics, this disclosure further acquires sound signals. In particular, this disclosure considers that the rotational noise of a UAV is mainly generated by the operation of its propellers, and the sound pressure level of the radial noise generated by the propellers is often greater than that of the axial noise. Conventional solutions also include UAV feature extraction based on sound signals; however, such solutions primarily acquire sound signals (axial noise) through a ground-based directional microphone array. Therefore, this disclosure acquires sound signals from both a first and a second direction to capture the rotational noise generated by the UAV as much as possible. Furthermore, environmental noise stripping is performed on the first and second sound signals respectively to obtain optimized first and second sound signals.In practice, ambient noise other than rotational noise is inevitably captured during sound signal acquisition. Therefore, this disclosure reduces the masking of rotational noise by ambient noise through ambient noise stripping. Next, based on the optimized first and second sound signals, UAV voiceprint features are generated. Finally, based on the UAV flight intention features, UAV heat source features, UAV morphological features, UAV flight trajectory features, and UAV voiceprint features, UAV identity information is generated. This method eliminates the need for detection and analysis of UAV control signals, achieving effective identification of UAVs with proprietary control protocols. Attached Figure Description

[0010] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0011] Figure 1 This is a flowchart of some embodiments of the UAV identification method based on small target detection according to the present disclosure;

[0012] Figure 2 This is a schematic diagram showing the positional relationship between the clear airspace and the ring-shaped restricted airspace;

[0013] Figure 3 This is a schematic diagram of the network structure of an unmanned aerial vehicle (UAV) positioning network;

[0014] Figure 4 This is a schematic diagram of the structure of some embodiments of drone recognition based on small target detection according to the present disclosure;

[0015] Figure 5 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0016] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0017] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0018] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0019] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0020] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0021] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] refer to Figure 1 The diagram illustrates a flow 100 of some embodiments of a drone identification method based on small target detection according to the present disclosure. This drone identification method based on small target detection includes the following steps:

[0023] Step 101: Collect target image sequences for the restricted airspace.

[0024] In some embodiments, the implementer of the UAV identification method based on small target detection (e.g., a computing device) can acquire a sequence of target image images for a target restricted airspace.

[0025] The target restricted airspace is a pre-designated airspace where drone flights are restricted. Specifically, the target restricted airspace can be the airspace corresponding to an area where access to privacy information within the area is restricted. For example, the target restricted airspace can be the unmanned aerial vehicle (UAV) controlled airspace stipulated in the "Interim Regulations on the Management of Unmanned Aerial Vehicle Flights." Alternatively, the target restricted airspace can also be a temporarily designated airspace based on actual restriction needs. Furthermore, due to the requirements of the aforementioned regulations, the flight altitude of light, small, and medium-sized UAVs is limited (not exceeding 300 meters above ground level). Therefore, the upper limit of the target restricted airspace altitude can be less than or equal to 300 meters above ground level. Moreover, considering that some modified UAVs may have higher flight altitudes, the upper limit of the target restricted airspace altitude can be adjusted adaptively.

[0026] Among them, light, small, and medium-sized UAVs are characterized by their small size, and therefore are typical small targets in the field of target detection.

[0027] The target image is a fused image composed of multispectral and visible light images. In particular, for restricted airspace, multiple consecutive target images can be acquired over time to form a target image sequence.

[0028] In practice, due to the small size of drones, simply acquiring visible light images for subsequent target detection may result in some targets not being detected. Therefore, this disclosure firstly utilizes both multispectral and visible light cameras to acquire multispectral and visible light images. Then, the multispectral and visible light images acquired under different spectra are overlaid and fused as different layers to form the target image.

[0029] It should be noted that the aforementioned computing devices can be either hardware or software. When the computing device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or as a single server or a single terminal device. When the computing device is software, it can be installed on the hardware devices listed above. It can be implemented as, for example, multiple software programs or software modules used to provide distributed services, or as a single software program or software module. No specific limitations are made here.

[0030] In some optional implementations of certain embodiments, prior to acquiring the target image sequence for the target restricted airspace, the method further includes:

[0031] Step S1: Construct a ring-shaped restricted airspace with the clear airspace as the center and the preset radius difference as the ring width.

[0032] The aforementioned airspace is the restricted airspace for drones. The preset radius difference is the difference between the inner and outer radii of the corresponding ring-shaped restricted airspace.

[0033] In practice, to ensure the effective identification and execution of corresponding operations (such as driving away) of drones, a restricted area is often set outside the airspace where drones are restricted from entering, serving as a window area for drone identification and operation. In particular, to guarantee the success rate of identification and operation, a specific value for a preset radius difference can be set according to actual needs. For example, the preset radius difference could be 2 km.

[0034] As an example, see Figure 2 The diagram shows the positional relationship between the clear airspace and the ring-shaped restricted airspace. In this diagram, the clear airspace 201 can be the inner circle of the ring-shaped restricted airspace 202, and the preset radius difference can be the radius difference between the corresponding outer circles of the clear airspace 201 and the ring-shaped restricted airspace 202.

[0035] Step S2: Using the preset arc length as the inner arc length, divide the above-mentioned ring-shaped restricted airspace into a set of fan-ring airspaces.

[0036] Each sector ring spatial domain in the aforementioned sector ring spatial domain set is assigned a scan identifier. The scan identifier indicates whether the corresponding sector ring spatial domain has been scanned in the current scan round. After the current scan round is completed, the scan identifiers corresponding to the sector ring spatial domains in the sector ring spatial domain set will be initialized according to the scan mode.

[0037] The scanning mode can include a preset scanning mode and a custom scanning mode. The preset scanning mode can indicate that each sector in the sector ring spatial domain set is scanned sequentially in a clockwise or counterclockwise direction. The custom scanning mode can indicate that the sector ring spatial domains in the sector ring spatial domain set are scanned according to preset scanning rules.

[0038] As an example, in the preset scanning mode, the scan flag corresponding to each sector ring airspace is initialized to 0, indicating that it has not been scanned. Once a sector ring airspace is scanned, the corresponding scan flag can be updated to 1, indicating that it has been scanned. A counter can also be set to count the number of sector ring airspaces scanned. Since the sector ring airspace is obtained by dividing the ring-shaped restricted airspace with a preset arc length as the inner arc length, the number of sector ring airspaces in the sector ring airspace set is fixed. When the counter value equals the number of sector ring airspaces in the sector ring airspace set, it indicates that the current scanning round has been completed. At this time, the scan flags corresponding to each sector ring airspace in the sector ring airspace set can be initialized to 0, and a new scanning round can begin. For example, see further details. Figure 2 For the ring-shaped restricted airspace 202, it can be divided into 8 sector-ring airspaces 203. Therefore, in the preset scanning mode, the 8 sector-ring airspaces 203 can be scanned one by one.

[0039] As another example, in custom scanning mode, since the upper limit of the drone's flight altitude is mainly around 300 meters above ground level, and considering that there may be obstacles blocking the fan-ring airspace, some fan-ring airspace may not need to be scanned, or only low-frequency scanning is required. In this case, the scanning flag corresponding to this type of fan-ring airspace can be initialized to 1. Therefore, when scanning the fan-ring area, this type of fan-ring airspace can be skipped, or selective fan-ring area scanning can be performed in different scanning rounds.

[0040] As another example, in custom scanning mode, since a single scanning cycle often requires a certain scanning time (e.g., scanning 360 degrees clockwise), and drones have movement characteristics, in order to improve scanning speed and reduce the probability of missed scans, the scanning speed can be increased at the expense of some scanning coverage. Specifically, in odd-numbered scanning cycles, the fan-ring airspace with odd numbers can be scanned, and in even-numbered scanning cycles, the fan-ring airspace with even numbers can be scanned. Therefore, the above purpose can be achieved by adjusting the scanning markers of the fan-ring airspace.

[0041] In particular, in the custom scan mode, not the entire sector ring airspace is scanned; therefore, the control threshold corresponding to the scan cycle can be adjusted adaptively.

[0042] Step S3: Determine the first fan-ring airspace along the scanning direction in the above-mentioned fan-ring airspace set under the current scan round as the target restricted flight area.

[0043] In practice, the first sector airspace with a corresponding scan identifier of 0 along the scan direction in the current scan round can be identified as the target restricted flight area.

[0044] Step 102: Based on the target image sequence and the target detection model, generate UAV flight intention features, UAV heat source features, UAV morphological features and UAV flight trajectory features.

[0045] In some embodiments, the aforementioned execution entity can generate UAV flight intention features, UAV heat source features, UAV morphological features, and UAV flight trajectory features based on the target image sequence and the target detection model.

[0046] The target detection model can be a machine learning model used to identify drones and extract corresponding features. Drone flight intention features represent the identified drone's flight intention. Drone heat source features represent the heat distribution of the identified drone. Drone morphological features represent the identified drone's morphological structure. Drone flight trajectory features represent the identified drone's flight trajectory.

[0047] In practice, firstly, due to the small size of drones, static identification alone has a high probability of missed detection. Therefore, combining (target) image sequences for drone identification reduces this probability. Secondly, the small size of drones may lead to misidentification of birds or other objects as drones. Therefore, multi-faceted target characterization is needed, considering flight intent, heat distribution, morphological structure, and flight trajectory. Specifically, while conventional target detection models such as YOLO (You Only Look Once) can be used, considering the need to further extract drone flight intent features, heat source features, morphological features, and flight trajectory features, four feature mappers (e.g., implemented using fully connected layers) can be set up in parallel on top of the YOLO model to extract these features.

[0048] Optionally, the target detection model includes: a UAV localization network, a UAV heat source feature extraction network, a UAV morphological feature extraction network, a UAV flight trajectory feature extraction network, and a UAV flight intention classifier.

[0049] The UAV positioning network consists of a first module, a second module, and a third module. For example, see... Figure 3 The diagram shows the network structure of the drone positioning network:

[0050] The first module is mainly used to perform multi-scale feature extraction on the input target image to output feature maps at four scales (feature map A1, feature map A2, feature map A3, and feature map A4, where the feature size of feature map A1 < the feature size of feature map A2 < the feature size of feature map A3 < the feature size of feature map A4). Specifically, the first module uses the FasterNet model as the backbone network.

[0051] The second module has four branches (branch 1, branch 2, branch 3, and branch 4). Each branch outputs processed feature maps at different scales.

[0052] Branch 1 takes feature map A1 as input and consists of a convolutional layer, a max-pooling layer, a max-pooling layer, a max-pooling layer, a concatenation layer, a convolutional layer, a first variable convolutional module, and an attention module. The input to the concatenation layer is the sum of the outputs of the first convolutional layer of branch 1 and the outputs of the three max-pooling layers. The input to the first variable convolutional module of branch 1 is the sum of the outputs of the second convolutional layer of branch 1 and the outputs of the second variable convolutional module of branch 2.

[0053] In this branch, the features from feature map D and the output of the first attention module of branch 3 are upsampled by an upsampling layer and then superimposed as the input to branch 4. Branch 4 consists of a first variable convolution module, an attention module, and a second variable convolution module. Specifically, after the features are processed by the attention module of branch 4, one path is input to the third module, and the other path is input to the second variable convolution module included in branch 4.

[0054] In this branch, the features from feature map C and the output of the first attention module of branch 2 are upsampled by an upsampling layer and then superimposed as the input to branch 3. Branch 3 consists of a first variable convolution module, an attention module, a second variable convolution module, and another attention module. Specifically, after the features are processed by the second attention module of branch 3, one path is input to the third module, and the other path is input to the second variable convolution module included in branch 3. The input of the second first variable convolution module of branch 3 is the superposition of the output of the first attention module of branch 3 and the output of the second variable convolution module of branch 4.

[0055] In this branch, the features from feature map B and the output of the second convolutional layer of branch 1 are upsampled by an upsampling layer and then superimposed as the input to branch 2. Branch 2 consists of a first variable convolutional module, an attention module, a second variable convolutional module, a first variable convolutional module, and an attention module. Specifically, after the features are processed by the second attention module of branch 2, one path is input to the third module, and the other path is input to the second variable convolutional module included in branch 2. The input of the second first variable convolutional module of branch 1 is the superposition of the output of the first attention module of branch 1 and the output of the second variable convolutional module of branch 3.

[0056] The third module includes four parallel decoupling locators, corresponding to branches 1, 2, 3, and 4, respectively. Specifically, the input to the first decoupling locator is the output of the attention module of branch 1. The input to the second decoupling locator is the output of the second attention module of branch 2. The input to the third decoupling locator is the output of the second attention module of branch 3. The input to the third decoupling locator is the output of the first attention module of branch 4.

[0057] The third module includes a decoupling locator that employs a dual-branch structure (a classification branch and a bounding box branch). The classification branch consists of convolutional layers, convolutional layers, and 2D convolutional layers. The bounding box branch consists of convolutional layers, convolutional layers, and 2D convolutional layers.

[0058] The convolutional layers in both the second and third modules consist of two-dimensional convolutional layers (Conv2D), two-dimensional batch normalization layers, and SiLU activation functions. The max-pooling layer in the second module is a two-dimensional max-pooling layer (MaxPool2D). The first variable convolutional module in the second module uses a C2F_DCN variable convolutional module. The attention module in the second module uses a coordinate attention mechanism.

[0059] In practice, firstly, the first module uses the FasterNet model as the backbone network, which effectively reduces the number of model parameters. Secondly, by setting four decoupled locators, target detection at different scales is achieved. In particular, the setting of the decoupled locator corresponding to branch 1 can effectively detect small targets. Through model structure optimization, the above target detection model can be deployed on low-computing-power platforms.

[0060] The UAV heat source feature extraction network consists of an upsampling network and a heatmap mapping layer. The upsampling network comprises three upsampling layers and a feature overlay layer. First, the three upsampling layers correspond to branches 2, 3, and 4 of the second module, respectively, to perform parallel upsampling of the features (local feature maps) output by branches 2, 3, and 4. The feature size of the upsampled features is kept consistent with the feature size of the output of branch 1 of the first module. Next, the feature overlay layer overlays the output of branch 1 and the three upsampled features as the output of the upsampling network. The heatmap mapping layer maps the output of the upsampling network to obtain a heatmap for a single target image. By processing each target image in the target image sequence, a heatmap sequence can be obtained, serving as the UAV heat source feature.

[0061] The UAV morphology feature extraction network consists of three convolutional layers. Since the first decoupling localizer in the third module is used for feature target detection at the smallest scale, the input to the UAV morphology feature extraction network is the local features defined by the rectangular bounding box of the first decoupling localizer in the third module. The three convolutional layers in the UAV morphology feature extraction network perform linear convolution processing on the local features defined by the rectangular bounding box of the first decoupling localizer in the third module to obtain the UAV morphology features.

[0062] The UAV flight trajectory feature extraction network generates UAV flight trajectory features through linear interpolation. Specifically, firstly, the decoupled locator in the third module of the UAV localization network outputs a bounding box defining the UAV. By sequentially processing the target images in the target image sequence, a time-varying sequence of UAV position coordinates is obtained. Then, the UAV flight trajectory feature extraction network interpolates the position coordinate sequence using linear interpolation to obtain the UAV flight trajectory features.

[0063] The UAV flight intent classifier comprises a position vector encoder, a position feature extractor, and an intent classifier. First, the position coordinates of the UAV flight trajectory features are converted into three-dimensional position coordinates in a geodetic coordinate system. Then, the position vector encoder converts two adjacent three-dimensional position coordinates into three position vectors (corresponding to the X-axis, Y-axis, and Z-axis directions), thus obtaining a position vector matrix for the UAV flight trajectory features. Next, the position feature extractor further extracts features from the position vector matrix. Finally, the extracted features are input into the intent classifier to obtain the UAV flight intent features. The position feature extractor uses a Seq2Seq model. The intent classifier employs a multi-classifier.

[0064] In some optional implementations of certain embodiments, the execution entity generates UAV flight intention features, UAV heat source features, UAV morphological features, and UAV flight trajectory features based on the target image sequence and target detection model, including:

[0065] Step S1: For each target image in the above target image sequence, perform drone detection on the target image through the above target drone positioning network to generate drone description information and obtain a drone description information set.

[0066] The UAV description information includes: UAV position coordinates and local feature maps. The size of the local feature maps is consistent with the size of the region of interest corresponding to the UAV position coordinates. The UAV position coordinates are in the image coordinate system.

[0067] Step S2: Generate the drone flight trajectory features based on the drone location coordinates included in the drone description information set and the drone flight trajectory feature extraction network.

[0068] Step S3: Generate the above-mentioned UAV heat source features based on the local feature maps included in the UAV description information set and the above-mentioned UAV heat source feature extraction network.

[0069] Step S4: Generate the above-mentioned UAV morphological features based on the local feature maps included in the UAV description information set and the above-mentioned UAV morphological feature extraction network;

[0070] Step S5: Generate the drone flight intention features based on the drone flight trajectory features and the drone flight intention classifier.

[0071] Step 103: Acquire the first sound signal and the second sound signal.

[0072] In some embodiments, the aforementioned executing entity may collect a first sound signal and a second sound signal.

[0073] The first sound signal and the second sound signal are collected along the first direction and the second direction, respectively. Both the first direction and the second direction are directed toward the restricted airspace of the target. The first direction is horizontal.

[0074] Optionally, the first sound signal is acquired by an airborne sound acquisition device, and the second sound signal is acquired by a ground-based sound acquisition device. The airborne sound acquisition device can be a trusted unmanned aerial vehicle (UAV) containing a directional microphone array. The ground-based sound acquisition device is a directional microphone array mounted on the ground.

[0075] In practice, to ensure the alignment of the first and second sound signals, the aforementioned executing entity can simultaneously control both the airborne and ground-based sound acquisition devices to acquire the first and second sound signals. Specifically, given the known trajectory characteristics of the UAV, the airborne sound acquisition device can be controlled to fly at the same altitude as the three-dimensional position coordinate point corresponding to the latest UAV position coordinates included in the UAV trajectory characteristics, and oriented towards that point to acquire the first sound signal. Similarly, the ground-based sound acquisition device can be controlled to oriented towards the three-dimensional position coordinate point corresponding to the latest UAV position coordinates included in the UAV trajectory characteristics to acquire the second sound signal.

[0076] Step 104: Perform environmental noise stripping on the first sound signal and the second sound signal respectively to obtain the optimized first sound signal and the optimized second sound signal.

[0077] In some embodiments, the aforementioned execution entity may perform environmental noise stripping on the first sound signal and the second sound signal respectively to obtain an optimized first sound signal and an optimized second sound signal.

[0078] The optimized first sound signal is the first sound signal after removing environmental noise. The optimized second sound signal is the second sound signal after removing environmental noise.

[0079] In practice, although both the first and second sound signals are acquired using directional microphone arrays, which reduces environmental noise interference to some extent, the presence of environmental noise in the acquisition direction still results in the inclusion of environmental noise in both the first and second sound signals. Since it is necessary to extract the rotational noise generated by the drone, it is essential to suppress environmental noise interference as much as possible. Specifically, Kalman filtering can be used to filter the noise from both the first and second sound signals separately, resulting in optimized first and second sound signals.

[0080] In some optional implementations of certain embodiments, the execution entity performs environmental noise stripping on the first sound signal and the second sound signal respectively to obtain optimized first sound signal and optimized second sound signal, including...

[0081] Step S1: Determine the signal difference between the first sound signal and the basic environmental noise to obtain the first noise-reduced sound signal.

[0082] The aforementioned basic environmental noise is the average environmental noise collected within the restricted airspace of the aforementioned target. The noise impact degree of the aforementioned basic environmental noise is controlled by the influence coefficient.

[0083] In practice, taking the first sound signal as an example, it can be expressed by the following formula: S(t) = s1(t) + s2(t). Here, s1(t) represents the environmental noise, and s2(t) represents the rotational noise generated by the UAV. Assuming that the environmental noise within the target restricted airspace changes slightly over a long period, the basic environmental noise s3(t) within the target restricted airspace can be collected in advance. This can be understood as the environmental noise within the target restricted airspace, varying based on the basic environmental noise. Furthermore, to control the influence of the basic environmental noise, an influence coefficient β can be set, with a value range of [0,1]. Therefore, the process of determining the signal difference between the first sound signal and the basic environmental noise to obtain the first noise-reduced sound signal can be expressed by the following formula: S(t) - βs3(t) = [s1(t) - βs3(t)] + s2(t). This method can quickly suppress noise signals to a certain extent.

[0084] Step S2: Encode the first noise-reduced audio signal using an encoder to obtain the first encoded signal features and the second encoded signal features.

[0085] The encoder described above consists of one one-dimensional convolutional module, one normalization layer, and a target convolutional layer. The first encoded signal feature is the output feature of the one-dimensional convolutional module included in the encoder. The second encoded signal feature is the output feature of the target convolutional layer included in the encoder. The one-dimensional convolutional module consists of a target convolutional layer, a PReLU activation function, a normalization layer, a deconvolutional layer, a PReLU activation function, a normalization layer, and a target convolutional layer. The kernel size of the target convolutional layer is 1×1.

[0086] Step S3: Perform feature processing on the second encoded signal features using an optimizer to obtain the third encoded signal features.

[0087] The optimizer consists of three sub-optimizers (optimizer A1, optimizer A2, and optimizer A3), connected serially. Each of the three optimizers includes five one-dimensional convolutional modules. Skip connections are established between any two adjacent odd-numbered one-dimensional convolutional modules (in practice, skip connections are used to prevent feature forgetting). For example, optimizer A1 includes one-dimensional convolutional modules C1, C2, C3, C4, and C5. The output of one-dimensional convolutional module C1 is the input of one-dimensional convolutional module C2. The output of one-dimensional convolutional module C2 is the input of one-dimensional convolutional module C3. The output of one-dimensional convolutional module C3 is the input of one-dimensional convolutional module C4. The output of one-dimensional convolutional module C4 is the input of one-dimensional convolutional module C5. The output of one-dimensional convolutional module C5 is the output of optimizer A1. Specifically, the output of one-dimensional convolutional module C1 serves as the input of one-dimensional convolutional module C3 via a skip connection between C1 and C3. The output of one-dimensional convolutional module C3 serves as the input of one-dimensional convolutional module C5 via a skip connection between C3 and C5. The same applies to optimizers A2 and A3. Furthermore, the output of optimizer A1 serves as the input of optimizer A2, and the output of optimizer A2 serves as the input of optimizer A3.

[0088] Step S4: Generate the optimized first sound signal using the decoder, the first encoded signal features, and the third encoded signal features.

[0089] The decoder comprises a first activation function, a target convolutional layer, a second activation function, and a one-dimensional convolutional module. The optimized first audio signal is the output signal of the one-dimensional convolutional module in the decoder. The third encoded signal features are processed sequentially by the first activation function, the target convolutional layer, and the second activation function in the decoder, and then subjected to a tensor product operation with the first encoded signal features, serving as the input features of the one-dimensional convolutional module in the decoder. The first activation function is the PreLU activation function.

[0090] In practice, three linearly configured sub-optimizers are used to achieve step-by-step signal feature extraction. To avoid the feature forgetting problem inherent in linear network structures, skip connections are established between the local network structures of the sub-optimizers. Furthermore, the feature forgetting problem is further overcome by superimposing the encoder's output features (first encoded signal features) with the optimizer's output features (third encoded signal features).

[0091] In some optional implementations of certain embodiments, environmental noise stripping is performed on the first sound signal and the second sound signal respectively to obtain optimized first sound signal and optimized second sound signal, further comprising:

[0092] Step S1: Determine the signal difference between the second sound signal and the basic environmental noise to obtain the second noise-reduced sound signal.

[0093] Step S2: Encode the second noise-reduced audio signal using the encoder to obtain the fourth and fifth encoded signal features.

[0094] The fourth encoded signal feature is the output feature of the one-dimensional convolutional module included in the encoder, which is the output feature of the first encoded signal feature. The fifth encoded signal feature is the output feature of the target convolutional layer included in the encoder.

[0095] Step S3: Perform feature processing on the fifth encoded signal features using an optimizer to obtain the sixth encoded signal features.

[0096] Step S4: Generate the optimized second sound signal using the decoder, the fourth encoded signal feature, and the sixth encoded signal feature.

[0097] The sixth encoded signal feature is processed sequentially by the first activation function, the target convolutional layer, and the second activation function included in the decoder. Then, it is subjected to a tensor product operation with the fourth encoded signal feature and used as the input feature of the one-dimensional convolutional module included in the decoder.

[0098] Step 105: Generate UAV voiceprint features based on the optimized first sound signal and the optimized second sound signal.

[0099] In some embodiments, the aforementioned executing entity can generate drone voiceprint features based on the optimized first sound signal and the optimized second sound signal.

[0100] Among them, the drone voiceprint feature represents the voiceprint feature corresponding to the rotational noise generated by the drone.

[0101] As an example, the optimized first sound signal and the optimized second sound signal can be downsampled to a preset length to serve as the voiceprint features of the drone.

[0102] As another example, the spectral correlation features (such as Mel-spectrum cepstral coefficients, linear prediction cepstral coefficients, and linear prediction coefficients) corresponding to the optimized first sound signal and the optimized second sound signal can be extracted separately as the drone's voiceprint features.

[0103] As another example, due to the fixed nature of the ground-based sound acquisition equipment and the influence of the positional relationship between the ground-based sound acquisition equipment and the target UAV on the second sound signal, the acquisition of the second sound signal is variable. In order to ensure the uniqueness of the obtained UAV voiceprint features, further analysis is required depending on the situation: (1) When the second direction is perpendicular to the ground, that is, the target UAV is located directly above the ground-based sound acquisition equipment, the above-mentioned sound signal downsampling or spectrum correlation feature extraction method can be used to process the optimized first sound signal and the optimized second sound signal respectively to obtain the UAV voiceprint features. (2) When the second direction is not perpendicular to the ground, firstly, the optimized second sound component needs to be decomposed into a sound signal component perpendicular to the ground through vector operation. Then, the above-mentioned sound signal downsampling or spectrum correlation feature extraction method is used to process the optimized first sound signal and the sound signal component respectively to obtain the UAV voiceprint features.

[0104] Step 106: Generate drone identity information based on drone flight intention characteristics, drone heat source characteristics, drone morphological characteristics, drone flight trajectory characteristics, and drone voiceprint characteristics.

[0105] In some embodiments, the aforementioned executing entity may generate drone identity information based on drone flight intention characteristics, drone heat source characteristics, drone morphological characteristics, drone flight trajectory characteristics, and drone voiceprint characteristics.

[0106] The drone identity information represents the physical attributes of the identified drone (target drone). This includes parameters such as the number of rotors, aircraft type, model number, flight intention type, and flight altitude. Aircraft type can include: light aircraft, small aircraft, and medium aircraft.

[0107] In practice, firstly, the flight intent type can be directly determined by the UAV's flight intent features; specifically, malicious flight intent can be statistically analyzed based on these features. Secondly, flight altitude can be obtained from the UAV's flight trajectory features. Since these features include the UAV's position coordinates, the flight altitude parameter can be determined by mapping from the image coordinate system to the geodetic coordinate system. Furthermore, to ensure the accuracy of the determined flight altitude, equipment such as a rangefinder can be used to orient and measure the distance to the target UAV and convert it into a vertical component as the flight altitude. Next, regarding the number of rotors and aircraft type, UAV thermal characteristics, UAV morphological characteristics, and UAV acoustic signature features can be combined. Since these three features have different feature scales and reside in different feature spaces, feature mapping can be used to project the UAV thermal characteristics, UAV morphological characteristics, and UAV acoustic signature features into the same feature space. Then, a fusion vector is constructed through feature concatenation and downsampling, and a multi-classifier is used to classify the target UAV's corresponding aircraft type. Furthermore, based on the constructed fusion vector, it can be matched with the model feature vector of the drone in the pre-constructed drone information database through similarity calculation to determine the corresponding drone model.

[0108] In some optional implementations of some embodiments, the above method further includes:

[0109] Step S1: Send an authentication request to the target drone.

[0110] The target drone is the drone corresponding to the aforementioned drone identity information. The aforementioned identity verification request includes: verification terminal identity information, session identifier, and request level. The session identifier represents the unique identifier of the identity verification request, the request level represents the verification level corresponding to the identity verification request, and the verification terminal identity information represents the identity of the terminal initiating the identity verification request. The aforementioned request levels include: a first request level, a second request level, and a third request level. The first request level indicates that only the identity information and flight access permissions corresponding to the target drone are obtained. The second request level indicates that, based on the first level request, the flight parameters of the target drone (e.g., current location, flight mode, flight altitude, remaining battery power) are further obtained. The third request level indicates that, based on the second request level, the identity information of the drone operator bound to the target drone is further obtained.

[0111] In practice, given the type of drone, the aforementioned execution entity can dynamically adjust to the corresponding communication protocol and communicate with the target drone under that protocol. In particular, since the target drone's communication address is unknown at the time of communication establishment, an authentication request can be sent to the target drone via broadcast.

[0112] Step S2: In response to the target drone failing to respond to the authentication request after a preset time period, the target drone is driven away.

[0113] In practice, for example, directional jamming devices can be used to trigger the target drone's automatic return-to-home mechanism, thereby driving the drone away. Alternatively, simulating and sending false control signals can also cause the target drone to land at its apex.

[0114] Step S3: Upon receiving the request response information returned by the target drone, perform drone authentication and flight access permission verification on the target drone based on the request response information.

[0115] The request-response information includes: an identity token and a list of granted permissions. The identity token represents the target drone's identity. The list of granted permissions represents the flight permissions already granted to the target drone. Specifically, the target drone's identity token and verification terminal identity information can be issued through the same trusted terminal.

[0116] In practice, drone authentication verifies the legitimacy and validity of identity tokens. Flight access verification checks the target drone's list of existing permissions to determine if the target drone has flight permissions for the target restricted airspace.

[0117] Step S4: In response to the aforementioned target drone failing drone authentication or flight access permission verification, drive away the aforementioned target drone.

[0118] In practice, the method of driving away drones can be found in step S2, and will not be repeated here. In addition, drone authentication can precede flight access verification. When the target drone fails drone authentication, flight access verification will not be performed, and the target drone will be driven away directly.

[0119] Step S5: In response to the above target drone passing drone authentication and flight access permission verification and the above request level being the third request level, perform operation object authentication on the drone operation object corresponding to the above target drone based on the above request response information.

[0120] In practice, when the request level is level three, the request response message will also include object identity information corresponding to the drone being operated. Therefore, it is possible to search the object identity information database to see if the object identity information included in the request response information exists, and whether the object identity information included in the request response information matches the target drone.

[0121] Step S6: In response to the failure to authenticate the target drone, drive away the aforementioned drone.

[0122] In practice, the expulsion method can be found in step S2, and will not be repeated here.

[0123] The above embodiments of this disclosure have the following beneficial effects: The UAV identification method based on small target detection in some embodiments of this disclosure achieves accurate identification of UAV identities. Specifically, the reason for inaccurate identification is that some UAVs use proprietary control protocols, resulting in poor identification and control performance of this solution. UAVs mainly possess two types of signals: control signals and image transmission signals. Because some UAVs use proprietary control protocols, the control signals cannot be effectively parsed, thus preventing effective identification of the UAV based on the detection of its corresponding radio signals. Therefore, the UAV identification method based on small target detection in some embodiments of this disclosure first acquires a target image sequence for a restricted flight airspace. The restricted flight airspace is a pre-defined airspace restricting UAV flight, and the target image is a fused image composed of multispectral and visible light images. For most drones, characterized by low-speed, low-altitude, and small size, this disclosure addresses this type of slow, small target by acquiring target images (fusion images composed of multispectral and visible light images) of the target in restricted airspace, thereby capturing drone features from multiple angles as much as possible from the image perspective. Secondly, based on the aforementioned target image sequence and target detection model, drone flight intention features, drone heat source features, drone morphological features, and drone flight trajectory features are generated. This is achieved through machine learning, combining only images to create a multi-angle drone profile depicting the drone's flight intention, heat source, morphology, and flight trajectory. Next, a first sound signal and a second sound signal are acquired, respectively, along a first direction and a second direction, both pointing towards the restricted airspace of the target. The first direction is horizontal. In practice, to further enrich the drone profile, considering that drones, especially common multi-rotor drones, exhibit rotational noise (mainly composed of thickness noise and load noise) during flight, and that different drones often have different rotational noise characteristics, this disclosure further acquires sound signals. In particular, this disclosure considers that the rotational noise of a UAV is mainly generated by the operation of its propellers, and the sound pressure level of the radial noise generated by the propellers is often greater than that of the axial noise. Conventional solutions also include UAV feature extraction based on sound signals; however, such solutions primarily acquire sound signals (axial noise) through a ground-based directional microphone array. Therefore, this disclosure acquires sound signals from both a first and a second direction to capture the rotational noise generated by the UAV as much as possible. Furthermore, environmental noise stripping is performed on the first and second sound signals respectively to obtain optimized first and second sound signals.In practice, ambient noise other than rotational noise is inevitably captured during sound signal acquisition. Therefore, this disclosure reduces the masking of rotational noise by ambient noise through ambient noise stripping. Next, based on the optimized first and second sound signals, UAV voiceprint features are generated. Finally, based on the UAV flight intention features, UAV heat source features, UAV morphological features, UAV flight trajectory features, and UAV voiceprint features, UAV identity information is generated. This method eliminates the need for detection and analysis of UAV control signals, achieving effective identification of UAVs with proprietary control protocols.

[0124] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a drone identification device based on small target detection. These device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, this drone identification device based on small target detection can be specifically applied to various electronic devices.

[0125] like Figure 4 As shown, a drone identification device 400 based on small target detection in some embodiments includes: a first acquisition unit 401, a first generation unit 402, a second acquisition unit 403, an environmental noise stripping unit 404, a second generation unit 405, and a third generation unit 406. The first acquisition unit 401 is configured to acquire a sequence of target images for a restricted flight airspace, wherein the restricted flight airspace is a pre-defined airspace restricting drone flight, and the target image is a fused image composed of a multispectral image and a visible light image; the first generation unit 402 is configured to generate drone flight intention features, drone heat source features, drone morphological features, and drone flight trajectory features based on the target image sequence and a target detection model; the second acquisition unit 403 is configured to acquire a first sound signal and a second sound signal, wherein the first sound signal and the second sound signal are acquired along a first direction and a second direction, respectively. Both the first direction and the second direction are directed toward the restricted airspace of the target, and the first direction is horizontal. The environmental noise stripping unit 404 is configured to strip environmental noise from the first sound signal and the second sound signal respectively to obtain an optimized first sound signal and an optimized second sound signal. The second generation unit 405 is configured to generate UAV voiceprint features based on the optimized first sound signal and the optimized second sound signal. The third generation unit 406 is configured to generate UAV identity information based on the UAV flight intention features, the UAV heat source features, the UAV morphological features, the UAV flight trajectory features, and the UAV voiceprint features.

[0126] It is understandable that the units described in the small target detection-based UAV identification device 400 are similar to the reference units. Figure 1 The steps in the described method correspond accordingly. Therefore, the operations, features, and beneficial effects described above for the method also apply to the UAV identification device 400 based on small target detection and the units contained therein, and will not be repeated here.

[0127] The following is for reference. Figure 5 It illustrates a schematic diagram of the structure of an electronic device (e.g., a computing device) suitable for implementing some embodiments of the present disclosure. Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality or scope of the embodiments of this disclosure. Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system and a computer program. The computer program includes program instructions that, when executed, cause the processor to perform any of the methods described above. The processor provides computational and control capabilities to support the operation of the entire computer device. The internal memory provides an environment for the execution of the computer program in the non-volatile storage medium; when executed by the processor, the computer program causes the processor to perform any of the methods described above. The network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present disclosure and does not constitute a limitation on the computer device to which the present disclosure is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0128] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0129] In one embodiment, the processor is used to run a computer program stored in a memory to perform the following steps: acquiring a target image sequence for a target restricted airspace, wherein the target restricted airspace is a pre-defined airspace restricting UAV flight, and the target image is a fused image composed of multispectral images and visible light images; generating UAV flight intention features, UAV thermal source features, UAV morphological features, and UAV flight trajectory features based on the target image sequence and a target detection model; acquiring a first sound signal and a second sound signal, wherein the first sound signal and the second sound signal are acquired along a first direction and a second direction, respectively, both the first direction and the second direction being towards the target restricted airspace, and the first direction being a horizontal direction; performing environmental noise stripping on the first sound signal and the second sound signal respectively to obtain an optimized first sound signal and an optimized second sound signal; generating UAV voiceprint features based on the optimized first sound signal and the optimized second sound signal; and generating UAV identity information based on the UAV flight intention features, the UAV thermal source features, the UAV morphological features, the UAV flight trajectory features, and the UAV voiceprint features.

[0130] This disclosure also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, and the method implemented when the program instructions are executed can be referred to the various embodiments of the methods described above.

[0131] The aforementioned computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. Alternatively, the aforementioned computer-readable storage medium may be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0132] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0133] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for identifying unmanned aerial vehicles (UAVs) based on small target detection, characterized in that, include: Collect a sequence of target images for a restricted airspace, wherein the restricted airspace is a pre-defined airspace that restricts the flight of UAVs, and the target image is a fused image composed of multispectral images and visible light images; Based on the target image sequence and the target detection model, drone flight intention features, drone heat source features, drone morphology features and drone flight trajectory features are generated. The target detection model includes: drone localization network, drone heat source feature extraction network, drone morphology feature extraction network, drone flight trajectory feature extraction network and drone flight intention classifier. A first sound signal and a second sound signal are collected, wherein the first sound signal and the second sound signal are collected along a first direction and a second direction, respectively, and both the first direction and the second direction are directed toward the target restricted airspace, and the first direction is a horizontal direction; Environmental noise is removed from the first sound signal and the second sound signal respectively to obtain the optimized first sound signal and the optimized second sound signal; Based on the optimized first sound signal and the optimized second sound signal, generate the drone's voiceprint features; Based on the drone's flight intention characteristics, thermal source characteristics, morphological characteristics, flight trajectory characteristics, and voiceprint characteristics, drone identity information is generated, wherein... The step of generating UAV flight intent features, UAV thermal source features, UAV morphological features, and UAV flight trajectory features based on the target image sequence and target detection model includes: For each target image in the target image sequence, drone detection is performed on the target image through the target drone localization network to generate drone description information, resulting in a drone description information set. The drone description information includes: drone position coordinates and local feature maps, wherein the feature map size corresponding to the local feature map is consistent with the region size of the region of interest corresponding to the drone position coordinates. The drone flight trajectory features are generated based on the drone location coordinates included in the drone description information set and the drone flight trajectory feature extraction network. The drone heat source features are generated based on the local feature maps included in the drone description information set and the drone heat source feature extraction network. The drone morphological features are generated based on the local feature maps included in the drone description information set and the drone morphological feature extraction network. The drone flight intention features are generated based on the drone flight trajectory features and the drone flight intention classifier.

2. The method according to claim 1, characterized in that, Before acquiring the target image sequence for the restricted airspace, the method further includes: A ring-shaped restricted airspace is constructed with the clear airspace as the center and a preset radius difference as the ring width, wherein the clear airspace is the airspace that restricts the drone from flying into; Using a preset arc length as the inner arc length, the annular restricted airspace is divided into a fan-ring airspace set. Each fan-ring airspace in the fan-ring airspace set is assigned a scanning identifier. The scanning identifier indicates whether the corresponding fan-ring airspace has been scanned in the current scanning round. After the current scanning round is completed, the scanning identifiers corresponding to the fan-ring airspaces in the fan-ring airspace set will be initialized according to the scanning mode. The first sector-ring airspace in the current scan cycle, along the scan direction, is determined as the target restricted flight area.

3. The method according to claim 2, characterized in that, The method further includes: Send an authentication request to the target drone, wherein the target drone is the drone corresponding to the drone identity information, and the authentication request includes: authentication terminal identity information, session identifier and request level, wherein the session identifier represents the unique identifier of the authentication request, the request level represents the authentication level corresponding to the authentication request, and the request level includes: first request level, second request level and third request level; If the target drone fails to respond to the authentication request after a preset period of time, the target drone will be driven away. Upon receiving the request response information returned by the target drone, the system performs drone authentication and flight access permission verification on the target drone based on the request response information. In response to the target drone failing drone authentication or flight access permission verification, the target drone is driven away; In response to the target drone passing drone authentication and flight access permission verification and the request level being the third request level, the drone operation object corresponding to the target drone is authenticated according to the request response information; In response to failure to authenticate the target drone, the drone is driven away.

4. The method according to claim 3, characterized in that, The first sound signal is acquired by an airborne sound acquisition device, and the second sound signal is acquired by a ground-based sound acquisition device. The step of removing environmental noise from the first and second sound signals to obtain optimized first and second sound signals includes: The signal difference between the first sound signal and the basic environmental noise is determined to obtain the first noise-reduced sound signal. The basic environmental noise is the average environmental noise collected in the target restricted airspace. The noise impact degree of the basic environmental noise is controlled by the influence coefficient. The first denoised audio signal is encoded by an encoder to obtain a first encoded signal feature and a second encoded signal feature. The encoder consists of a one-dimensional convolutional module, a normalization layer and a target convolutional layer. The first encoded signal feature is the output feature of the one-dimensional convolutional module included in the encoder. The second encoded signal features are processed by an optimizer to obtain the third encoded signal features. The optimizer consists of three sub-optimizers connected in series. Each of the three sub-optimizers includes five one-dimensional convolutional modules. The optimized first audio signal is generated by the decoder, the first encoded signal feature, and the third encoded signal feature. The decoder consists of a first activation function, a target convolutional layer, a second activation function, and a one-dimensional convolutional module. The optimized first audio signal is the output signal of the one-dimensional convolutional module included in the decoder. The third encoded signal feature is processed sequentially by the first activation function, the target convolutional layer, and the second activation function included in the decoder, and then subjected to a tensor product operation with the first encoded signal feature to serve as the input feature of the one-dimensional convolutional module included in the decoder.

5. The method according to claim 4, characterized in that, The step of removing environmental noise from the first sound signal and the second sound signal respectively to obtain the optimized first sound signal and the optimized second sound signal further includes: Determine the signal difference between the second sound signal and the basic ambient noise to obtain the second noise-reduced sound signal; The encoder is used to encode the second noise-reduced audio signal to obtain the fourth and fifth encoded signal features. The fifth encoded signal features are processed by an optimizer to obtain the sixth encoded signal features. The optimized second sound signal is generated using the decoder, the fourth encoded signal feature, and the sixth encoded signal feature.

6. A drone identification device based on small target detection, characterized in that, include: The first acquisition unit is configured to acquire a sequence of target images for a target restricted airspace, wherein the target restricted airspace is a pre-defined airspace that restricts the flight of UAVs, and the target image is a fused image composed of multispectral images and visible light images. The first generation unit is configured to generate UAV flight intention features, UAV heat source features, UAV morphological features and UAV flight trajectory features based on the target image sequence and the target detection model. The target detection model includes: UAV localization network, UAV heat source feature extraction network, UAV morphological feature extraction network, UAV flight trajectory feature extraction network and UAV flight intention classifier. The second acquisition unit is configured to acquire a first sound signal and a second sound signal, wherein the first sound signal and the second sound signal are acquired along a first direction and a second direction, respectively, and both the first direction and the second direction are directed toward the target restricted airspace, and the first direction is a horizontal direction; An ambient noise stripping unit is configured to strip ambient noise from the first sound signal and the second sound signal respectively, to obtain an optimized first sound signal and an optimized second sound signal. The second generation unit is configured to generate UAV voiceprint features based on the optimized first sound signal and the optimized second sound signal; The third generation unit is configured to generate drone identity information based on the drone's flight intention features, the drone's heat source features, the drone's morphological features, the drone's flight trajectory features, and the drone's voiceprint features. The step of generating UAV flight intent features, UAV thermal source features, UAV morphological features, and UAV flight trajectory features based on the target image sequence and target detection model includes: For each target image in the target image sequence, drone detection is performed on the target image through the target drone localization network to generate drone description information, resulting in a drone description information set. The drone description information includes: drone position coordinates and local feature maps, wherein the feature map size corresponding to the local feature map is consistent with the region size of the region of interest corresponding to the drone position coordinates. The drone flight trajectory features are generated based on the drone location coordinates included in the drone description information set and the drone flight trajectory feature extraction network. The drone heat source features are generated based on the local feature maps included in the drone description information set and the drone heat source feature extraction network. The drone morphological features are generated based on the local feature maps included in the drone description information set and the drone morphological feature extraction network. The drone flight intention features are generated based on the drone flight trajectory features and the drone flight intention classifier.

7. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 5.

8. A computer-readable medium, characterized in that, It stores a computer program thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images