Unmanned aerial vehicle identification method based on YOLO v10-DID model

Abstract

The invention relates to the field of unmanned aerial vehicle identification, and provides an unmanned aerial vehicle identification method based on a YOLO v10-DID model, the YOLO v10-DID model in the method is established based on a YOLO v10 target detection model, the YOLO v10-DID model comprises an OFIE module, an MFF module and a DT module, the OFIE module is introduced to carry out optical flow estimation, motion information (optical flow information) between each pair of continuous frames in an image frame sequence can be effectively captured, and the motion information can be accurately identified. Gaussian filtering smoothing processing is carried out on optical flow information, the influence of noise motion information on a detection result can be reduced, target motion information and visual information are fused through an attention channel fusion module, the complementary advantages of the target motion information and the visual information are played, comprehensive and deep feature extraction of the unmanned aerial vehicle target is realized, and the detection accuracy is improved. And the detection precision and robustness of the YOLO v10 model on the moving unmanned aerial vehicle target are obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of drone identification, and in particular to a drone identification method based on a YOLO v10-DID model. Background Art

[0002] In recent years, drones have rapidly become popular in commercial and civilian fields. They have brought convenience to public utilities and personal life, but at the same time have also raised security issues such as privacy protection. Therefore, strengthening the management and supervision of drones has become an urgent issue that needs to be addressed.

[0003] Among existing technologies, drone identification technology can monitor airspace in real time, detect and locate drones entering sensitive areas in a timely manner to prevent them from posing a threat to people's privacy. In addition, the technology can also record the trajectory and activities of drones entering sensitive areas, which helps monitor and manage airspace resources.

[0004] Traditional drone identification methods include:

[0005] Vision-based recognition method: This method uses a camera to capture images or videos and extracts the features of the drone, such as shape, color, and texture, through image processing and computer vision technology. Then, the drone is identified and tracked through target detection and classification algorithms (such as deep learning convolutional neural networks). However, due to the small size of drones, large changes in flight altitude and speed, diverse appearance and structure, and the complexity of the activity environment, this method has obvious limitations in extracting features, detecting and tracking drone targets.

[0006] Single-sensor drone identification method based on radar, infrared, and sound waves: Radar identification has difficulty in distinguishing small and low-speed drones, and it is easy to mistake them for other flying objects; infrared identification is affected by ambient temperature changes and thermal radiation interference, and the identification accuracy may fluctuate under different temperature conditions; sound wave identification is sensitive to background noise, and other sounds or echoes in the environment may interfere with the drone's sound wave signal, affecting the identification accuracy.

[0007] UAV identification method integrating multiple sensors: This method can effectively make up for the limitations of a single sensor and improve identification performance by integrating data from sensors such as vision, radar, infrared, and sound waves, and using complex feature fusion algorithms. However, this method needs to address challenges such as algorithm complexity, data consistency, real-time, and cost. In addition, its deployment and maintenance costs are high, and it requires complex professional equipment and technical support. Summary of the invention

[0008] In view of the shortcomings of the prior art, the object of the present invention is to provide a drone identification method based on the YOLO v10-DID (detect indrone) model. The YOLO v10-DID model in the method is built based on the YOLO v10 target detection model, including an OFIE (Optical Flow Information Extraction) module, an MFF (Multimodal Feature Fusion) module and a DT (detect) module; the identification steps include:

[0009] S1: Split the video to be detected acquired by the camera into a continuous image frame sequence, that is, extract each frame of the video at a fixed time interval to form a time-consistent image frame sequence;

[0010] S2: Use the OFIE module to extract optical flow information from the image frame sequence to obtain the optical flow frame sequence between every two frames of images;

[0011] S3: Convert the generated optical flow frame sequence into a python array format corresponding to the image frame sequence;

[0012] S4: Delete the first frame of the image frame sequence to ensure a one-to-one correspondence between the number of optical flow frames and image frames;

[0013] S5: Use Gaussian filtering to smooth the optical flow frame sequence;

[0014] S6: combining the two channels of the processed optical flow frame sequence with the three channels of the corresponding image frame sequence to form a multimodal optical flow-image information matrix containing five channels, which is input into the MFF module to generate a comprehensive feature representation;

[0015] S7: Input the fused features into the DT module to identify and locate the UAV target;

[0016] S8: Output the results of YOLO v10-DID model recognition, including the category, location and confidence score of the drone.

[0017] Furthermore, the OFIE module is a layered stacked optical flow estimation network, which is composed of four sub-networks: FlowNetC, FlowNetSA, FlowNetS and FlowNet-SD. The architecture of FlowNetC contains two independent convolutional network branches, which respectively extract features from the input image pairs, and then calculate the similarity between feature maps through the correlation layer to capture large displacement motion information. FlowNetC is responsible for processing large displacement motion in image pairs, FlowNetSA and FlowNetS are responsible for refining optical flow estimation, and FlowNet-SD is responsible for processing small displacement motion in image pairs.

[0018] Furthermore, the FlowNetSA module introduces a self-attention mechanism layer between the conv3 layer and the conv3_1 layer of the FlowNetS module.

[0019] Furthermore, the feature map Where C is the number of channels of the feature map, W is the width of the feature map, and H is the height of the feature map; the self-attention mechanism layer first generates query, key, and value matrices through three 1×1 convolutions:

[0020]

[0021] Then use pytorch's view function to The query matrix Q and key matrix K of shape are transformed into A matrix Q of shape 1 And the matrix K 1 , and use pytorch's permute function to A matrix Q of shape 1 Convert to A matrix Q of shape 2 :

[0022]

[0023] Q 2 =Q 1 .permute(1,0)

[0024] Finally, calculate the energy matrix E:

[0025]

[0026] The element at the (i, j)th position in E is the influence of the jth element of the input feature map on the ith element, thereby obtaining the dependency relationship between any two elements; Softmax normalization is performed on each row of the energy matrix E to obtain the attention matrix:

[0027]

[0028] The element at the (i, j)th position in A is the influence of the jth element of the input feature map on the ith element, and the sum of each row of A is 1; use the view function to convert the C×W×H value matrix V into a C×(W×H) shape matrix V 1 :

[0029] V 1 =V.view(C,W×H)

[0030] Calculate the final output matrix Y of this module:

[0031] Y=γ(A×V 1 T ) T .view(C,W,H)+X

[0032] Among them, the coefficient γ is obtained through learning, with an initial value of 0, and the output is the original feature map X. As the learning deepens, the coefficient γ adds a weighted attention feature to the original feature map X, and the global dependency relationship between any two positions in the feature map can be obtained.

[0033] Furthermore, the workflow of the MFF module is as follows: the five-channel optical flow-image information is converted into a feature map X of the C channel through a convolutional layer and a ReLU activation function. conv , then X conv Perform maximum pooling and average pooling to obtain a global description of each channel;

[0034] X conv The size of is C×W×H, where C is the number of channels of the feature map, W is the width of the feature map, and H is the height of the feature map. For each channel c, the formula for global maximum pooling is:

[0035] X max (c) = max X conv (i,j,c)

[0036] Among them, X conv (i, j, c) represents the value of the input feature map at position (i, j) and channel c, X max (c) is the global maximum value of channel c, indicating the most significant feature in this channel; the result of global maximum pooling is a vector of size C, indicating the global maximum value of each channel;

[0037] Average pooling extracts the overall feature information of each channel by calculating the global average of each channel feature map; average pooling captures the global distribution of features in the channel:

[0038]

[0039] Among them, X conv (i, j, c) represents the value of the input feature map at position (i, j) and channel c, X avg (c) is the global average of channel c; the result of global average pooling is a vector of size C, which represents the global average of each channel; the results of global maximum pooling and average pooling are concatenated together to obtain a vector X of size 2C fused , mapped back to C channels through two fully connected layers, and normalized with the Sigmoid activation function to ensure that the output weight value is between 0 and 1, and obtain X weights , X weights With X conv Multiply each channel by channel to adjust the response strength of each channel, highlight important features, and suppress irrelevant or redundant information.

[0040] Furthermore, the DT module is a standard YOLO v10 neural network structure. The neural network structure of YOLO v10 starts from the input features and first extracts features through multiple convolutional layers. The model initially contains several basic convolutional layers and C2f modules for extracting low-level features. Subsequently, the C2f / CIB module and the SCDown module are introduced to extract higher-level features and expand the perception field of view through downsampling to capture a wider range of contextual information. After the feature extraction part, the network enters the SPPF module, which further enriches the feature expression through multi-scale spatial pyramid pooling. Then, the PSA module aggregates the features and integrates information at different levels to enhance the overall semantic understanding ability. In the Neck part, an upsampling operation is first performed to restore a higher-resolution feature map, and then the features at different levels are fused through multiple Concat operations. These fused feature maps are further processed by the C2f / CIB module to obtain features with stronger expressive ability.

[0041] Furthermore, the optical flow data in step S2 is a series of floating point numbers, and each optical flow data contains two values, which are the displacement of the pixel in the horizontal and vertical directions respectively.

[0042] Further, in step S5, the calculation method for smoothing the optical flow frame sequence using Gaussian filtering is:

[0043] Create a Gaussian kernel matrix G(x,y), whose formula is:

[0044]

[0045] Among them, (x, y) represents the distance between the pixel and the center of the kernel, and σ represents the standard deviation;

[0046] The pixel matrix I of the two channels of the optical flow frame 1 and I 2 Perform convolution operations separately and apply the Gaussian kernel to I 1 and I 2 , the convolution formula is:

[0047]

[0048] Among them, I' 1 (x,y) and I' 2 (x,y) is the processed pixel matrix, and k is the radius of the Gaussian kernel.

[0049] Furthermore, the image frame sequence in step S1 is a static image frame of a video in jpg or png format, and the fixed time interval in step S1 is Second.

[0050] Furthermore, the optical flow frame sequence in step S2 is a file in flo format, in which each element corresponds to the optical flow information between two adjacent image frames.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. The present invention introduces the OFIE module for optical flow estimation, which can effectively capture the motion information (optical flow information) between each pair of consecutive frames in the image frame sequence. The optical flow information is smoothed by Gaussian filtering to reduce the influence of noise motion information on the detection results. The target motion information and visual information are fused through the attention channel fusion module to give play to their complementary advantages, realize comprehensive and deep feature extraction of UAV targets, and significantly improve the detection accuracy and robustness of the YOLO v10 model for moving UAV targets.

[0053] 2. The present invention can achieve efficient drone detection by relying only on ordinary cameras, reducing the cost and complexity of the system. Each processing step (such as optical flow estimation, Gaussian filtering, feature fusion, etc.) has high computational efficiency, can capture and process the target's motion information and visual information in real time, and has good timeliness and responsiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the specific implementation of the present invention or the technical solution in the prior art, the drawings required for describing the specific implementation or the prior art will be briefly introduced below. Obviously, the drawings described below are only one implementation of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1This is a schematic diagram of the OFIE module structure in the YOLO v10-DID model.

[0056] Figure 2 This is a schematic diagram of the MFF module structure in the YOLO v10-DID model.

[0057] Figure 3 This is a schematic diagram of the network model structure of YOLO v10-DID.

[0058] Figure 4 It is a schematic diagram of the working process of the drone identification of the present invention. DETAILED DESCRIPTION

[0059] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the technical scheme in the specific implementation mode of the present invention is clearly and completely described below to further illustrate the present invention. Obviously, the specific implementation mode described is only a part of the implementation mode of the present invention, rather than all styles.

[0060] In the drone identification method based on the YOLO v10-DID model of the present invention, the YOLO v10-DID model is built based on the YOLO v10 target detection model, including an OFIE module, an MFF module and a DT module (such as Figure 3 shown).

[0061] like Figure 1 As shown in the figure, the OFIE module is a layered stacked optical flow estimation network, which is composed of multiple sub-networks (including FlowNetC, FlowNetSA, FlowNetS and FlowNet-SD). FlowNetC uses a correlation layer to capture large displacement motion in image pairs, FlowNetSA and FlowNetS further refine the optical flow estimation, and FlowNet-SD specializes in small displacement motion, forming a multi-scale optical flow estimation process from coarse to fine, from large displacement to small displacement. The entire network optimizes the optical flow estimation results layer by layer by stacking these sub-networks, first pre-training on the FlyingChairs dataset, and then fine-tuning on the FlyingThings3D dataset, and finally achieving high-precision optical flow estimation.

[0062] Among them, FlowNetC, FlowNetS and FlowNet-SD are standard FlowNet2 (FlowNet2 is a deep learning model for optical flow estimation. It uses the cascade and fusion of multiple sub-networks to utilize fine-grained and coarse-grained flow information to achieve higher accuracy and robustness) module structures. The architecture of FlowNetC contains two independent convolutional network branches, which extract features from the input image pairs respectively. After that, the similarity between feature maps is calculated through the correlation layer to effectively capture large displacement motion information. FlowNetS uses a single convolutional network to directly extract features from the input image pairs and estimate optical flow through continuous convolution operations. FlowNetS has a simpler design and higher computational efficiency, which is suitable for processing scenes with small displacement or simple motion. FlowNet-SD is optimized on the basis of FlowNetS, and the structure is improved specifically for small displacement scenes. It adds more refined convolution operations to better capture detailed motion and small displacement changes.

[0063] The present invention designs a FlowNetSA module for drone target recognition. This module introduces a self-attention mechanism layer for drone detection between the conv3 layer and the conv3_1 layer of the standard FlowNetS module. (i.e., the shape of X is C×W×H), where C is the number of channels of the feature map, W is the width of the feature map, and H is the height of the feature map. The self-attention mechanism layer first generates the query, key, and value matrices through three 1×1 convolutions:

[0064]

[0065] Then use pytorch's view function to convert The query matrix Q and key matrix K of shape are transformed into A matrix Q of shape 1 And the matrix K 1 , and use pytorch's permute function to A matrix Q of shape 1 Convert to A matrix Q of shape 2 :

[0066]

[0067] Q 2 =Q 1 .permute(1,0)

[0068] Calculate the energy matrix E:

[0069]

[0070] The significance of this step is that the element at the (i, j)th position in E is the influence of the jth element of the input feature map on the ith element, thereby obtaining the dependency relationship between any two elements. Perform Softmax normalization on each row of the energy matrix E (Softmax normalization is a function that converts a real vector into a probability distribution. It maps each element between 0 and 1 and ensures that the sum of all elements is 1. It is often used in the output layer of multi-category classification problems to predict the probability of each category) to obtain the attention matrix:

[0071]

[0072] The element at the (i, j)th position in A is the influence of the jth element of the input feature map on the ith element, and the sum of each row of A is 1. Use the view function to convert the C×W×H value matrix V into a C×(W×H) shape matrix V 1 :

[0073] V 1 =V.view(C,W×H)

[0074] Calculate the final output matrix Y of this module:

[0075] Y=γ(A×V 1 T ) T .view(C,W,H)+X

[0076] Among them, the coefficient γ is obtained through learning, with an initial value of 0, and the output is the original feature map X. As the learning deepens, the coefficient γ adds weighted attention features to the original feature map X, and the global dependency between any two positions in the feature map can be obtained. By introducing the self-attention mechanism, FlowNetSA can adaptively weight each position of the feature map and capture the motion relationship between the global and local in the image. For long-distance or high-altitude scenes, FlowNetSA's self-attention module can effectively focus on small areas in the image and strengthen the feature representation of these areas, which helps the network to more accurately identify the dynamic characteristics of drones and improve the detection accuracy of drones.

[0077] The MFF module neural network structure is as follows Figure 2 As shown, the optical flow-image information of the five channels passes through the convolution layer and the RELU activation function to obtain the optical flow-image feature map of C channels (C is initially three and can be adjusted later), denoted as X conv Next, for X conv Perform maximum pooling and average pooling to obtain a global description of each channel. The role of maximum pooling is to extract the most significant features by taking the maximum value in each region of each channel feature map. For global maximum pooling, the maximum value in each channel is regarded as the global description of the channel. Let X conv The size of is C×W×H, where C is the number of channels of the feature map, W is the width of the feature map, and H is the height of the feature map. For each channel c, the formula for global maximum pooling is:

[0078] X max (c) = max X conv (i,j,c)

[0079] Among them, X conv (i, j, c) represents the value of the input feature map at position (i, j) and channel c, X max (c) is the global maximum of channel c, indicating the most significant feature in that channel. The result of global maximum pooling is a vector of size C, indicating the global maximum of each channel.

[0080] Average pooling extracts the overall feature information of each channel by calculating the global average of the feature map of each channel. Average pooling captures the global distribution of features in the channel:

[0081]

[0082] Among them, X conv (i, j, c) represents the value of the input feature map at position (i, j) and channel c, X avg (c) is the global average of channel c. The result of global average pooling is a vector of size C, which represents the global average of each channel. The results of global maximum pooling and average pooling are concatenated together to obtain a vector X of size 2C. fused , mapped back to C channels through two fully connected layers, and normalized with the Sigmoid activation function (the Sigmoid activation function is a nonlinear function widely used in neural networks. Its output value range is between (0,1), and it can map any value to between 0 and 1. It is often used as the activation function of the output layer in classification problems) to ensure that the output weight value is between 0 and 1. weights , X weights With X convMultiply each channel by channel to adjust the response strength of each channel, highlight important features, and suppress irrelevant or redundant information.

[0083] In the drone detection task, the MFF module fuses optical flow and image information based on the channel attention mechanism, which can effectively combine the dynamic and static features of the drone. Optical flow information captures the motion characteristics of the drone, and image information provides appearance details. The maximum pooling and average pooling are used to extract significant features and global features, and then adaptive weight adjustment is used to highlight key information and suppress irrelevant noise. This fusion method improves the robustness of the model in complex scenes and can detect drones more accurately, especially when the lighting changes, the background is complex, or the drone moves quickly.

[0084] The DT module is a standard YOLO v10 neural network structure. The YOLO v10 neural network structure starts with input features and first extracts features through multiple convolutional layers. The model initially contains several basic convolutional layers (marked as Conv) and C2f modules for extracting low-level features. Subsequently, the C2f / CIB module and SCDown module are introduced to extract higher-level features and expand the receptive field of view through downsampling to capture a wider range of contextual information. After the feature extraction part, the network enters the SPPF (Spatial Pyramid Pooling Fast) module, which further enriches the feature expression through multi-scale spatial pyramid pooling. Next, the PSA (Path Aggregation Network) module aggregates the features and integrates information at different levels to enhance the overall semantic understanding ability. In the Neck part, first upsampling is performed to restore higher-resolution feature maps, and then the features at different levels are fused through multiple Concat operations. These fused feature maps are further processed by the C2f / CIB module to obtain features with stronger expressive power. The network uses two prediction heads during training, one using a one-to-many label assignment strategy and the other using a one-to-one label assignment strategy, to exploit the rich supervisory signals of the one-to-many label assignment strategy during training, and the prediction results of the one-to-one label assignment strategy during inference to achieve efficient inference without NMS.

[0085] like Figure 4 As shown in the figure, the drone identification method based on the YOLO v10-DID model specifically includes the following steps:

[0086] S1: Split the video to be detected acquired by the camera into a continuous image frame sequence, that is, extract each frame of the video to form an image frame sequence with consistent time sequence; the image frame sequence is a static image frame of the video in jpg or png format.

[0087] S2: Use OFIE module to extract optical flow information from the image frame sequence to obtain the optical flow frame sequence between every two frames of image; the optical flow frame sequence is a file in flo format, where each element corresponds to the optical flow information between two adjacent image frames. For example, a video contains N frames, numbered F. 1 、F 12 ,……F N , then the optical flow frame sequence contains N-1 optical flow frames, which are used to represent F 1 With F 2 The optical flow between 2 With F 3 The optical flow between, ... F N-1 With F N The optical flow data is a series of floating point numbers (usually 32-bit floating point numbers), and each optical flow data contains two values, which are the displacement of the pixel in the horizontal and vertical directions.

[0088] S3: Convert the generated optical flow frame sequence into the python array format corresponding to the image frame sequence; the first and second dimensions of the two represent the pixel matrix height and width respectively, the third dimension of the optical flow frame sequence has two channels, representing the horizontal and vertical movement distances of the pixels from the previous frame to the next frame (in pixels), and the third dimension of the image frame sequence has three channels, representing the red, green, and blue colors of the pixels respectively.

[0089] S4: Delete the first frame of the image frame sequence to ensure a one-to-one correspondence between the number of optical flow frames and image frames;

[0090] S5: Use Gaussian filtering to smooth the optical flow frame sequence; the calculation method is:

[0091] Create a Gaussian kernel matrix G(x,y), whose formula is:

[0092]

[0093] Among them, (x, y) represents the distance between the pixel and the center of the kernel, and σ represents the standard deviation, which determines the degree of smoothing;

[0094] The pixel matrix I of the two channels of the optical flow frame 1 and I 2 Perform convolution operations separately and apply the Gaussian kernel to I 1 and I 2 , the convolution formula is:

[0095]

[0096] Among them, I' 1 (x,y) and I' 2(x,y) is the processed pixel matrix, and k is the radius of the Gaussian kernel.

[0097] S6: combining the two channels of the processed optical flow frame sequence with the three channels of the corresponding image frame sequence to form a multimodal optical flow-image information matrix containing five channels, which is input into the MFF module to generate a comprehensive feature representation;

[0098] S7: Input the fused features into the DT module to identify and locate the UAV target;

[0099] S8: Output the results of YOLO v10-DID model recognition, including the category, location and confidence score of the drone.

[0100] The above describes the main technical features and basic principles of the present invention and the related advantages. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the concept or basic features of the present invention. Therefore, no matter from which point of view, the above specific embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention.

[0101] In addition, it should be understood that although the present specification is described according to various implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each implementation method may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A drone identification method based on the YOLO v10-DID model, characterized in that: The YOLO v10-DID model is built based on the YOLO v10 target detection model, including the OFIE module, MFF module, and DT module; the recognition steps include: S1: Split the video to be detected acquired by the camera into a continuous image frame sequence, that is, extract each frame of the video at a fixed time interval to form a time-consistent image frame sequence; S2: Use the OFIE module to extract optical flow information from the image frame sequence to obtain the optical flow frame sequence between every two frames of images; S3: Convert the generated optical flow frame sequence into a python array format corresponding to the image frame sequence; S4: Delete the first frame of the image frame sequence to ensure a one-to-one correspondence between the number of optical flow frames and image frames; S5: Use Gaussian filtering to smooth the optical flow frame sequence; S6: combining the two channels of the processed optical flow frame sequence with the three channels of the corresponding image frame sequence to form a multimodal optical flow-image information matrix containing five channels, which is input into the MFF module to generate a comprehensive feature representation; S7: Input the fused features into the DT module to identify and locate the UAV target; S8: Output the results of YOLO v10-DID model recognition, including the category, location and confidence score of the drone.

2. The method for identifying drones based on the YOLO v10-DID model according to claim 1, characterized in that: The OFIE module is a layered optical flow estimation network, which is composed of four sub-networks: FlowNetC, FlowNetSA, FlowNetS and FlowNet-SD. The architecture of FlowNetC contains two independent convolutional network branches, which respectively extract features from the input image pairs, and then calculate the similarity between feature maps through the correlation layer to capture large displacement motion information. FlowNetC is responsible for processing large displacement motion in image pairs, FlowNetSA and FlowNetS are responsible for refining optical flow estimation, and FlowNet-SD is responsible for processing small displacement motion in image pairs.

3. The method for identifying drones based on the YOLOv10-DID model according to claim 2, characterized in that: The FlowNetSA module introduces a self-attention mechanism layer between the conv3 layer and the conv3_1 layer of the FlowNetS module.

4. The method for identifying drones based on the YOLOv10-DID model according to claim 3, characterized in that: The feature map Where C is the number of channels of the feature map, W is the width of the feature map, and H is the height of the feature map; the self-attention mechanism layer first generates query, key, and value matrices through three 1×1 convolutions: Then use pytorch's view function to The query matrix Q and key matrix K of shape are transformed into The shape of the matrix Q1 and matrix K1, and use the pytorch permute function The matrix Q1 of shape is transformed into Matrix Q2 of shape: Q2 = Q1.permute(1,0) Finally, calculate the energy matrix E: The element at the (i, j)th position in E is the influence of the jth element of the input feature map on the ith element, thereby obtaining the dependency relationship between any two elements; Softmax normalization is performed on each row of the energy matrix E to obtain the attention matrix: The element at the (i, j)th position in A is the influence of the jth element of the input feature map on the ith element, and the sum of each row of A is 1; use the view function to convert the C×W×H value matrix V into a C×(W×H) shape matrix V1: V1=V.view(C,W×H) Calculate the final output matrix Y of this module: Y=γ(A×V1 T ) T .view(C,W,H)+X Among them, the coefficient γ is obtained through learning, with an initial value of 0, and the output is the original feature map X. As the learning deepens, the coefficient γ adds a weighted attention feature to the original feature map X, and the global dependency relationship between any two positions in the feature map can be obtained.

5. The method for identifying drones based on the YOLOv10-DID model according to claim 1, characterized in that: The workflow of the MFF module is as follows: the five-channel optical flow-image information is converted into a feature map X of the C channel through a convolutional layer and a ReLU activation function. conv , then X conv Perform maximum pooling and average pooling to obtain a global description of each channel; X conv The size of is C×W×H, where C is the number of channels of the feature map, W is the width of the feature map, and H is the height of the feature map. For each channel c, the formula for global maximum pooling is: X max (c)=max X conv (i,j,c) Among them, X conv (i, j, c) represents the value of the input feature map at position (i, j) and channel c, X max (c) is the global maximum value of channel c, indicating the most significant feature in this channel; the result of global maximum pooling is a vector of size C, indicating the global maximum value of each channel; Average pooling extracts the overall feature information of each channel by calculating the global average of each channel feature map; average pooling captures the global distribution of features in the channel: Among them, X conv (i, j, c) represents the value of the input feature map at position (i, j) and channel c, X avg (c) is the global average of channel c; the result of global average pooling is a vector of size C, which represents the global average of each channel; the results of global maximum pooling and average pooling are concatenated together to obtain a vector X of size 2C fused , mapped back to C channels through two fully connected layers, and normalized with the Sigmoid activation function to ensure that the output weight value is between 0 and 1, and obtain X weights , X weights With X conv Multiply each channel by channel to adjust the response strength of each channel, highlight important features, and suppress irrelevant or redundant information.

6. The method for identifying unmanned aerial vehicles based on the YOLO y10-DID model according to claim 1, characterized in that: The DT module is a standard YOLO v10 neural network structure. The neural network structure of YOLO v10 starts from the input features and first extracts features through multiple convolutional layers. The model initially contains several basic convolutional layers and C2f modules for extracting low-level features. Subsequently, the C2f / CIB module and the SCDown module are introduced to extract higher-level features and expand the perception field of view through downsampling to capture a wider range of contextual information. After the feature extraction part, the network enters the SPPF module, which further enriches the feature expression through multi-scale spatial pyramid pooling. Then, the PSA module aggregates the features and integrates information at different levels to enhance the overall semantic understanding ability. In the Neck part, an upsampling operation is first performed to restore a higher-resolution feature map, and then the features at different levels are fused through multiple Concat operations. These fused feature maps are further processed by the C2f / CIB module to obtain features with stronger expressive ability.

7. The method for identifying drones based on the YOLO v10-DID model according to claim 1, characterized in that: The optical flow data in step S2 is a series of floating point numbers, and each optical flow data contains two values, which are the displacement of the pixel in the horizontal and vertical directions respectively.

8. The method for identifying drones based on the YOLOv10-DID model according to claim 1, characterized in that: The calculation method for smoothing the optical flow frame sequence using Gaussian filtering in step S5 is: Create a Gaussian kernel matrix G(x,y), whose formula is: Among them, (x, y) represents the distance between the pixel and the center of the kernel, and σ represents the standard deviation; The pixel matrices I1 and I2 of the two channels of the optical flow frame are convolved respectively, and the Gaussian kernel is applied to I1 and I2. The convolution formula is: Where I'1x,y and I'2x,y are the processed pixel matrices, and k is the radius of the Gaussian kernel.

9. The method for identifying drones based on the YOLOv10-DID model according to claim 1, characterized in that: The image frame sequence in step S1 is a static image frame of a video in jpg or png format, and the fixed time interval in step S1 is Second.

10. The method for identifying drones based on the YOLOv10-DID model according to claim 1, characterized in that: The optical flow frame sequence in step S2 is a file in flo format, in which each element corresponds to the optical flow information between two adjacent image frames.

Images