Detection method of black flight unmanned aerial vehicle
Through a one-way fusion strategy, high, medium and low-level feature extraction was carried out, and the black flying drone detection model was established, which solved the accuracy and speed of the black flying drone detection around the power facilities, and achieved efficient and accurate detection results.
Patent Information
- Application Number
- CN202510170443.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-17
- Publication Date
- 2025-07-04
AI Technical Summary
The prior art is difficult to quickly and accurately detect and identify black-flying drones around power facilities, resulting in difficulty in protecting power facilities safety.
A one-way fusion strategy is used to extract high, medium and low-level feature, and a black-fly drone detection model is established through deformable convolution and simple feature stitching and upsampling operations, and a legal drone simulates the flight images of black-fly drones for training to achieve fast and accurate detection.
It improves the accuracy and speed of Black Flying UAV detection, reduces the computing volume and memory requirements, reduces the risk of network instability, and avoids information redundancy.
Smart Images
Figure CN120259718A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for detecting unlicensed drones, belonging to the technical field of power system safety. Background Art
[0002] With the growth of drone applications, the phenomenon of unauthorized "unlicensed flight" has gradually become a major issue of social concern. The so-called "unlicensed flight" drones refer to drones that fly without legal registration, approval or permission. The existence of such drones not only threatens public safety but also poses significant risks to the facilities and operations of specific industries. Especially in the field of power systems, the potential threat of "unlicensed flight" drones is particularly prominent.
[0003] Equipment such as power system transmission lines, substations, and high-voltage towers are its core components. These facilities are widely distributed, and many are located in remote or complex terrains, making it difficult to conduct round-the-clock manual patrols and monitoring. Therefore, the safety protection of power facilities faces huge challenges, especially under the threat of "unlicensed flight" drones. The flight altitude of drones is close to power facilities, especially transmission lines and high-voltage towers. The metal structure is likely to attract drones to approach, increasing the danger. In addition, due to their flexible, concealed, and easy-to-operate characteristics, drones may be used for malicious destruction or stealing sensitive information, threatening the safety of the power grid.
[0004] In view of the threat of "unlicensed flight" drones to power facilities and the limitations of existing technologies in real-time monitoring and intervention, there is an urgent need for a method to actively detect "unlicensed flight" drones, which can automatically identify unlicensed drones and give warnings to ensure the safe and stable operation of power facilities. Summary of the Invention
[0005] The technical problem to be solved by the present invention is: how to quickly and accurately detect unlicensed drones.
[0006] To solve the above technical problem, the technical solution proposed by the present invention is: a method for detecting unlicensed drones, comprising the following steps:
[0007] Step 1: Use a legal drone to simulate the flight situation of an unlicensed drone around power facilities and capture and collect N simulated unlicensed drone flight images; preprocess the N simulated unlicensed drone flight images, and collect the preprocessed N simulated unlicensed drone flight images to form an unlicensed drone flight image dataset , as shown in the following formula (1),
[0008] (1)
[0009] In formula (1), They are respectively the first simulated black flight UAV flight image pixel map, the second simulated black flight UAV flight image pixel map to the Nth simulated black flight UAV flight image pixel map among the N preprocessed simulated black flight UAV flight images; the are all pixel maps with a height of H and a width of W; to are respectively the coordinates in to the coordinates on the pixel values; to are respectively the coordinates in to the coordinates on the pixel values; to are respectively the coordinates in to the coordinates on the pixel values; to are respectively the coordinates in to the coordinates on the pixel values; to are respectively the coordinates in to the coordinates on the pixel values; to are respectively the coordinates in to the coordinates on the pixel values;
[0010] Step 2: Perform first feature extraction on the above-mentioned with a height of H and a width of W through deformable convolution with the number of convolution kernels being in sequence, to obtain N first simulated black flight UAV flight image first feature maps with a height of , a width of , and a number of channels of ; the , the second simulated black flight UAV flight image first feature map to the Nth simulated black flight UAV flight image first feature map ; the , to are shown in the following formula (2),
[0011] (2)
[0012] In formula (2), are respectively the first sub-feature maps of the first channels of the first simulated black flight UAV flight images obtained by performing feature extraction through deformable convolutions with the number of convolutional kernels being , the first sub-feature maps of the second channels of the first simulated black flight UAV flight images to the first sub-feature maps of the th channels of the first simulated black flight UAV flight images; are respectively the first sub-feature maps of the first channels of the second simulated black flight UAV flight images obtained by performing feature extraction through deformable convolutions with the number of convolutional kernels being , the first sub-feature maps of the second channels of the second simulated black flight UAV flight images to the first sub-feature maps of the th channels of the second simulated black flight UAV flight images; are respectively the first sub-feature maps of the first channels of the Nth simulated black flight UAV flight images obtained by performing feature extraction through deformable convolutions with the number of convolutional kernels being , the first sub-feature maps of the second channels of the Nth simulated black flight UAV flight images to the first sub-feature maps of the th channels of the Nth simulated black flight UAV flight images;
[0013] Step 3: Perform second feature extraction on the first feature map of the first simulated black flight UAV flight image to obtain the second feature map of the first simulated black flight UAV flight image ; The specific content of performing second feature extraction on the first feature map of the first simulated black flight UAV flight image is as follows:
[0014] Step 3.1: The first feature map of the first simulated black flight UAV flight image is evenly divided into G first sub-feature maps of the first simulated black flight UAV flight images with a height of , a width of , and a number of channels of , the first sub-feature map of the first simulated black flight UAV flight image , the second sub-feature map of the first simulated black flight UAV flight image , … to the Gth sub-feature map of the first simulated black flight UAV flight image , where the , , … to are as shown in the following formula (3), (3)
[0015] The , , … to Substitute the filtering assignment features into the following formula (4) in sequence to obtain the first filtered sub-feature map of the first simulated drone flight image for black flight , the second filtered sub-feature map of the first simulated drone flight image for black flight , … to the G-th filtered sub-feature map of the first simulated drone flight image for black flight ,
[0016] (4)
[0017] In formula (4), is a non-linear activation function;
[0018] Step 3.2: Perform pooling on the first filtered sub-feature map of the first simulated drone flight image for black flight in the horizontal and vertical directions,
[0019] (5)
[0020] In formula (5), and are respectively the first filtered horizontal sub-feature map and the first filtered vertical sub-feature map of the first simulated drone flight image for black flight obtained by performing pooling on the in the horizontal and vertical directions; the height of the remains unchanged at , the width is pooled to a single value, and the number of channels remains unchanged at ; the height of the is pooled to a single value, the width remains unchanged at , and the number of channels remains unchanged at ; the is the feature at the coordinate on the c-th channel in the ;
[0021] Perform splicing, channel compression, and fusion on the and through the following formula (6) to form the first filtered spliced sub-feature map of the first simulated drone flight image for black flight ,
[0022] (6)
[0023] In formula (6), is the first weight matrix; is an operation of splicing multiple feature maps along a specified dimension, that is, splicing two feature maps , along the channel dimension to form a new feature map;
[0024] The first filtered and spliced sub-feature map has a dimension of , and a number of channels of ;
[0025] Perform weight assignment and normalization on each channel in the and sum them up to obtain the second feature map of the first simulated black-flying UAV flight image ,
[0026] (7)
[0027] In formula (7), is to stack the two-dimensional feature maps of all channels in each subgroup on a new dimension to form a complete three-dimensional feature; are respectively the first channel, the second channel to the channel in the first filtered and spliced sub-feature map after weight assignment and normalization to obtain the first filtered and spliced first-channel sub-feature map, the first filtered and spliced second-channel sub-feature map to the first filtered and spliced channel sub-feature map of the first simulated black-flying UAV flight image; c is the number of channels; is the first filtered and spliced c-channel weight sub-feature map of the first simulated black-flying UAV flight image obtained by weight assignment of the c-channel in the first filtered and spliced sub-feature map ; and are respectively the mean and standard deviation of the c-channel in the first filtered and spliced sub-feature map ; is a stability factor, which is an empirical constant; is the feature value of the c-channel in; is the c-channel weight matrix of the first simulated black-flying UAV flight image; is the second filtered and spliced sub-feature map of the first simulated black-flying UAV flight image obtained by the second convolution; is the second weight matrix; is the convolution bias term;
[0028] Step 4: Repeat the content of Step 3 to perform second feature extraction on the second feature map of the second simulated black-flying UAV flight image , the first feature map of the third simulated black-flying UAV flight image , … to the first feature map of the Nth simulated black-flying UAV flight image in sequence to obtain the second feature map of the first simulated black-flying UAV flight image , the second feature map of the flight image of the third simulated unauthorized drone to the second feature map of the flight image of the Nth simulated unauthorized drone ;
[0029] Perform first feature fusion on the above-mentioned , ,..., to successively form the first fused feature map of the flight image of the first simulated unauthorized drone , the first fused feature map of the flight image of the second simulated unauthorized drone to the first fused feature map of the flight image of the Nth simulated unauthorized drone , (8);
[0030] Step 5: Perform second feature fusion on the first fused feature map of the flight image of the first simulated unauthorized drone to form the second fused feature map of the flight image of the first simulated unauthorized drone ,
[0031] (9)
[0032] In formula (9), , , are learnable weights automatically obtained through training, represents channel-wise weighting, , , are the convolutional kernel weights for extracting high-level, middle-level, and low-level feature maps respectively, , , are bias terms, is a convolution with a stride of 2, is an upsampling operation, is a concatenation operation; , , are the low-level, middle-level, and high-level feature maps generated by performing convolution operations on respectively; is the feature obtained by upsampling the fused middle feature map , is the feature map obtained by upsampling ; is the feature obtained by concatenating the upsampled feature with the low-level feature , is the feature obtained by concatenating the upsampled high-level feature with the middle-level feature Features obtained by splicing and and are the low-resolution fusion feature, medium-resolution fusion feature, and high-resolution feature obtained after channel compression, respectively
[0033] The second fusion feature map of the first simulated illegal drone flight image has a height of and a width of and a number of channels of ;
[0034] Perform global average pooling on each channel of the and splice the results to form the global feature vector of the first simulated illegal drone flight image ,
[0035] (10)
[0036] In formula (10), are the feature values after global average pooling of the sub-feature maps of the first channel, the second channel to the channel in the respectively;
[0037] Step 6: Repeat the content of Step 5 to process the first fusion feature map of the second simulated illegal drone flight image to the first fusion feature map of the Nth simulated illegal drone flight image to obtain the global feature vector of the second simulated illegal drone flight image to the global feature vector of the Nth simulated illegal drone flight image , and collect the global feature vector of the first simulated illegal drone flight image , the global feature vector of the second simulated illegal drone flight image to the global feature vector of the Nth simulated illegal drone flight image to form the set Z of global feature vectors of illegal drone flights;
[0038] Step 7: Establish an illegal drone detection model, as shown in the following formula (11),
[0039] (11)
[0040] In formula (11), is the detection result of the illegal drone detection model; is the detection category of the mth object among a total of K detection objects; is the detection position of the m-th object among a total of K detection objects; is the maximum probability selection function; is the normalization function; is the classification weight of the black flying UAV detection model; is the global feature vector of the black flying UAV flight input into the black flying UAV detection model; is the weight vector of the fully connected layer in the black flying UAV detection model; is the bias term of the fully connected layer in the black flying UAV detection model; is the weight matrix of the regressor in the black flying UAV detection model; is the bias term of the regressor in the black flying UAV detection model;
[0041] Divide the global feature vector set Z of the black flying UAV flight into a training set, a test set, and a validation set according to an empirical ratio, and substitute the training set, the test set, and the validation set into the black flying UAV detection model for model training to obtain a trained black flying UAV detection model; the trained black flying UAV detection model takes the global feature vector of the UAV flight image obtained by processing the UAV flight image through steps 2 to 5 as input and outputs a UAV detection image marked with the category and position of the UAV;
[0042] Step 8: When it is necessary to detect black flying UAVs flying around power facilities, take real-time detection UAV flight images of the UAVs around the power facilities in real time; process the real-time detection UAV flight images through steps 2 to 5 to obtain a global feature vector of the real-time detection UAV flight image, and input the global feature vector of the real-time detection UAV flight image into the trained black flying UAV detection model to obtain a real-time UAV detection image marked with the category and position of the UAV, and the black flying UAVs flying around the power facilities can be detected according to the real-time UAV detection image.
[0043] Furthermore, the specific formula for the first feature extraction in step 2 is as shown in the following formula (12),
[0044] (12)
[0045] In formula (12), is the convolution weight corresponding to the sampling coordinate on the -th simulated black flying UAV flight image among the N preprocessed simulated black flying UAV flight images; is the coordinate on the -th simulated black flying UAV flight image among the N preprocessed simulated black flying UAV flight images The value after offset of the c-th convolution kernel input to the deformable convolution through the sampling coordinates ; , and are the padding value, convolution kernel size, and convolution stride of the deformable convolution respectively.
[0046] Furthermore, the specific process of model training in step 7 is as follows:
[0047] Perform forward propagation and backward propagation on the black flying UAV detection model through the following formula (13), continuously optimize the model parameters, so that the model gradually learns the mapping relationship between the input features and the target output;
[0048] (13)
[0049] In formula (13), M represents the total number of detection targets, and K represents the total number of categories; is the classification loss, used to measure the difference between the predicted category and the true category ; is the regression loss, used to measure the difference between the predicted position and the true position ; represents the weighted sum of the classification loss and the regression loss, and are weight coefficients, used to balance the importance of the classification and regression tasks; is the classification weight of the black flying UAV detection model, is the regression weight in the black flying UAV detection model, and are the gradients of the model parameters calculated according to the total loss through the chain rule, representing the direction and magnitude of weight update; t is the current iteration number, is the classification weight of the current iteration, is the classification weight of the next iteration, is the regression weight of the current iteration, is the regression weight of the next iteration, is the learning rate;
[0050] Execute the forward propagation and backward propagation steps on the black flying UAV detection model in a loop until the loss function converges or reaches the maximum number of iterations.
[0051] Advantages of the present invention: The detection method of the black-flying UAV of the present invention extracts high, medium, and low-level features. The fusion direction adopts unidirectional fusion here, that is, simple operations (such as splicing, upsampling, convolution, etc.) are used, without additional complex weight learning or attention mechanism. Compared with bidirectional fusion, it reduces the computational amount and memory requirements, and also avoids the problem of information redundancy that may be introduced in the bidirectional transmission of features. It is easier to converge during the training process and reduces the risk of network instability. Using a simple fusion strategy of channel splicing and compression, the operation has a low computational amount and does not involve complex feature interaction or additional attention mechanism, so it is more efficient. The splicing operation directly combines multi-scale features together, retains the original information of each feature map, and avoids feature loss. Using the fused features to train the black-flying UAV detection model greatly improves the detection accuracy and speed of the black-flying UAV detection model. Description of the Drawings
[0052] Figure 1 is a flowchart of a detection method for a black-flying UAV of the present invention. Detailed Embodiments
[0053] The following further describes a detection method for a black-flying UAV of the present invention in conjunction with the drawings and specific embodiments
[0054] Embodiment
[0055] The detection method for the black-flying UAV in this example, as Figure 1 shown, includes the following steps:
[0056] Step 1: Use a legal UAV to simulate the flight situation of a black-flying UAV around power facilities, and capture and collect N simulated black-flying UAV flight images; preprocess the N simulated black-flying UAV flight images, and collect the preprocessed N simulated black-flying UAV flight images to form a black-flying UAV flight image dataset , as shown in the following formula (1),
[0057] (1)
[0058] In formula (1), are respectively the first simulated black-flying UAV flight image pixel map, the second simulated black-flying UAV flight image pixel map to the Nth simulated black-flying UAV flight image pixel map in the preprocessed N simulated black-flying UAV flight images; are all pixel maps with a height of H and a width of W; to are respectively the coordinates in to the coordinates on the pixel values; to respectively are the coordinates in to the coordinates the pixel values on; to respectively are the coordinates in to the coordinates the pixel values on; to respectively are the coordinates in to the coordinates the pixel values on; to respectively are the coordinates in to the coordinates the pixel values on; to respectively are the coordinates in to the coordinates the pixel values on;
[0059] Step 2: Through the deformable convolution with the number of convolution kernels being perform the first feature extraction on the with height H and width W in sequence, and obtain N first feature maps of the first simulated black fly drone flight images with height and width and the number of channels being ; The specific formula for the first feature extraction is shown in the following formula (12), to the Nth first feature map of the Nth simulated black fly drone flight image ;
[0060]
[0061] In formula (12), is the convolution weight corresponding to the sampling coordinate on the th simulated black fly drone flight image among the N preprocessed simulated black fly drone flight images; is the coordinate on the th simulated black fly drone flight image among the N preprocessed simulated black fly drone flight images, and the value after offset of the input to the cth convolution kernel of the deformable convolution through the sampling coordinate ; , and They are the padding value, convolution kernel size, and convolution stride of the deformable convolution, respectively.
[0062] , to As shown in Equation (2) below,
[0063] (2)
[0064] In Equation (2), are respectively the first sub-feature maps of the first channel of the first simulated black flight UAV flight image obtained by feature extraction through deformable convolution with the number of convolution kernels being from the first sub-feature map of the second channel of the first simulated black flight UAV flight image to the first sub-feature maps of the th channel of the first simulated black flight UAV flight image; are respectively the first sub-feature maps of the first channel of the second simulated black flight UAV flight image obtained by feature extraction through deformable convolution with the number of convolution kernels being from the first sub-feature map of the second channel of the second simulated black flight UAV flight image to the first sub-feature maps of the th channel of the second simulated black flight UAV flight image; are respectively the first sub-feature maps of the first channel of the Nth simulated black flight UAV flight image obtained by feature extraction through deformable convolution with the number of convolution kernels being from the first sub-feature map of the second channel of the Nth simulated black flight UAV flight image to the first sub-feature maps of the th channel of the Nth simulated black flight UAV flight image;
[0065] Step 3: Perform second feature extraction on the first feature map of the first simulated black flight UAV flight image to obtain the second feature map of the first simulated black flight UAV flight image ; The specific content of performing second feature extraction on the first feature map of the first simulated black flight UAV flight image is as follows:
[0066] Step 3.1: The first feature map of the first simulated black flight UAV flight image is evenly divided into G first sub-feature maps of the first simulated black flight UAV flight image with a height of , a width of , and a channel number of , the second sub-feature map of the first simulated black flight UAV flight image and , … to the G-th sub-feature map of the first simulated illegal drone flight image , , , … to As shown in the following formula (3), (3)
[0067] Substitute , , … to into the following formula (4) in turn to filter and assign features, and obtain the first filtered sub-feature map of the first simulated illegal drone flight image , the second filtered sub-feature map of the first simulated illegal drone flight image , … to the G-th filtered sub-feature map of the first simulated illegal drone flight image ,
[0068] (4)
[0069] In formula (4), is a non-linear activation function;
[0070] Step 3.2: Perform horizontal and vertical pooling on the first filtered sub-feature map of the first simulated illegal drone flight image ,
[0071] (5)
[0072] In formula (5), and are respectively the first filtered horizontal sub-feature map and the first filtered vertical sub-feature map of the first simulated illegal drone flight image obtained through horizontal and vertical pooling; The height of remains unchanged as , the width is pooled into a single value, and the number of channels remains unchanged as ; The height of is pooled into a single value, the width remains unchanged as , and the number of channels remains unchanged as ; is the feature at the coordinate
[0073] on the c-th channel in and are spliced, channel-compressed and fused through the following formula (6) to form the first filtered spliced sub-feature map of the first simulated illegal drone flight image ,
[0074] (6)
[0075] In formula (6), is the first weight matrix; is an operation of splicing multiple feature maps along a specified dimension, that is, splicing two feature maps , along the channel dimension to form a new feature map;
[0076] The first filtered spliced sub-feature map has a dimension of , and the number of channels is ;
[0077] By the following formula (7), weight assignment, normalization and summarization are performed on each channel in to obtain the second feature map of the first simulated black flight drone flight image ,
[0078] (7)
[0079] In formula (7), is to stack the two-dimensional feature maps of all channels in each subgroup in a new dimension to form a complete three-dimensional feature; are respectively the first channel, the second channel to the channel in the first filtered spliced sub-feature map after weight assignment and normalization to obtain the first filtered spliced first channel sub-feature map, the first filtered spliced second channel sub-feature map to the first filtered spliced channel sub-feature map; c is the number of channels; is the first filtered spliced c-channel weighted sub-feature map of the first simulated black flight drone flight image obtained by weight assignment of the c-channel in the first filtered spliced sub-feature map ; and are respectively the mean and standard deviation of the c-channel in the first filtered spliced sub-feature map ; is the stability factor, which is an empirical constant; is the eigenvalue of the c-channel in is the first simulated black flight drone flight image's c-channel weight matrix; is the second filtered spliced sub-feature map of the first simulated black flight drone flight image obtained by the second convolution; is the second weight matrix; is the convolution bias term;
[0080] Step 4: Repeat the content of Step 3 to perform second feature extraction on the second feature map of the flight image of the second simulated black flight drone , the first feature map of the flight image of the third simulated black flight drone , … to the first feature map of the flight image of the Nth simulated black flight drone to obtain the second feature map of the flight image of the first simulated black flight drone , the second feature map of the flight image of the third simulated black flight drone to the second feature map of the flight image of the Nth simulated black flight drone ;
[0081] Perform first feature fusion on , , … to to successively form the first fused feature map of the flight image of the first simulated black flight drone , the first fused feature map of the flight image of the second simulated black flight drone to the first fused feature map of the flight image of the Nth simulated black flight drone , (8);
[0082] Step 5: Perform second feature fusion on the first fused feature map of the flight image of the first simulated black flight drone to form the second fused feature map of the flight image of the first simulated black flight drone ,
[0083] (9)
[0084] In formula (9), , , are learnable weights automatically obtained through training, represents channel-wise weighting, , , are the convolutional kernel weights for extracting high-level, middle-level, and low-level feature maps respectively, , , are the bias terms, is a convolution with a stride of 2, is an upsampling operation, is a concatenation operation; , , are the low-level, middle-level, and high-level feature maps generated by performing a convolution operation on respectively; is the fused middle feature map The features obtained by upsampling, are the feature maps obtained by upsampling ; are the features obtained by concatenating the upsampled features with the low-level features ; are the features obtained by concatenating the upsampled high-level features with the intermediate-level features ; , , are the low-resolution fusion features, medium-resolution fusion features, and high-resolution features obtained by channel compression, respectively;
[0085] The height of the first simulated black flight UAV flight image second fusion feature map is , the width is , and the number of channels is ;
[0086] Each channel of is globally averaged pooled and the results are concatenated to form the first simulated black flight UAV flight image global feature vector ,
[0087] (10)
[0088] In formula (10), are respectively The feature values after global average pooling of the sub-feature maps of the first channel, the second channel to the channel in;
[0089] Step 6: Repeat the content of Step 5 to process the second simulated black flight UAV flight image first fusion feature map to the Nth simulated black flight UAV flight image first fusion feature map to obtain the second simulated black flight UAV flight image global feature vector to the Nth simulated black flight UAV flight image global feature vector , and collect the first simulated black flight UAV flight image global feature vector , the second simulated black flight UAV flight image global feature vector to the Nth simulated black flight UAV flight image global feature vector to form the black flight UAV flight global feature vector set Z;
[0090] Step 7: Establish a black flight UAV detection model, as shown in the following formula (11),
[0091] (11)
[0092] In formula (11), is the detection result of the unlicensed drone detection model; is the detection category of the m-th object among a total of K detection objects; is the detection position of the m-th object among a total of K detection objects; is the maximum probability selection function; is the normalization function; is the classification weight of the unlicensed drone detection model; is the global flight feature vector of the unlicensed drone input into the unlicensed drone detection model; is the weight vector of the fully connected layer in the unlicensed drone detection model; is the bias term of the fully connected layer in the unlicensed drone detection model; is the weight matrix of the regressor in the unlicensed drone detection model; is the bias term of the regressor in the unlicensed drone detection model;
[0093] The global flight feature vector set Z of the unlicensed drone is divided into a training set, a test set, and a validation set according to an empirical ratio, and the training set, the test set, and the validation set are substituted into the unlicensed drone detection model for model training to obtain a trained unlicensed drone detection model; the trained unlicensed drone detection model takes the global feature vector of the drone flight image obtained by processing the drone flight image through steps 2 to 5 as input and outputs a drone detection image with the category and position of the drone marked;
[0094] The specific process of model training is as follows:
[0095] Perform forward propagation and backward propagation on the unlicensed drone detection model through the following formula (13), continuously optimize the model parameters, so that the model gradually learns the mapping relationship between the input features and the target output;
[0096] (13)
[0097] In formula (13), M represents the total number of detection targets, and K represents the total number of categories; is the classification loss, used to measure the predicted category and the true category difference; is the regression loss, used to measure the predicted position and the true position difference; represents the weighted sum of the classification loss and the regression loss, and is the weight coefficient, which is used to balance the importance of classification and regression tasks; is the classification weight of the black flight UAV detection model, is the regression weight in the black flight UAV detection model, and are based on the total loss The gradient of the model parameters is calculated through the chain rule, indicating the direction and magnitude of weight update; t is the current iteration number, is the classification weight of the current iteration, is the classification weight of the next iteration, is the regression weight of the current iteration, is the regression weight of the next iteration, is the learning rate;
[0098] The forward propagation and backward propagation steps are repeatedly executed on the black flight UAV detection model until the loss function converges or the maximum number of iterations is reached.
[0099] Step 8: When it is necessary to detect black flight UAVs flying around power facilities, the real-time detection UAV flight images of the UAVs flying around the power facilities are taken in real time; the real-time detection UAV flight images are processed through Steps 2 to 5 to obtain the global feature vectors of the real-time detection UAV flight images. The global feature vectors of the real-time detection UAV flight images are input into the trained black flight UAV detection model, and the real-time UAV detection images marked with the categories and positions of the UAVs can be obtained. According to the real-time UAV detection images, the black flight UAVs flying around the power facilities can be detected.
Claims
1. A detection method for unlicensed drones, characterized in that: Including the following steps: Step 1: Use a legal drone to simulate the flight of an unlicensed drone around power facilities and capture N simulated unlicensed drone flight images; preprocess the N simulated unlicensed drone flight images and collect the preprocessed N simulated unlicensed drone flight images to form an unlicensed drone flight image dataset , as shown in the following formula (1) （1） In formula (1), are respectively the first simulated flight image pixel map, the second simulated flight image pixel map to the Nth simulated flight image pixel map of the N simulated flight images of black flying drones after the preprocessing; the are all pixel maps with a height of H and a width of W; to are respectively the coordinates in to the coordinate on the pixel values; to are respectively the coordinates in to the coordinate on the pixel values; to are respectively the coordinates in to the coordinate on the pixel values; to are respectively the coordinates in to the coordinate on the pixel values; to are respectively the coordinates in to the coordinate on the pixel values; to are respectively the coordinates in to the coordinate on the pixel values; Step 2: Use deformable convolution with the number of convolutional kernels being to perform first feature extraction on the with height H and width W in sequence, obtaining N first feature maps of the first simulated black fly drone flight images with height , width , and number of channels , the first feature map of the second simulated black fly drone flight image , the first feature map of the third simulated black fly drone flight image to the first feature map of the Nth simulated black fly drone flight image ; the , to are as shown in the following formula (2). （2） In formula (2), are respectively the first sub-feature map of the first channel of the first simulated black flying UAV flight image, the first sub-feature map of the second channel of the first simulated black flying UAV flight image to the first sub-feature map of the th channel of the first simulated black flying UAV flight image, which are obtained by performing feature extraction on the deformable convolution with the number of convolution kernels being are respectively the first sub-feature map of the first channel of the second simulated black flying UAV flight image, the first sub-feature map of the second channel of the second simulated black flying UAV flight image to the first sub-feature map of the th channel of the second simulated black flying UAV flight image, which are obtained by performing feature extraction on the deformable convolution with the number of convolution kernels being are respectively the first sub-feature map of the first channel of the Nth simulated black flying UAV flight image, the first sub-feature map of the second channel of the Nth simulated black flying UAV flight image to the first sub-feature map of the th channel of the Nth simulated black flying UAV flight image, which are obtained by performing feature extraction on the deformable convolution with the number of convolution kernels being Step 3: Perform second feature extraction on the first feature map of the first simulated black flight UAV flight image to obtain the second feature map of the first simulated black flight UAV flight image ; The specific content of performing second feature extraction on the first feature map of the first simulated black flight UAV flight image is as follows: Step 3.1: The first feature map of the first simulated unauthorized drone flight image According to the number of channels It is evenly divided into G first sub-feature maps of the first simulated unauthorized drone flight image with a height of and a width of and the number of channels is , the second sub-feature map of the first simulated unauthorized drone flight image , the... to the G-th sub-feature map of the first simulated unauthorized drone flight image , , ... to are as shown in the following formula (3), (3) Substitute the , , … to into the following formula (4) in turn to filter the assignment features, and obtain the first filtered sub-feature map of the first simulated black flight UAV flight image , the second filtered sub-feature map of the first simulated black flight UAV flight image , … to the Gth filtered sub-feature map of the first simulated black flight UAV flight image . （4） In formula (4), is a non-linear activation function; Step 3.2: Perform horizontal and vertical pooling on the first filtered sub-feature map of the first simulated unauthorized drone flight image through the following formula (5) to obtain （5） In formula (5), and are respectively the first filtered horizontal sub-feature map and the first filtered vertical sub-feature map of the first simulated flight image of the black flying drone obtained by pooling in the horizontal and vertical directions; the has a constant height of , the width is pooled into a single value, and the number of channels remains unchanged at ; the has a height pooled into a single value, a constant width of , and the number of channels remains unchanged at ; the is the feature at the coordinate on the c-th channel in the ; ; The and are spliced, channel-compressed and fused to form a first filtered spliced sub-feature map of the first simulated black flight drone flight image , （6） In formula (6), is the first weight matrix; is an operation of concatenating multiple feature maps along a specified dimension, that is, concatenating two feature maps , along the channel dimension to form a new feature map; The first filtered and spliced sub-feature map has a dimension of , and a number of channels of ; Perform weight assignment and normalization on each channel in the following formula (7) and sum them up to obtain the second feature map of the first simulated flight image of the black-flying drone , （7） In formula (7), is to stack the two-dimensional feature maps of all channels in each subgroup in a new dimension to form a complete three-dimensional feature; are respectively the first filtered and spliced sub-feature maps in the first channel, the second channel to the channel after weight assignment and normalization to obtain the first filtered and spliced first channel sub-feature map, the first filtered and spliced second channel sub-feature map to the first filtered and spliced channel sub-feature map; c is the number of channels; is the first filtered and spliced c-channel weight sub-feature map of the first simulated black flight UAV flight image, which is the c-channel of the first filtered and spliced sub-feature map after weight assignment; and are respectively the mean and standard deviation of the c-channel in the first filtered and spliced sub-feature map ; is the stability factor, which is an empirical constant; is the feature value of the c-channel in; is the c-channel weight matrix of the first simulated black flight UAV flight image; is the second filtered and spliced sub-feature map of the first simulated black flight UAV flight image obtained by second convolution; is the second weight matrix; is the convolution bias term; Step 4: Repeat the content of Step 3 to perform second feature extraction on the second feature map of the flight image of the second simulated black-flying drone , the first feature map of the flight image of the third simulated black-flying drone , … to the first feature map of the flight image of the Nth simulated black-flying drone to successively obtain the second feature map of the flight image of the first simulated black-flying drone , the second feature map of the flight image of the third simulated black-flying drone to the second feature map of the flight image of the Nth simulated black-flying drone ; Perform first feature fusion on the , , … to to successively form the first fusion feature map of the first simulated black flight UAV flight image , the first fusion feature map of the second simulated black flight UAV flight image to the first fusion feature map of the Nth simulated black flight UAV flight image , (8); Step 5: Perform second feature fusion on the first fusion feature map of the first simulated black flight UAV flight image through the following formula (9) to form the second fusion feature map of the first simulated black flight UAV flight image , （9） In formula (9), , , are learnable weights automatically obtained through training, represents channel-wise weighting, , , are the convolutional kernel weights for extracting high-level, middle-level, and low-level feature maps respectively, , , are bias terms, is a convolution with a stride of 2, is an upsampling operation, is a concatenation operation; , , are the low-level, middle-level, and high-level feature maps generated by performing a convolution operation on respectively; is the feature obtained by upsampling the fused middle feature map , is the feature map obtained by upsampling ; is the feature obtained by concatenating the upsampled feature with the low-level feature , is the feature obtained by concatenating the upsampled high-level feature with the middle-level feature ; , , are the low-resolution fused feature, medium-resolution fused feature, and high-resolution feature obtained after channel compression respectively; The height of the first fused feature map of the simulated drone flight images is , the width is , and the number of channels is ; Perform global average pooling on each channel of the following formula (10) and splice the results to form a first global feature vector of the simulated black flight drone flight image and , （10） In formula (10), are respectively the eigenvalues after global average pooling of the sub-feature maps of the first channel, the second channel to the channel in the Step 6: Repeat the content of Step 5 to process the first fusion feature map of the second simulated black flight UAV flight image to the first fusion feature map of the Nth simulated black flight UAV flight image to obtain the global feature vector of the second simulated black flight UAV flight image to the global feature vector of the Nth simulated black flight UAV flight image , and collect the global feature vector of the first simulated black flight UAV flight image , the global feature vector of the second simulated black flight UAV flight image to the global feature vector of the Nth simulated black flight UAV flight image to form a set Z of global feature vectors of black flight UAV flights; Step 7: Establish a black-flying UAV detection model as shown in the following formula (11). （11） In formula (11), is the detection result of the black flight UAV detection model; is the detection category of the m-th object among a total of K detection objects; is the detection position of the m-th object among a total of K detection objects; is the maximum probability selection function; is the normalization function; is the classification weight of the black flight UAV detection model; is the global flight feature vector of the black flight UAV input into the black flight UAV detection model; is the weight vector of the fully connected layer in the black flight UAV detection model; is the bias term of the fully connected layer in the black flight UAV detection model; is the weight matrix of the regressor in the black flight UAV detection model; is the bias term of the regressor in the black flight UAV detection model; Divide the global feature vector set Z of the black-flying UAV flight into a training set, a test set, and a validation set according to an empirical ratio, and substitute the training set, the test set, and the validation set into the black-flying UAV detection model for model training to obtain a trained black-flying UAV detection model; the trained black-flying UAV detection model takes the global feature vector of the UAV flight image obtained by processing the UAV flight image through Steps 2 to 5 as input and takes the UAV detection image marked with the category and position of the UAV as output. Step 8: When it is necessary to detect black-flying UAVs flying around power facilities, take real-time detection UAV flight images of the UAVs around the power facilities in real time; process the real-time detection UAV flight images through Steps 2 to 5 to obtain the global feature vector of the real-time detection UAV flight images, and input the global feature vector of the real-time detection UAV flight images into the trained black-flying UAV detection model to obtain a real-time UAV detection image marked with the category and position of the UAV, and the black-flying UAVs flying around the power facilities can be detected according to the real-time UAV detection image.
2. The detection method of the unlicensed drone according to claim 1, characterized in that: The specific formula for the first feature extraction in Step 2 is as shown in the following formula (12). （12） In formula (12), is the convolution weight corresponding to the sampling coordinates on the th simulated flight image of black - flying drones among the N pre - processed simulated flight images of black - flying drones; is the coordinate on the th simulated flight image of black - flying drones among the N pre - processed simulated flight images of black - flying drones, and the value after offset of the c - th convolution kernel input to the deformable convolution through the sampling coordinates ; , and are respectively the padding value, the convolution kernel size, and the convolution stride of the deformable convolution.
3. The detection method of the unlicensed drone according to claim 1 or 2, characterized in that: The specific process of model training in Step 7 is as follows: Perform forward propagation and backward propagation on the black-flying UAV detection model through the following formula (13), and continuously optimize the model parameters, so that the model gradually learns the mapping relationship between the input features and the target output. （13） In Equation (13), M represents the total number of detection targets, and K represents the total number of categories; is the classification loss, which is used to measure the predicted category and the true category difference; is the regression loss, which is used to measure the predicted position and the true position difference; represents the weighted sum of the classification loss and the regression loss, and are weight coefficients, which are used to balance the importance of the classification and regression tasks; is the classification weight of the black-flying UAV detection model, is the regression weight in the black-flying UAV detection model, and are the gradients of the model parameters calculated according to the total loss through the chain rule, indicating the direction and magnitude of the weight update; t is the current iteration number, is the classification weight for the current iteration, is the classification weight for the next iteration, is the regression weight for the current iteration, is the regression weight for the next iteration, is the learning rate; Loop through the forward propagation and backward propagation steps for the black-flying UAV detection model until the loss function converges or reaches the maximum number of iterations.