Low-altitude unmanned aerial vehicle identification and countermeasure method and system based on visual fusion

Abstract

The application discloses a low-altitude unmanned aerial vehicle identification and countermeasure method and system based on visual fusion, and relates to the technical field of unmanned aerial vehicle identification and countermeasure. The application comprises the following steps: setting a multi-module visual module, wherein the multi-module visual module comprises an RGB module, a depth module and an infrared module; based on the multi-module visual module, fusing multi-modal visual features, pre-processing acquired multi-modal data, extracting initial features, and normalizing the modal feature set through a processing method. The application extracts multi-modal features, sets initial coefficients according to the extracted multi-modal features, weights and fuses the extracted data, acquires identification feedback data through a classification probability and a conversion function, sets a feedback threshold, and countermeasures through a target countermeasure method when the feedback threshold is reached, so that the unmanned aerial vehicle is identified by multiple modules, the identification accuracy of the unmanned aerial vehicle is improved, and the effect of multiple modules working together to identify the unmanned aerial vehicle is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle identification countermeasures, in particular to a low-altitude unmanned aerial vehicle identification and countermeasure method and system based on visual fusion. BACKGROUND

[0002] Low-altitude unmanned aerial vehicles are small unmanned aerial vehicles that fly at low altitudes and are widely used in civilian, commercial and military fields. They perform tasks through remote control or autonomous navigation systems, have the characteristics of flexibility, efficiency and low cost, and are suitable for close-range operations. Small size and weight, small size design, easy to carry and deploy, mainly used for farmland monitoring, pesticide spraying, medical material transportation, disaster monitoring, search and rescue operations, traffic management, aerial photography, reconnaissance and the like. The identification and countermeasures of low-altitude unmanned aerial vehicles refer to detecting and identifying unmanned aerial vehicles flying at low altitudes through technical means, and taking appropriate measures to interfere with, control or shoot down them to prevent potential safety hazards or threats.

[0003] The patent with publication number CN118799390A discloses a method for identifying key nodes of a group of unmanned aerial vehicles based on a spatio-temporal graph convolution, a medium and equipment. First, a group of unmanned aerial vehicle target detection algorithm is used to detect the positions of each unmanned aerial vehicle in the group; then a group of unmanned aerial vehicle target tracking algorithm is used to extract the motion trajectories of each unmanned aerial vehicle in the group; then a group of unmanned aerial vehicle spatio-temporal graph model is constructed, and a spatio-temporal graph convolution network is used to extract the deep spatio-temporal features of each member unmanned aerial vehicle; finally, a fully connected layer is used for graph node classification, thereby realizing the identification of key nodes of the group of unmanned aerial vehicles. Compared with traditional target identification methods, the present application can model the spatio-temporal information of the target simultaneously, and learn a classification function that is more suitable for the actual task using a data-driven approach. It has the advantages of strong operability, high identification accuracy and strong adaptability, and can provide more reliable information support for subsequent countermeasures, which is of great significance for the countermeasures against the group of unmanned aerial vehicles.

[0004] The above-mentioned and similar technical solutions have the problem that when identifying unmanned aerial vehicles, conventional methods rely on a single visual feature, such as analyzing the acquired images. However, in complex scenarios, with the increase of interference factors such as insufficient light or animal interference, the identification accuracy of the unmanned aerial vehicles is low, which cannot guarantee the efficiency of subsequent countermeasures. SUMMARY

[0005] The present application aims to provide a low-altitude unmanned aerial vehicle identification and countermeasure method and system based on visual fusion to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a low-altitude unmanned aerial vehicle identification and countermeasure method based on visual fusion, comprising:

[0007] Set a multi-module vision module, the multi-module vision module includes an RGB module, a depth module and an infrared module;

[0008] Based on the multi-module vision module, the multi-modal visual features are fused, based on the multi-module vision module, the obtained multi-modal data is preprocessed, and preliminary features are extracted, the modal feature set is normalized by a processing method, and a normalized data set is obtained;

[0009] Based on the normalized data set, multi-modal feature extraction is performed by an extraction method, and an extracted feature set is obtained;

[0010] Set an initial weight coefficient to obtain an initial coefficient term, based on the extracted feature set, the extracted feature set is weighted and fused by the initial coefficient term to obtain a multi-modal fusion feature;

[0011] Based on the multi-modal fusion feature, information recognition is performed by a classification probability acquisition method, the classification probability acquisition method includes a classification probability acquisition formula:

[0012] ;

[0013] Wherein is a conversion function, is a transposed weight matrix, is a bias term, is a multi-modal fusion feature, is a classification probability;

[0014] Set a target output category, and obtain a target result term based on the classification probability calculation result;

[0015] Based on the target result term, recognition feedback data is obtained, a feedback threshold is set, and it is judged whether the recognition feedback data reaches the feedback threshold, when the feedback threshold is reached, countermeasures are taken by a target countermeasure method;

[0016] The initial coefficient term is adjusted by an adjustment method, and then the multi-modal fusion feature is updated to obtain an updated feature term, the classification probability is reacquired based on the updated feature term, and the updated feedback data is obtained, and the specific recognition effect is performed under different conditions.

[0017] The processing method includes:

[0018] RGB image preprocessing:

[0019] ;

[0020] Wherein is an original RGB image, and are the mean and standard deviation of the RGB image respectively;

[0021] Depth image preprocessing:

[0022] ;

[0023] wherein is the original depth image, and are the mean and standard deviation of the depth image, respectively;

[0024] Infrared image preprocessing:

[0025] ;

[0026] wherein is the original infrared image, and are the mean and standard deviation of the infrared image, respectively.

[0027] The extraction method includes a convolutional neural network extraction, and the extraction method includes:

[0028] A convolution operation formula is set, and the convolution operation formula is:

[0029] ;

[0030] wherein is the normalized data, is the convolution kernel weight, is the bias term, is the output feature map;

[0031] Based on the output feature map, a ReLU activation function is used to obtain an activation information item, maximum pooling is performed based on the activation information item, and full connection is performed to obtain a full connection data layer. The full connection data layer is normalized to obtain an extraction feature set.

[0032] The extraction feature set includes RGB features, depth features, and infrared features, and the acquisition method of the multi-modal fusion features includes:

[0033] A multi-modal fusion feature formula is set:

[0034] ;

[0035] wherein , , are initial coefficient items, , , are the RGB features, depth features, and infrared features, respectively.

[0036] The setting method of the feedback threshold includes:

[0037] Set the first feedback value and the second feedback value, the first feedback value and the second feedback value are fixed values, based on the target output category, obtain at least one category first value and category second value;

[0038] Set the target identification item, the target identification item is included in the target output category, based on the combination result of the target identification item and the first feedback value and the second feedback value, obtain the feedback threshold value.

[0039] The target countermeasure method comprises:

[0040] Set the countermeasure combination module, the countermeasure combination module includes the interference module and the electronic module;

[0041] Based on the identification feedback data, obtain the distance data, obtain the countermeasure distance item, set the countermeasure interval, the countermeasure interval corresponds to the interference module and the electronic module respectively, based on the countermeasure distance item, select the target module as the selected countermeasure module.

[0042] Further, the adjustment method comprises:

[0043] Obtain the light intensity data through the target device, obtain the real-time intensity item;

[0044] Divide the light intensity into at least two division range items, the division range item corresponds to different light intensity range respectively;

[0045] Set the main module and the secondary module, and set the main weight proportion value, the main weight proportion value corresponds to the division range item respectively;

[0046] Determine the comparison result of the real-time intensity item and the division range item, obtain the proportion value of the main module, and then obtain the target proportion value, the secondary module allocates the remaining proportion value, and then adjusts the initial coefficient item.

[0047] Further, a low-altitude unmanned aerial vehicle identification and countermeasure system based on visual fusion uses the low-altitude unmanned aerial vehicle identification and countermeasure method based on visual fusion, comprising:

[0048] Composition module: set the multi-module visual module, the multi-module visual module includes RGB module, depth module and infrared module;

[0049] Processing module: based on the multi-module visual module, fuse multi-modal visual features, based on the multi-module visual module, preprocess the obtained multi-modal data, extract preliminary features, normalize the modal feature set through the processing method, and obtain the normalized data set;

[0050] The identification module: based on the normalized data set, the multi-modal feature extraction is carried out through the extraction method, and the extraction feature set is obtained; the initial weight coefficient is set, and the initial coefficient term is obtained; based on the extraction feature set, the extraction feature set is weighted and fused through the initial coefficient term, and the multi-modal fusion feature is obtained; based on the multi-modal fusion feature, information recognition is carried out through the classification probability acquisition method;

[0051] The countermeasure module: the target output category is set, the target result term is obtained based on the classification probability calculation result; the identification feedback data is obtained based on the target result term, the feedback threshold is set, and it is judged whether the identification feedback data reaches the feedback threshold; when the feedback threshold is reached, the countermeasure is carried out through the target countermeasure method;

[0052] The adjustment module: the initial coefficient term is adjusted through the adjustment method, and then the multi-modal fusion feature is updated to obtain the updated feature term; the classification probability is re-acquired based on the updated feature term, and the updated feedback data is obtained; the targeted identification effect is carried out under different conditions.

[0053] Compared with the prior art, the beneficial effects of the present application are:

[0054] The low-altitude unmanned aerial vehicle identification and countermeasure method and system based on visual fusion, through the RGB image module, the depth image module and the infrared image module to form a multi-module visual module, through the preprocessing of the obtained multi-modal data, the preliminary features are extracted, and the normalization processing is carried out, and the multi-modal features are extracted, according to the extracted multi-modal features, the initial coefficient is set, the extracted data is weighted and fused, and the multi-modal fusion feature is obtained, then through the acquisition of the classification probability, the identification feedback data is obtained with the conversion function, and the feedback threshold is set, when the feedback threshold is reached, the countermeasure is carried out through the target countermeasure method, so as to realize the effect of multi-module cooperation for identifying unmanned aerial vehicle, and improve the identification accuracy of unmanned aerial vehicle.

[0055] At the same time, by dividing the range of light intensity, setting the main module and the secondary module, and setting the main weight proportion value, the main weight proportion value corresponds to the divided range respectively, by judging the comparison result of the real-time intensity and the divided range, the proportion value of the main module is obtained, and then the target proportion value is obtained, the secondary module is evenly distributed to the remaining proportion value, and then the initial coefficient term is adjusted, the weight coefficient of different modules is increased or decreased under different conditions, so that the identification of unmanned aerial vehicle is more targeted. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 It is the overall flowchart of the present application;

[0057] Figure 2 It is the normalized data set acquisition flowchart of the present application;

[0058] Figure 3 A schematic diagram of a multi-modal data composition of the present application;

[0059] Figure 4 A schematic diagram of a multi-modal fusion feature acquisition process of the present application;

[0060] Figure 5 A schematic diagram of an extraction method process of the present application. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0062] Conventional unmanned aerial vehicle recognition methods usually rely on analyzing acquired RGB images. These methods aim to extract visual features in the images, such as the shape, texture, color, etc. of the unmanned aerial vehicle. Through training a machine learning model, such as a convolutional neural network, automatic recognition of the unmanned aerial vehicle can be achieved. However, this single visual feature-dependent recognition method has inherent vulnerability. In an ideal experimental environment, such as sufficient light, simple background, and obvious unmanned aerial vehicle features, this kind of method can often achieve high recognition accuracy. However, in real-world applications, the scene is often much more complex than the experimental environment, and various interference factors will significantly reduce the recognition effect. Light conditions are a key factor affecting the quality of RGB images. In environments with insufficient light, such as at night or on cloudy days, the signal-to-noise ratio of the image will be greatly reduced, resulting in image blurring, detail loss, and even color distortion. These problems will directly affect the extraction of visual features, making it difficult for recognition algorithms based on RGB images to accurately determine whether an unmanned aerial vehicle exists or to incorrectly identify other objects as unmanned aerial vehicles. For example, at night, the lights of urban landscapes can form patterns similar to unmanned aerial vehicles, causing the algorithm to misjudge. In addition to light conditions, animal interference is also an important factor that cannot be ignored. In some scenarios, such as airports, wildlife reserves, etc., birds or other animals can appear in the image at the same time as the unmanned aerial vehicle. The shape, size, and movement pattern of these animals can be similar to those of the unmanned aerial vehicle, causing visual confusion and thus reducing recognition accuracy.

[0063] The technical scheme provided in the application comprises a multi-module vision module composed of an RGB image module, a depth image module and an infrared image module, preliminary features are extracted and normalized by preprocessing the obtained multi-modal data, multi-modal features are extracted, initial coefficients are set according to the extracted multi-modal features, the extracted data is weighted and fused to obtain multi-modal fusion features, and then recognition feedback data is obtained by means of classification probability and a conversion function, and a feedback threshold is set, and countermeasures are taken by means of target countermeasures when the feedback threshold is reached, so as to realize the effect of multi-module cooperation in recognizing unmanned aerial vehicles, improve the recognition accuracy of unmanned aerial vehicles, and specifically, as shown in Figure 1 , the steps S100-S700 are included.

[0064] Step S100: A multi-module vision module is set, and the multi-module vision module comprises an RGB module, a depth module and an infrared module.

[0065] It should be noted that the multi-module vision module is composed of at least three parts, namely an RGB image module, a depth image module and an infrared image module, and the multi-mode vision module is obtained by combining the multiple modules, so that the best data acquisition effect is realized by adjusting the mutual cooperation of the multiple modules in visual acquisition.

[0066] Step S200: As shown in Figure 2 , multi-modal vision features are fused based on the multi-module vision module, the obtained multi-modal data is preprocessed based on the multi-module vision module to extract preliminary features, and the modal feature set is normalized by a processing method to obtain a normalized data set.

[0067] It should be noted that, as shown in Figure 3 , when the multi-modal data is preprocessed, the multi-modal data is composed of an RGB image module, a depth image module and an infrared image module, so the data of the three modules needs to be analyzed and processed respectively to extract preliminary features, and the three modules are normalized by a corresponding processing method to obtain a normalized data set composed of three normalized data groups, and the processing method comprises:

[0068] RGB image preprocessing:

[0069]

[0070] wherein is an original RGB image, and are the mean and standard deviation of the RGB image respectively;

[0071] Depth image preprocessing:

[0072] ​ ;

[0073] in This is the original depth map. and These are the mean and standard deviation of the depth map, respectively.

[0074] Infrared image preprocessing:

[0075] ;

[0076] in The original infrared image, and These represent the mean and standard deviation of the infrared image, respectively.

[0077] Example 1

[0078] In the specific implementation process, when the acquired RGB image is [100, 150, 200], the depth image data is [5, 10, 15], and the infrared image is [30, 50, 70], the mean and standard deviation of the RGB image can be obtained according to the RGB image preprocessing formula, the depth image preprocessing formula, and the infrared image preprocessing formula. and The mean and standard deviation of the depth maps are 150 and 50, respectively. and The mean and standard deviation of the infrared images are 10 and 5, respectively. and The values ​​are 50 and 20 respectively, at which point the RGB image is normalized:

[0079] ;

[0080] Depth image normalization:

[0081] ;

[0082] Infrared image normalization:

[0083] .

[0084] Step S300: Based on the normalized dataset, multimodal feature extraction is performed using extraction methods to obtain the extracted feature set.

[0085] like Figure 5 As shown, it should be noted that the extraction methods include convolutional neural network extraction, and the extraction methods include:

[0086] The convolution operation formula is defined as follows:

[0087] ;

[0088] in The data is after normalization. For convolution kernel weights, For bias terms, Output feature map;

[0089] The convolutional network neural structure consists of three convolutional layers with a stride of 1. The kernel size is set to 3×3, and the output feature dimension is 128×128×32. After the convolution operation, the feature map dimension is 128×128×32. The kernel size is set to a 3*3 matrix, and the bias term... The value is 0.1, and the kernel weights are:

[0090] ;

[0091] Based on the output feature map, the ReLU activation function is used to obtain activation information terms. Max pooling is then performed on the activation information terms, followed by a fully connected layer to obtain a fully connected data layer. The fully connected data layer is then normalized to obtain the extracted feature set.

[0092] Specifically, when using the ReLU activation function, the formula is:

[0093] ;

[0094] Max pooling formula:

[0095] ;

[0096] The pooling kernel size is 2×2, and the step size is 2;

[0097] Fully connected layer formula:

[0098] ;

[0099] in This is the weight matrix of the fully connected layer. For bias terms, For the input feature map, It is a function;

[0100] Normalize the output of the fully connected layer:

[0101] ;

[0102] Given an input feature map dimension of 8192 (8×8×128) and a fully connected layer output dimension of 128, the weight matrix... With dimensions of 8192×128, the extracted feature set is obtained after normalizing the fully connected layer.

[0103] Example 2

[0104] When the normalized RGB image is [-1, 0, 1], the convolution kernel weights are:

[0105] At this point, according to the convolution operation formula, the convolution kernel weights are... Multiply the input image element-wise and sum the results, then add a bias term. The calculation result is: ;

[0106] Using the ReLU activation function:

[0107] ;

[0108] We set up max pooling with a 2×2 kernel size and a stride of 2. Since the input feature map is relatively small, the output after pooling is: ;

[0109] Fully connected layer formula:

[0110] ;

[0111] in This is the weight matrix of the fully connected layer. For bias terms, For the input feature map, For a function, set The bias term is [0.5, 0.6, 0.7]. The input feature is 0.1. The output value of the fully connected layer is 0.7. At this point, according to the normalization formula:

[0112] ;

[0113] ;

[0114] Step S400: As Figure 4 As shown, initial weight coefficients are set to obtain initial coefficient terms. Based on the extracted feature set, the extracted feature set is weighted and fused using the initial coefficient terms to obtain multimodal fusion features.

[0115] It should be noted that the extracted feature set includes RGB features, depth features, and infrared features. Methods for obtaining multimodal fusion features include:

[0116] Define the multimodal fusion feature formula:

[0117] ;

[0118] in , , For the initial coefficient term, , , The initial coefficients for RGB features, depth features, and infrared features are set to 0.5, 0.3, and 0.2, respectively.

[0119] Example 3

[0120] In the specific implementation process, when the RGB features, depth features, and infrared features are obtained as follows: , , At that time, based on the set initial coefficients and the multimodal fusion feature formula, the calculation process is as follows:

[0121] ;

[0122] The calculation result is:

[0123] .

[0124] Step S500: Based on multimodal fusion features, information is identified using a classification probability acquisition method.

[0125] It should be noted that the methods for obtaining classification probabilities include the formula for obtaining classification probabilities:

[0126] ;

[0127] in For transformation functions, This is the weight matrix after transpose. For bias terms, For multimodal fusion features, For classification probability;

[0128] The transformation functions are the linear calculation function and the Softmax calculation function, respectively.

[0129] Linear computation functions: ;

[0130] Softmax calculation function: .

[0131] Set the target output category, and obtain the target result item based on the classification probability calculation result.

[0132] It should be noted that the output categories are drone, bird, and background.

[0133] Step S600: Based on the target result item, obtain the identification feedback data, set the feedback threshold, determine whether the identification feedback data reaches the feedback threshold, and when the feedback threshold is reached, countermeasures are taken through the target countermeasure method.

[0134] It is important to note that the method for setting the feedback threshold includes: setting a first feedback value and a second feedback value, both of which are fixed values, with the first and second feedback values ​​being 50% and 90% respectively; obtaining at least one first value and one second value for a category based on the target output category; and setting a target identification item, which is included in the target output category. Based on the combination of the target identification item and the first and second feedback values, the feedback threshold is obtained. The set target identification item is the drone, and since the drone is included in the target output category, the feedback threshold is 50% and 90% of the drone probability.

[0135] It should be noted that the target countermeasure method includes: setting up a countermeasure combination module, which includes an interference module and an electronic module; acquiring distance data based on identification feedback data to obtain a countermeasure distance item; setting countermeasure intervals, with each countermeasure interval corresponding to the interference module and the electronic module respectively; there are two countermeasure intervals, one corresponding to the interference module and the other to the electronic module, and the distances of the countermeasure intervals are 0-500m and above 500m respectively; and selecting the target module as the selected countermeasure module based on the countermeasure distance item.

[0136] Step S700: Adjust the initial coefficient terms by adjusting the adjustment method, and then update the multimodal fusion features to obtain the updated feature terms.

[0137] It is important to note that the classification probability is re-acquired based on the updated feature terms to obtain updated feedback data. Targeted recognition is then performed under different conditions. The adjustment methods include: acquiring light intensity data from the target device to obtain a real-time intensity term; dividing the light intensity into ranges to obtain two range terms, each corresponding to a different light intensity range, with the ranges being 0-100 lux and above 100 lux respectively; setting a main module and secondary modules, with the main module being an RGB module and the secondary modules being a depth module and an infrared module, while simultaneously setting a sovereign weight percentage, corresponding to the range terms, with sovereign weights of 60% and 30% respectively; comparing the real-time intensity term with the range terms to obtain the proportion value of the main module, and then obtaining the target proportion value; the remaining proportion values ​​are evenly distributed among the secondary modules, thereby adjusting the initial coefficient terms.

[0138] In the specific implementation process, when the real-time light intensity is 20 lux, the corresponding intensity range is between 0-100 lux, and the corresponding sovereign weight ratio is 30%. At this time, the sub-modules allocate the remaining weight ratio on an average basis, that is, the RGB weight coefficient is 30%, the depth weight coefficient and the infrared weight coefficient are both 35%.

[0139] A low-altitude UAV identification and countermeasure system based on visual fusion utilizes the aforementioned visual fusion-based low-altitude UAV identification and countermeasure method, comprising: a component module: setting up a multi-module visual module, including an RGB module, a depth module, and an infrared module; a processing module: fusing multimodal visual features based on the multi-module visual modules, preprocessing the acquired multimodal data based on the multi-module visual modules, extracting preliminary features, and normalizing the modal feature set using a processing method to obtain a normalized dataset; a recognition module: extracting multimodal features based on the normalized dataset using an extraction method to obtain an extracted feature set; setting initial weight coefficients to obtain initial coefficient terms, and based on the extracted... The feature set is weighted and fused using initial coefficient terms to obtain multimodal fusion features. Based on these features, information recognition is performed using a classification probability acquisition method. The countermeasure module sets a target output category and calculates the target result based on the classification probability. Based on the target result, recognition feedback data is obtained, a feedback threshold is set, and it is determined whether the recognition feedback data reaches the threshold. If the threshold is reached, countermeasures are performed using a target countermeasure method. The adjustment module adjusts the initial coefficient terms using an adjustment method to update the multimodal fusion features, obtaining updated feature terms. Based on these updated feature terms, the classification probability is re-acquired to obtain updated feedback data, enabling targeted recognition under different conditions.

[0140] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A low-altitude unmanned aerial vehicle identification and countermeasure method based on visual fusion, characterized in that, The application relates to a multi-modal feature fusion method and device. The multi-modal visual module comprises an RGB module, a depth module and an infrared module; The multi-modal visual module is used for fusing multi-modal visual features, and the multi-modal visual module is used for normalizing the obtained multi-modal data to obtain a normalized data set; Based on the normalized data set, multi-modal feature extraction is performed through an extraction method to obtain an extracted feature set; An initial weight coefficient is set to obtain an initial coefficient term, and the extracted feature set is weighted and fused through the initial coefficient term to obtain multi-modal fusion features; Based on the multi-modal fusion features, information recognition is performed through a classification probability acquisition method, and the classification probability acquisition method comprises a classification probability acquisition formula: ; wherein is a transformation function, is a transposed weight matrix, is a bias term, is a multi-modal fusion feature, is a classification probability; A target output category is set, and a target result term is obtained based on a classification probability calculation result; Based on the target result term, recognition feedback data is obtained, a feedback threshold is set, and it is judged whether the recognition feedback data reaches the feedback threshold; when the feedback threshold is reached, countermeasures are taken through a target countermeasure method; The initial coefficient term is adjusted through an adjusting method, and the multi-modal fusion features are updated to obtain an updated feature term; the classification probability is reacquired based on the updated feature term to obtain updated feedback data, and targeted recognition is performed under different conditions; The feedback threshold setting method comprises: A first feedback value and a second feedback value are set, the first feedback value and the second feedback value are fixed values, and at least one category first value and a category second value are obtained based on the target output category; A target recognition term is set, the target recognition term is contained in the target output category, and a feedback threshold is obtained based on the combination result of the target recognition term and the first feedback value and the second feedback value; The target countermeasure method comprises: A countermeasure combination module is set, and the countermeasure combination module comprises an interference module and an electronic module; Distance data is acquired based on the recognition feedback data to obtain a countermeasure distance term, a countermeasure interval is set, the countermeasure interval corresponds to the interference module and the electronic module respectively, and a target module is selected as a selected countermeasure module based on the countermeasure distance term; The adjusting method comprises: Light intensity data is acquired through a target device to obtain a real-time intensity term; The light intensity is divided into at least two division range terms, and the division range terms correspond to different light intensity ranges respectively; A main module and a secondary module are set, and a main weight proportion value is set, the main weight proportion value corresponds to the division range term respectively; The proportion value of the main module is obtained by judging the comparison result of the real-time intensity term and the division range term, and then a target proportion value is obtained, the secondary module is evenly distributed with the remaining proportion value, and then the initial coefficient term is adjusted; The extraction method comprises a convolutional neural network extraction, and the extraction method comprises: A convolution operation formula is set, and the convolution operation formula is: ; wherein is the normalized data, is a convolution kernel weight, is a bias term, is an output feature map; Based on the output feature map, an activation information term is obtained by using a ReLU activation function, maximum pooling is performed based on the activation information term, full connection is performed, a full connection data layer is obtained, the full connection data layer is normalized, and the extracted feature set is obtained.

2. The low-altitude unmanned aerial vehicle identification and countermeasure method based on visual fusion according to claim 1, characterized in that: The processing method comprises: RGB image preprocessing; ; wherein is the original RGB image, and are the mean and standard deviation of the RGB image, respectively; Depth image pre-processing: ; wherein is the original depth image, and are the mean and standard deviation of the depth image, respectively; Infrared image preprocessing: ; wherein is the original infrared image, and are the mean and standard deviation of the infrared image, respectively.

3. The low-altitude unmanned aerial vehicle identification and countermeasure method based on visual fusion according to claim 1, characterized in that: The extracted feature set comprises RGB features, depth features and infrared features, and the multi-modal fusion feature acquisition method comprises: A multi-modal fusion feature formula is set: ; wherein , , is an initial coefficient term, , , RGB features, depth features, and infrared features, respectively.

4. A low-altitude unmanned aerial vehicle identification and countermeasure system based on visual fusion, characterized in that: The low-altitude unmanned aerial vehicle identification and countermeasure method based on visual fusion according to any one of claims 1-3 is used, comprising: A component module: a multi-module visual module is set, and the multi-module visual module includes an RGB module, a depth module and an infrared module; A processing module: based on the multi-module visual module, multi-modal visual features are fused, based on the multi-module visual module, the obtained multi-modal data is normalized to obtain a normalized data set; An identification module: based on the normalized data set, multi-modal feature extraction is performed through an extraction method to obtain an extracted feature set; an initial weight coefficient is set to obtain an initial coefficient term, based on the extracted feature set, the extracted feature set is weighted and fused through the initial coefficient term to obtain multi-modal fusion features; based on the multi-modal fusion features, information recognition is performed through a classification probability acquisition method; A countermeasure module: a target output category is set, and a target result term is obtained based on the classification probability calculation result; based on the target result term, identification feedback data is obtained, a feedback threshold is set, and it is judged whether the identification feedback data reaches the feedback threshold; when the feedback threshold is reached, a target countermeasure method is used for countermeasures; An adjustment module: the initial coefficient term is adjusted through an adjustment method, and then the multi-modal fusion features are updated to obtain an updated feature term; based on the updated feature term, the classification probability is re-acquired to obtain updated feedback data, and targeted identification is performed under different conditions.

Images