A method and system for intercepting unmanned aerial vehicles (UAVs) based on multispectral index fusion sensing and laser adaptive soft and hard damage.

Abstract

This invention discloses a method and system for intercepting unmanned aerial vehicles (UAVs) based on multispectral index fusion perception and laser adaptive soft / hard damage. It relates to the field of anti-UAV security and control technology. The method includes: simultaneously acquiring multispectral raw channel images of the target area using a multispectral camera and performing preprocessing; generating a spectral index fusion image based on the preprocessed multispectral raw channel images; constructing a two-stream network and inputting the preprocessed multispectral raw channel images and the spectral index fusion image into the two-stream network for high-precision detection to obtain detection results; extracting multi-dimensional features from the detection results to identify the target type and threat level; dynamically calculating laser parameters based on the target type and threat level, with a servo turntable performing continuous aiming and simultaneously initiating soft / hard damage modes to execute interception. This invention integrates multispectral index enhanced perception and laser adaptive damage, solving the problems of low UAV detection accuracy and target confusion in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of anti-drone security and control technology, and more specifically to an anti-drone interception method and system based on multispectral index fusion perception and laser adaptive soft and hard damage. Background Technology

[0002] Currently, civilian consumer-grade and industrial-grade drones are rapidly becoming widespread. At the same time, illegal intrusion, malicious reconnaissance, and carrying dangerous goods are frequent violations of flight regulations, posing a serious threat to airport airspace, core security areas, national defense facilities, and public safety. High-precision, adaptive, and all-weather anti-drone interception technology has become a core necessity for low-altitude security and prevention.

[0003] Current mainstream anti-drone technologies mainly suffer from the following technical shortcomings: Target detection suffers from poor robustness and failure in complex scenes. Traditional anti-drone systems often rely on single visible light images for target detection, making them susceptible to interference from changes in lighting, fog, haze, and camouflage. In complex low-altitude scenes with mixed vegetation, water bodies, and buildings, small drone targets are easily confused with the background, resulting in high false negative and false positive rates. Some multispectral detection solutions simply fuse multi-band raw images without incorporating spectral indices for background suppression and target enhancement, failing to fully leverage the material differentiation and environmental interference resistance advantages of multispectral imaging, thus limiting target identification accuracy.

[0004] The rigidity of laser destruction strategies and low efficiency of existing laser counter-drone systems mean that they generally use fixed laser parameters (power, pulse width, irradiation time) and do not dynamically adapt to the type, material, and threat level of the drone. This can easily lead to energy waste, incomplete destruction, or over-destruction. Soft destruction (signal interference) and hard destruction (physical ablation) are independent of each other and cannot be automatically switched according to the target threat level, making it difficult to meet the dual needs of compliant handling of civilian drones and efficient interception of high-risk drones.

[0005] The lack of a closed-loop process leads to insufficient interception reliability. Most systems only achieve one-way execution of "detection-destruction" without a real-time assessment of the damage effect or a parameter feedback mechanism. This makes it impossible to confirm whether the target has been effectively dealt with, which can easily lead to repeated operations and missed targets. As a result, the overall combat efficiency and interception success rate of the system are low.

[0006] The existing high-precision multispectral detection technology and laser damage interception system are not integrated and coupled, and the detection results cannot be synchronized to the damage module in real time. It is difficult to achieve a coherent response of "rapid perception - precise aiming - adaptive response", which cannot meet the actual combat requirements of real-time interception of low-altitude UAVs.

[0007] Therefore, there is an urgent need to develop an anti-drone interception system that integrates multispectral index-enhanced perception, adaptive laser parameter configuration, intelligent switching between soft and hard damage, and closed-loop feedback of damage effects to address the aforementioned shortcomings of existing technologies. Summary of the Invention

[0008] In view of this, the present invention provides a method and system for intercepting anti-drone based on multispectral index fusion perception and laser adaptive soft and hard damage. It deeply couples multispectral perception and laser damage to form a closed loop of intelligent collaboration throughout the entire process. The system has fast response, strong adaptability and high interception success rate. It can be widely used in scenarios such as airports, core security and national defense airspace, and comprehensively improve the intelligence level and combat capability of low-altitude anti-drone prevention and control.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for intercepting unmanned aerial vehicles (UAVs) based on multispectral index fusion sensing and laser adaptive soft and hard damage includes: Step 1: Simultaneously acquire multispectral raw channel images of the target area using a multispectral camera and perform preprocessing; Step 2: Generate a spectral index fusion image based on the preprocessed multispectral original channel image; Step 3: Construct a dual-stream network, and input the preprocessed multispectral original channel image and the spectral index fused image into the dual-stream network respectively for high-precision detection to obtain the detection results; Step 4: Extract multi-dimensional features from the detection results to identify the target type and threat level; Step 5: Calculate laser parameters dynamically based on target type and threat level, the servo turntable completes continuous aiming, and simultaneously activates soft / hard damage mode to perform interception.

[0010] Preferably, in step 1, visible light, near-infrared, and thermal infrared multi-band signals of the target area are simultaneously acquired by a multispectral camera. Gaussian filtering is used to suppress noise in the acquired signals. Based on the visible light image, multi-band image spatial registration is completed using SIFT and RANSAC algorithms. The gray values ​​of each band signal are mapped to the [0,1] interval for data normalization to obtain standardized multispectral data.

[0011] Preferably, the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) are calculated based on the standardized multispectral data. A threshold d is set on the NDVI image and binary segmentation is performed to generate a mask image for distinguishing between vegetated and non-vegetated areas. Based on the mask image, the NDVI and NDWI features are fused to obtain a spectral index fused image, as shown in the following formula: ; in, The image pixel values ​​for NDVI. The image pixel values ​​for NDWI These are the pixel values ​​of the mask image.

[0012] Preferably, the dual-stream network is a dual-stream feature extraction convolutional neural network, comprising: a backbone network, a feature fusion module, and a detection output module connected in sequence; the backbone network comprises two completely independent feature extraction backbone networks with identical structures; the high-precision detection process includes: The preprocessed multispectral raw channel image is input into the first backbone network to extract the basic visual features of the target and output the multispectral raw feature map. The spectral index fused image is input into the second backbone network to extract the target's spectral differences, background suppression, and small target enhancement features, and output the spectral index enhancement feature map. The feature fusion module performs channel concatenation and feature weighting fusion on the output feature map to obtain a fused feature map; The system performs target classification, bounding box regression, and confidence calculation on the fused feature map, and outputs detection results including target location and confidence score.

[0013] Preferably, step 4 includes: extracting multi-dimensional static and dynamic features from the target area corresponding to the detection result; extracting the target contour and aspect ratio shape features using the Canny edge detection algorithm; extracting the target heat source location and temperature gradient thermal radiation features in the thermal infrared image using the connected component analysis algorithm; calculating the multispectral band reflectance ratio to obtain the target material spectral features, thus completing static feature extraction; and calculating the target inter-frame pixel change using the inter-frame difference algorithm, and iteratively solving the target flight speed, acceleration, and trajectory curvature using the Kalman filter algorithm, thus completing dynamic feature extraction.

[0014] Preferably, the static features and dynamic features are combined to construct an n-dimensional target feature vector. Retrieve pre-stored standard feature template vectors for civilian or military drones. The feature matching degree S is calculated using the cosine similarity algorithm, and the calculation formula is as follows: ; Where i is the index of the feature component, and n is the total number of dimensions of the multi-dimensional feature vector; Preset similarity matching threshold When the matching degree When the target matches the corresponding standard feature template of the same type, civilian drones are classified as low threat level, and military / drones with dangerous payloads are classified as high threat level; when the matching degree At that time, the target is marked as an unknown target and manual review is triggered.

[0015] Preferably, step 5 includes: obtaining the real-time distance r between the target and the servo turntable based on the target type, threat level, and target location information in the detection results; and calculating the equivalent radius r0 of the key parts of the target using the target bounding box size. Laser parameters are dynamically calculated based on a laser damage energy density threshold algorithm. The laser energy density E must meet the target damage requirements. The calculation formula is as follows: ; Where E is the minimum energy density required for the critical part of the target, P is the laser output power, t is the laser irradiation time, and r0 is the equivalent radius of the critical part of the target; the laser output power P and irradiation time t are calculated by reversing the above formula.

[0016] Preferably, it also includes: using a Kalman filter algorithm to predict the target's position at the next moment, the calculation process of which is as follows: Based on the target's current location ,speed Predict the position at the next moment and speed The formula is: ; ; in, This refers to the frame interval time. Obtain the actual position of the target using a servo turntable. Calculate position error Combined with Kalman gain Update the target position and velocity using the following formula: ; ; Collect servo turntable attitude angle changes It corrects the turntable angle command in real time and simultaneously activates the corresponding damage mode to perform interception.

[0017] Preferably, an anti-drone interception system based on multispectral index fusion perception and laser adaptive soft and hard damage includes: Multispectral index fusion sensing module: used to acquire visible light, near infrared, and thermal infrared multi-band signals of the target area and generate a spectral index fusion image; Multimodal data preprocessing module: used to perform Gaussian filtering for noise reduction, SIFT+RANSAC registration, and interval normalization on the acquired multispectral signals, and output standardized multispectral data; Detection module: Used to construct a dual-stream network, and input the preprocessed multispectral raw channel image and spectral index fused image into the dual-stream network respectively for high-precision detection to obtain the detection result; Multi-dimensional feature identification module: used to extract static and dynamic features from the detection results, and identify target type and threat level through cosine similarity algorithm; The parameter dynamic configuration module dynamically calculates the core parameters of laser output power and irradiation time based on the target type, threat level, target size, and distance using an energy density algorithm. Oriented tracking and aiming module: used to predict the target trajectory using Kalman filtering and achieve continuous aiming via a servo turntable; Adaptive soft and hard damage execution module: used to automatically switch damage modes and perform interception operations based on the threat level; Effect evaluation feedback module: used to collect multispectral signals after interception, compare with target features to determine the damage effect, and if the damage is insufficient, to provide feedback to optimize laser parameters.

[0018] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method and system for anti-drone interception based on multispectral index fusion perception and laser adaptive soft and hard damage. Through multispectral index fusion and a dual-stream network detection architecture, it significantly improves the detection accuracy of small drone targets in complex scenarios such as vegetation, water bodies, and haze, effectively suppresses background interference and environmental clutter, breaks through the robustness bottleneck of traditional single-spectrum detection, and achieves stable target perception in all weather and multiple scenarios. The system dynamically calculates laser parameters and adaptively switches between soft and hard damage modes based on target type and threat level. For civilian drones, it uses signal interference soft damage, while for high-risk drones, it performs ablation hard damage on key parts, balancing compliant handling and efficient interception, avoiding energy waste and excessive damage. Relying on Kalman filter trajectory prediction and gyroscope attitude compensation, it achieves continuous and accurate targeting. Combined with damage effect evaluation and a closed-loop feedback mechanism, it verifies the handling effect and optimizes parameters in real time, ensuring thorough and complete interception. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 The method flowchart provided by the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, this invention discloses an anti-drone interception method based on multispectral index fusion sensing and laser adaptive soft and hard damage, including: Step 1: Simultaneously acquire multispectral raw channel images of the target area using a multispectral camera and perform preprocessing; Step 2: Generate a spectral index fusion image based on the preprocessed multispectral original channel image; Step 3: Construct a dual-stream network, and input the preprocessed multispectral original channel image and the spectral index fused image into the dual-stream network respectively for high-precision detection to obtain the detection results; Step 4: Extract multi-dimensional features from the detection results to identify the target type and threat level; Step 5: Calculate laser parameters dynamically based on target type and threat level, the servo turntable completes continuous aiming, and simultaneously activates soft / hard damage mode to perform interception.

[0023] Specifically, in step 1, visible light, near-infrared, and thermal infrared multi-band signals of the target area are simultaneously acquired by a multispectral camera. Gaussian filtering is used to suppress noise in the acquired signals. Based on the visible light image, the multi-band image spatial registration is completed using SIFT and RANSAC algorithms. The gray values ​​of each band signal are mapped to the [0,1] interval for data normalization to obtain standardized multispectral data.

[0024] In the same specific embodiment of the present invention, the registration reference is determined by using the visible light image after Gaussian filtering and denoising as the sole reference image for multi-band image spatial registration, and the objects to be registered are the synchronously acquired near-infrared and thermal infrared images.

[0025] SIFT feature extraction extracts SIFT scale-invariant feature points and corresponding feature descriptors from visible light reference images and near-infrared / thermal infrared images to be registered, solving the feature matching problem caused by band differences and scale changes in multi-band images.

[0026] The initial feature matching is based on the similarity calculation of SIFT feature descriptors, and completes the coarse matching of feature points between the visible light reference image and other band images to obtain the initial set of matching point pairs.

[0027] RANSAC mismatch removal uses the RANSAC algorithm for iterative calculation, removing mismatched outliers from the initial matching point pairs and retaining precisely matched inliers. Based on the inliers, the spatial transformation matrix of the image to be registered relative to the visible light reference image is solved.

[0028] Based on the solved spatial transformation matrix, spatial coordinate alignment is performed on near-infrared and thermal infrared images through spatial remapping and coordinate transformation, achieving pixel-level spatial alignment of all multi-band images with the visible light reference image, thus completing the spatial registration of multi-band images.

[0029] Specifically, the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) are calculated based on the standardized multispectral data. A threshold d is set on the NDVI image and binary segmentation is performed to generate a mask image for distinguishing between vegetated and non-vegetated areas. Based on the mask image, the NDVI and NDWI features are fused to obtain a spectral index fused image, as shown in the following formula: ; in, The image pixel values ​​for NDVI. The image pixel values ​​for NDWI These are the pixel values ​​of the mask image.

[0030] In a specific embodiment of the present invention, binary segmentation is performed to determine the standardized NDVI image pixel by pixel. If the NDVI pixel value is greater than d, it is determined to be a vegetation area, and the corresponding pixel value in the mask image is... =1; NDVI pixel value ≤ d, determined as a non-vegetated area (water body, building, airspace, etc.), the corresponding pixel value in the mask image. =0.

[0031] Specifically, the dual-stream network is a dual-stream feature extraction convolutional neural network, comprising: a backbone network, a feature fusion module, and a detection output module connected in sequence; the backbone network includes two completely independent feature extraction backbone networks with identical structures; the high-precision detection process includes: The preprocessed multispectral raw channel image is input into the first backbone network to extract the basic visual features of the target and output the multispectral raw feature map. The spectral index fused image is input into the second backbone network to extract the target's spectral differences, background suppression, and small target enhancement features, and output the spectral index enhancement feature map. The feature fusion module performs channel concatenation and feature weighting fusion on the output feature map to obtain a fused feature map; The system performs target classification, bounding box regression, and confidence calculation on the fused feature map, and outputs detection results including target location and confidence score.

[0032] In a specific embodiment of the present invention, the dual-stream feature extraction convolutional neural network uses two isomorphic lightweight convolutional neural networks as the backbone network. The two networks have completely identical structures and their parameters are trained independently. The backbone network is preferably any one of the CSPDarkNet series, ResNet18, ResNet34, or MobileNetV3.

[0033] Specifically, step 4 includes: extracting multi-dimensional static and dynamic features from the target area corresponding to the detection result; using the Canny edge detection algorithm to extract the target contour and aspect ratio shape features; using the connected component analysis algorithm to extract the target heat source location and temperature gradient thermal radiation features in the thermal infrared image; calculating the multispectral band reflectance ratio to obtain the target material spectral features, thus completing static feature extraction; using the inter-frame difference algorithm to calculate the target inter-frame pixel change; and combining the Kalman filter algorithm to iteratively solve the target flight speed, acceleration, and trajectory curvature, thus completing dynamic feature extraction.

[0034] In a specific embodiment of the present invention, dynamic feature extraction takes a multispectral image sequence as input. First, the initial motion information of the target is obtained through an inter-frame difference algorithm. Then, the target's flight speed and acceleration are calculated by iterative optimization using a Kalman filter algorithm. Finally, the trajectory curvature is calculated based on the motion trajectory to complete the dynamic feature extraction. The specific process and calculation are as follows: Preprocessing and initial target position acquisition: Input a multispectral image sequence (synchronized visible light, near-infrared, and thermal infrared image sequences) after Gaussian filtering, denoising, and spatial registration. Let the frame rate of the image sequence be f, and the frame interval be... (Usually f=30fps, i.e.) ).

[0035] The inter-frame difference algorithm is used to calculate the pixel changes of the target across frames, thus locating the target's initial position in each frame. Three consecutive frames are selected. (frame k-1) (frame k) (Frame k+1), the formula for calculating the inter-frame difference is: ; ; ; Where Q is an empirical threshold (Q=20, to adapt to the gray range of multispectral images). Based on the target motion region segmentation results, the target center coordinates are extracted through connected component analysis to obtain the initial position of the target in the k-th frame. Construct the initial position sequence .

[0036] Kalman filter algorithm initialization (adapting to the target motion model)

[0037] The target's flight process can be approximated as uniformly accelerated linear motion (CA model). Combining the state estimation characteristics of the Kalman filter algorithm, the core parameters such as the system state vector and state transition matrix are defined, and initialization is completed, as follows: The target's position, velocity, and acceleration in the x and y directions of the image coordinate system are taken as state variables to construct a 6-dimensional state vector. :

[0038] in, , The optimal position of the target in the k-th frame. , Let x and y be the flight velocities in the k-th frame. , Let x be the flight acceleration in the x and y directions of the k-th frame.

[0039] State transition matrix (6×6 matrix), constructed based on a uniformly accelerated motion model, reflecting the transmission relationship of state variables between adjacent frames:

[0040] Observation matrix (2×6 matrix), only the target position is observed, ignoring direct observations of velocity and acceleration:

[0041] Process noise covariance matrix (6×6 matrix), reflecting random noise (such as airflow interference) during the target's motion, with the following values:

[0042] Observation noise covariance matrix (2×2 matrix), reflecting the positional error of inter-frame differential positioning, with the following values:

[0043] Initial state vector Based on the initial position of the first 3 frames, the initial velocity and acceleration are fitted and initialized as follows: ; Initial error covariance matrix (6×6 matrix), initialized as follows:

[0044] Kalman filter iterative solution for velocity and acceleration: Based on the initialization parameters, an iterative "prediction → update" process is performed on each frame of the image to progressively optimize the target position, velocity, and acceleration. The specific calculation process is as follows (taking the k-th frame iteration as an example, k≥4): State prediction (predicting the state of frame k) Predict the state vector of the k-th frame (Based on the optimal state of frame k-1) ): ; Expand the calculation (taking the x-direction as an example, the y-direction is similar): ; ; ; Predicting the error covariance matrix of the k-th frame : ; in, State transition matrix The transpose of the matrix, Let be the error covariance matrix of the (k-1)th frame.

[0045] Obtain the observation value of the kth frame (The initial position of the target obtained based on inter-frame difference) ): ; Calculate Kalman gain (6×2 matrix), weighing the predicted values ​​against the observed values: ; in, Observation matrix The transpose of the matrix, This represents finding the inverse of a matrix.

[0046] Calculate the optimal state vector for the k-th frame. (Includes optimal position, velocity, and acceleration):

[0047] Unfold and extract velocity and acceleration (taking the x-direction as an example): ; ; in, Represents the Kalman gain matrix The element in the i-th row and j-th column; velocity in the y-direction. acceleration The calculation method is consistent with that in the x-direction.

[0048] Update the error covariance matrix of the k-th frame. In preparation for the next frame iteration: ; in, It is a 6×6 identity matrix.

[0049] Repeat the steps above to perform iterative calculations for each frame of the image sequence to obtain the optimal velocity of the target in each frame. acceleration This generates velocity and acceleration sequences, enabling the extraction of dynamic features of velocity and acceleration.

[0050] The trajectory curvature is calculated based on the iterative results. The trajectory curvature reflects the degree of bending of the target's flight trajectory. Based on the position and velocity sequences obtained from the Kalman filter iteration, the curvature formula is used for calculation. The specific process is as follows: Extract the optimal position from three consecutive frames (frames k-1, k, and k+1). , , Calculate the first derivative (velocity component) and second derivative (acceleration component) of the target trajectory: First derivative (velocity): , ; Second derivative (acceleration): , ; trajectory curvature The calculation formula is as follows (derived based on the curvature formula for plane curves): ; Perform the above calculation on each frame (k≥2, k≤N-1, N is the total number of frames in the image sequence) to obtain the trajectory curvature sequence. This completes the dynamic feature extraction of trajectory curvature.

[0051] The velocity sequence, acceleration sequence, and calculated trajectory curvature sequence obtained by Kalman filtering iteration are integrated into a multi-dimensional dynamic feature set of the target, which is used for subsequent target type identification and threat level determination, thus completing the entire dynamic feature extraction process.

[0052] Specifically, the static and dynamic features are combined to construct an n-dimensional target feature vector. Retrieve pre-stored standard feature template vectors for civilian or military drones. The feature matching degree S is calculated using the cosine similarity algorithm, and the calculation formula is as follows: ; Where i is the index of the feature component, and n is the total number of dimensions of the multi-dimensional feature vector; Preset similarity matching threshold When the matching degree When the target matches the corresponding standard feature template of the same type, civilian drones are classified as low threat level, and military / drones with dangerous payloads are classified as high threat level; when the matching degree At that time, the target is marked as an unknown target and manual review is triggered.

[0053] In a specific embodiment of the present invention, a preset similarity matching threshold is used. When the matching degree When the target matches the corresponding standard feature template, it is determined that the target is of the same type. The classification criteria are as follows: Civilian drones have no dangerous payloads, their flight purposes are mostly non-malicious, causing only airspace interference and posing no direct security risk, therefore they are classified as low threat level; Military / drones carrying dangerous payloads possess reconnaissance, attack, or dangerous goods carrying capabilities, threatening public safety and national defense security, posing a direct security hazard, therefore they are classified as high threat level. At that time, the target is marked as an unknown target and manual review is triggered.

[0054] Specifically, step 5 includes: obtaining the real-time distance r between the target and the servo turntable based on the target type, threat level, and target location information in the detection results; and calculating the equivalent radius r0 of the key parts of the target using the target bounding box size. Laser parameters are dynamically calculated based on a laser damage energy density threshold algorithm. The laser energy density E must meet the target damage requirements. The calculation formula is as follows: ; Where E is the minimum energy density required for the critical part of the target, P is the laser output power, t is the laser irradiation time, and r0 is the equivalent radius of the critical part of the target; the laser output power P and irradiation time t are calculated by reversing the above formula.

[0055] In a specific embodiment of the present invention, the core parameters of the bounding box of the target in the current frame are extracted from the dual-stream network detection results: bounding box width w (in image coordinates, unit: pixels) and bounding box height h (in image coordinates, unit: pixels). The bounding box completely encloses the target area with the target center as the reference, and is denoted as (w,h).

[0056] By combining the pixel resolution α of the multispectral image (unit: m / pixel, i.e., the distance of each pixel in the actual scene, preset α=0.01m / pixel, which can be calibrated according to camera parameters), the pixel size of the bounding box is converted into the actual physical size:

[0057] Where W is the actual width of the target (unit: m) and H is the actual height of the target (unit: m).

[0058] Based on the target type identification results (civilian / military / unknown target), determine the location and size proportion of key components: Civilian drones: The key components are the battery / motor, which are usually located in the center of the target area. Their actual size is about 30% to 50% of the actual width of the target. Taking the average of 40%, the actual width of the key component is w0 = W × 0.4. Military / UAVs with hazardous payloads: The key parts are the engine / ammunition compartment, and their actual size is about 40% to 60% of the actual width of the target. Taking the average of 50%, the actual width of the key parts is w0 = W × 0.5. Unknown target: Based on the principle of conservatism, take the middle value. The actual width of the key part w0 = W × 0.45.

[0059] The key part of the target is approximated as a regular circle (the core of laser damage is to focus on the center point of the key part, and the circular equivalent can simplify the energy density calculation), and its diameter is equal to the actual width w0 of the key part.

[0060] Based on the actual width w0 of the critical part, the equivalent radius r0 (unit: m) is calculated using the following formula:

[0061] Where β is the size ratio of the key part (civilian β=0.4, military β=0.5, unknown β=0.45), α is the pixel resolution, and w is the pixel width of the bounding box.

[0062] If the bounding box pixel width of a civilian drone is w = 50 pixels, the pixel resolution is α = 0.01m / pixel, and β = 0.4, then: .

[0063] Furthermore, in this embodiment, the preset low-threat target E1 is 100-200 J / cm². 2 High-threat targets E2 = 300-500 J / cm² 2 P is the laser output power, t is the laser irradiation time, and r0 is the equivalent radius of the key part of the target. Based on the above formula, the laser output power P and irradiation time t are calculated. Low-threat civilian UAVs correspond to low output power P1=100-300W and short irradiation time t1=1-3s, while high-threat military / UAVs with dangerous payloads correspond to high output power P2=500-1000W and long irradiation time t2=3-5s. At the same time, the laser pulse width is determined to be 100-500ns and the repetition frequency is 1-10kHz. Specifically, this also includes: using the Kalman filter algorithm to predict the target's position at the next moment, the calculation process of which is as follows: Based on the target's current location ,speed Predict the position at the next moment and speed The formula is: ; ; in, This refers to the frame interval time. Obtain the actual position of the target using a servo turntable. Calculate position error Combined with Kalman gain Update the target position and velocity using the following formula: ; ; Collect servo turntable attitude angle changes It corrects the turntable angle command in real time and simultaneously activates the corresponding damage mode to perform interception.

[0064] In one specific embodiment of the present invention, the receiver of the UAV is irradiated by a laser transmitting antenna to interfere with its signal reception, forcing the UAV to return or make an emergency landing; the high-threat military / UAV with dangerous payload activates hard damage mode, activates a high-power laser according to the calculated laser parameters, focuses the laser on the target engine, battery and other key parts, and achieves ablation and destruction through thermal effect, so that the UAV loses its flight capability and completes the interception.

[0065] Specifically, an anti-drone interception system based on multispectral index fusion sensing and laser adaptive soft and hard damage includes: Multispectral index fusion sensing module: used to acquire visible light, near infrared, and thermal infrared multi-band signals of the target area and generate a spectral index fusion image; Multimodal data preprocessing module: used to perform Gaussian filtering for noise reduction, SIFT+RANSAC registration, and interval normalization on the acquired multispectral signals, and output standardized multispectral data; Detection module: Used to construct a dual-stream network, and input the preprocessed multispectral raw channel image and spectral index fused image into the dual-stream network respectively for high-precision detection to obtain the detection result; Multi-dimensional feature identification module: used to extract static and dynamic features from the detection results, and identify target type and threat level through cosine similarity algorithm; The parameter dynamic configuration module dynamically calculates the core parameters of laser output power and irradiation time based on the target type, threat level, target size, and distance using an energy density algorithm. Oriented tracking and aiming module: used to predict the target trajectory using Kalman filtering and achieve continuous aiming via a servo turntable; Adaptive soft and hard damage execution module: used to automatically switch damage modes and perform interception operations based on the threat level; Effect evaluation feedback module: used to collect multispectral signals after interception, compare with target features to determine the damage effect, and if the damage is insufficient, to provide feedback to optimize laser parameters.

[0066] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0067] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intercepting unmanned aerial vehicles (UAVs) based on multispectral index fusion sensing and laser adaptive soft and hard damage, characterized in that, include: Step 1: Simultaneously acquire multispectral raw channel images of the target area using a multispectral camera and perform preprocessing; Step 2: Generate a spectral index fusion image based on the preprocessed multispectral original channel image; Step 3: Construct a dual-stream network, and input the preprocessed multispectral original channel image and the spectral index fused image into the dual-stream network respectively for high-precision detection to obtain the detection results; Step 4: Extract multi-dimensional features from the detection results to identify the target type and threat level; Step 5: Calculate laser parameters dynamically based on target type and threat level, the servo turntable completes continuous aiming, and simultaneously activates soft / hard damage mode to perform interception.

2. The anti-drone interception method based on multispectral index fusion sensing and laser adaptive soft and hard damage as described in claim 1, characterized in that, In step 1, visible light, near-infrared, and thermal infrared multi-band signals of the target area are simultaneously acquired by a multispectral camera. Gaussian filtering is used to suppress noise in the acquired signals. Based on the visible light image, the multi-band image spatial registration is completed using SIFT and RANSAC algorithms. The gray values ​​of each band signal are mapped to the [0,1] interval for data normalization to obtain standardized multispectral data.

3. The anti-drone interception method based on multispectral index fusion sensing and laser adaptive soft and hard damage according to claim 2, characterized in that, Based on the standardized multispectral data, the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Water Index (NDWI) are calculated. A threshold d is set on the NDVI image, and binary segmentation is performed to generate a mask image to distinguish between vegetated and non-vegetated areas. Based on the mask image, the NDVI and NDWI features are fused to obtain a spectral index fused image, as shown in the following formula: ； in, The image pixel values ​​for NDVI. The image pixel values ​​for NDWI These are the pixel values ​​of the mask image.

4. The anti-drone interception method based on multispectral index fusion sensing and laser adaptive soft and hard damage according to claim 3, characterized in that, The dual-stream network is a dual-stream feature extraction convolutional neural network, comprising: a backbone network, a feature fusion module, and a detection output module connected in sequence; the backbone network includes two completely independent feature extraction backbone networks with identical structures; the high-precision detection process includes: The preprocessed multispectral raw channel image is input into the first backbone network to extract the basic visual features of the target and output the multispectral raw feature map. The spectral index fused image is input into the second backbone network to extract the target's spectral differences, background suppression, and small target enhancement features, and output the spectral index enhancement feature map. The feature fusion module performs channel concatenation and feature weighting fusion on the output feature map to obtain a fused feature map; The system performs target classification, bounding box regression, and confidence calculation on the fused feature map, and outputs detection results including target location and confidence score.

5. The anti-drone interception method based on multispectral index fusion sensing and laser adaptive soft and hard damage as described in claim 1, characterized in that, Step 4 includes: extracting multi-dimensional static and dynamic features from the target area corresponding to the detection results; using the Canny edge detection algorithm to extract the target contour and aspect ratio shape features; using the connected component analysis algorithm to extract the target heat source location and temperature gradient thermal radiation features in the thermal infrared image; calculating the multispectral band reflectance ratio to obtain the target material spectral features, thus completing static feature extraction; using the inter-frame difference algorithm to calculate the target inter-frame pixel change; and combining the Kalman filter algorithm to iteratively solve the target flight speed, acceleration, and trajectory curvature, thus completing dynamic feature extraction.

6. The anti-drone interception method based on multispectral index fusion sensing and laser adaptive soft and hard damage according to claim 5, characterized in that, The static and dynamic features are combined to construct an n-dimensional target feature vector. Retrieve pre-stored standard feature template vectors for civilian or military drones. The feature matching degree S is calculated using the cosine similarity algorithm, and the calculation formula is as follows: ； Where i is the index of the feature component, and n is the total number of dimensions of the multi-dimensional feature vector; Preset similarity matching threshold When the matching degree When the target matches the corresponding standard feature template of the same type, civilian drones are classified as low threat level, and military / drones with dangerous payloads are classified as high threat level; when the matching degree At that time, the target is marked as an unknown target and manual review is triggered.

7. The anti-drone interception method based on multispectral index fusion sensing and laser adaptive soft and hard damage according to claim 6, characterized in that, Step 5 includes: obtaining the real-time distance r between the target and the servo turntable based on the target type, threat level, and target location information in the detection results; and calculating the equivalent radius r0 of the key parts of the target based on the target bounding box size. Laser parameters are dynamically calculated based on a laser damage energy density threshold algorithm. The laser energy density E must meet the target damage requirements. The calculation formula is as follows: ； Where E is the minimum energy density required for the critical part of the target, P is the laser output power, t is the laser irradiation time, and r0 is the equivalent radius of the critical part of the target; the laser output power P and irradiation time t are calculated by reversing the above formula.

8. The anti-drone interception method based on multispectral index fusion sensing and laser adaptive soft and hard damage according to claim 7, characterized in that, Also includes: The Kalman filter algorithm is used to predict the target's position at the next moment. The calculation process is as follows: Based on the target's current location ,speed Predict the position at the next moment. and speed The formula is: ； ； in, This refers to the frame interval time. Obtain the actual position of the target using a servo turntable. Calculate position error Combined with Kalman gain Update the target position and velocity using the following formula: ； ； Collect servo turntable attitude angle changes It corrects the turntable angle command in real time and simultaneously activates the corresponding damage mode to perform interception.

9. A counter-drone interception system based on multispectral index fusion sensing and laser adaptive soft and hard damage, characterized in that, include: Multispectral index fusion sensing module: used to acquire visible light, near infrared, and thermal infrared multi-band signals of the target area and generate a spectral index fusion image; Multimodal data preprocessing module: used to perform Gaussian filtering for noise reduction, SIFT+RANSAC registration, and interval normalization on the acquired multispectral signals, and output standardized multispectral data; Detection module: Used to construct a two-stream network, and input the preprocessed multispectral raw channel image and spectral index fused image into the two-stream network respectively for high-precision detection to obtain the detection result; Multi-dimensional feature identification module: used to extract static and dynamic features from the detection results, and identify target type and threat level through cosine similarity algorithm; The parameter dynamic configuration module dynamically calculates the core parameters of laser output power and irradiation time based on the target type, threat level, target size, and distance using an energy density algorithm. Oriented tracking and aiming module: used to predict the target trajectory using Kalman filtering and achieve continuous aiming through a servo turntable; Adaptive soft and hard damage execution module: used to automatically switch damage modes and perform interception operations based on the threat level; Effect evaluation feedback module: used to collect multispectral signals after interception, compare with target features to determine the damage effect, and if the damage is insufficient, to provide feedback to optimize laser parameters.

Images