Power transmission channel unmanned aerial vehicle target spectrum adaptive matching method and system

Abstract

The application provides a power transmission channel unmanned aerial vehicle target spectrum adaptive matching method and system, relates to the technical field of image processing, and comprises the following steps: acquiring a multispectral image sequence and channel structure prior data, projecting the channel structure prior data to the multispectral image sequence to generate a channel structure mask; extracting a motion area deviating from the channel structure mask to obtain a candidate multispectral response sequence, performing structure reflection stripping on the candidate multispectral response sequence according to edge neighborhood spectrum in the channel structure mask to obtain structure stripped spectral data; determining the penetration reliability degree of each spectral band according to the structure stripped spectral data to generate spectral band credible distribution data; generating target real spectral data from the structure stripped spectral data by using the spectral band credible distribution data, and outputting an unmanned aerial vehicle matching result. Through structure reflection stripping and spectral band credibility reconstruction, the application improves the unmanned aerial vehicle recognition accuracy and long-term supervision stability under complex weather.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for adaptive spectral matching of UAV targets in power transmission channels. Background Technology

[0002] In the airspace surveillance of ultra-high voltage transmission channels, multispectral monitoring equipment is often used to observe intruding drones from a distance. Since the transmission channel is surrounded by structural backgrounds such as conductors, towers, and fittings, multispectral observation can utilize the differentiated responses of the target at different wavelengths to provide multidimensional spectral information for distinguishing drones from the background.

[0003] Currently, conventional multispectral target matching methods typically pre-define fixed combinations of spectral bands and construct a single target spectral template. During target recognition, these methods directly extract multispectral response data of the moving area within the monitoring field of view, calculate the similarity between the acquired spectral band response values ​​and the fixed target spectral template, and output the target matching status based on the calculation results.

[0004] However, under dynamic weather conditions such as fog or backlighting, the penetrating power of light of different wavelengths varies, and local reflections or hot spots are easily generated at the edges of conductors and fittings. The method described above, which directly extracts the response of the moving area and compares it with a fixed template, does not consider the differentiation of imaging quality in each spectral band and the spurious response caused by the structural background. This leads to the inclusion of noisy data from the failed spectral band in the comparison, which can easily misidentify conductor reflections or hot spots on fittings affected by the environment as targets, resulting in mismatches and missed matches. Furthermore, erroneous samples are continuously written into the historical template library, making it difficult to maintain stable identification in long-term continuous monitoring. Summary of the Invention

[0005] A first aspect of this application provides an adaptive spectral matching method for UAV targets in power transmission channels, comprising: Acquire multispectral image sequences and channel structure prior data, and project the channel structure prior data onto the multispectral image sequences to generate a channel structure mask; The motion region deviating from the channel structure mask is extracted to obtain candidate multi-band response sequences. The candidate multi-band response sequences are then subjected to structural reflectance stripping based on the edge neighborhood spectrum in the channel structure mask to obtain structural stripping spectral data. Based on the structure stripping spectral data, the penetration reliability of each spectral band is determined, and reliable spectral band distribution data is generated; The target's true spectral data is generated from the structure-extracted spectral data using the reliable distribution data of the spectral band, and the UAV matching result is output.

[0006] Optionally, in one possible implementation of the first aspect, projecting the channel structure prior data onto the multispectral image sequence to generate a channel structure mask includes: Acquire the current observation attitude and focal length parameters of the monitoring equipment; Based on the current observation posture and the focal length parameters, the three-dimensional structure wireframe in the prior data of the channel structure is projected onto the two-dimensional imaging plane of the multispectral image sequence to obtain the structure projection profile; An initial structural mask is generated based on the pixel distribution of the structural projection contour in the two-dimensional imaging plane. Extract the edge pixel set of the initial structure mask, and determine the number of tolerance pixels to be extended outward based on the wind deflection amplitude of the transmission line and the calibration error of the monitoring equipment; The number of tolerance pixels is shifted outward along the normal direction of the edge pixel set to generate a structural extension region; The extended region of the structure is added to the outer edge of the initial structure mask to obtain the channel structure mask.

[0007] Optionally, in one possible implementation of the first aspect, generating an initial structural mask based on the pixel distribution of the structural projection profile in the two-dimensional imaging plane includes: Extract the spatial depth value corresponding to each structural line segment in the three-dimensional structural wireframe; The structural projection contour is layered according to the spatial depth value to obtain multiple depth level contours, and the local imaging region corresponding to any of the depth level contours is determined in the two-dimensional imaging plane. Obtain the radiance difference between adjacent pixels within the local imaging region; Determine the statistical mean and standard deviation of the radiance difference within the local imaging region, and determine the pixel position corresponding to the radiance difference that exceeds the range of the statistical mean plus or minus the standard deviation as the actual structural boundary; The set of projected boundary pixels corresponding to the depth level contour in the two-dimensional imaging plane is updated to the pixel position corresponding to the actual structural boundary to obtain the corrected level contour. The initial structure mask is generated by combining the modified layer contours corresponding to each depth layer contour.

[0008] Optionally, in one possible implementation of the first aspect, the extraction of motion regions deviating from the channel structure mask to obtain candidate multi-band response sequences includes: The difference in radiance between adjacent frames in the multispectral image sequence at the same pixel coordinates is determined to obtain a motion difference image; The pixels covered by the channel structure mask are removed from the motion difference image to obtain the residual motion region; Extract the set of spatially adjacent moving pixels with the same direction of motion from the residual motion region to obtain the independent motion region; The coordinates of the independent motion region in each spectral band of the multispectral image sequence are determined, the radiance value at the coordinates of the coordinates is extracted, and the candidate multispectral response sequence is generated.

[0009] Optionally, in one possible implementation of the first aspect, extracting a set of spatially adjacent moving pixels with the same direction of motion in the residual motion region to obtain an independent motion region includes: Extract the set of spatially adjacent candidate pixels in the residual motion region, and obtain the centroid position sequence of the candidate pixel set in multiple consecutive frames of images; Determine the displacement direction between adjacent frames in the centroid position sequence; When the displacement direction does not alternately reverse between consecutive frames, and the centroid position sequence does not exhibit a state of back-and-forth movement along the same trajectory, the candidate pixel set is determined to be a valid set of moving pixels. The set of effective moving pixels is taken as the independent motion region.

[0010] Optionally, in one possible implementation of the first aspect, the step of performing structural reflectance stripping on the candidate multi-band response sequence based on the edge neighborhood spectrum in the channel structure mask to obtain structurally stripped spectral data includes: Starting from the boundary of the channel structure mask, extend at least three adjacent pixel rows outward from the channel structure mask to obtain the edge neighborhood band; Extract the spectral response values ​​of each pixel within the edge neighborhood band in the multispectral image sequence to generate neighborhood spectral distribution data; The decreasing radiance distribution of the edge neighborhood band along the direction away from the boundary of the channel structure mask is determined based on the neighborhood spectral distribution data. Edge radiation subtraction reference data is constructed based on the described decreasing radiance distribution. Calculate the normal distance of each moving pixel in the candidate multi-band response sequence from the boundary of the channel structure mask; match the corresponding radiance subtraction value in the edge radiance subtraction reference data according to the normal distance, and subtract the radiance subtraction value from the corresponding moving pixel in the candidate multi-band response sequence to obtain the structure stripping spectral data.

[0011] Optionally, in one possible implementation of the first aspect, determining the decreasing radiance distribution of the edge neighborhood band along the direction away from the channel structure mask boundary based on the neighborhood spectral distribution data includes: The edge neighborhood band is divided into a group of pixel sub-bands arranged along the normal direction of the channel structure mask boundary; For any of the aforementioned pixel sub-bands, obtain the radiance value of the center pixel within the pixel sub-band in a single spectral band; Based on the radiance value of each pixel sub-band and the normal distance of the pixel sub-band from the boundary of the channel structure mask, a negative correlation relationship is determined as radiance extends with normal distance. The radiance drop amplitude between adjacent pixel sub-bands is determined based on the negative correlation correspondence. The radiance decrease amplitude of each individual spectral band is combined to generate the radiance decrease distribution state.

[0012] Optionally, in one possible implementation of the first aspect, determining the penetration reliability of each spectral band based on the structural stripping spectral data and generating reliable spectral band distribution data includes: Extract the local contrast and cross-band response difference magnitude of each spectral band in the structure stripping spectral data; Obtain the ambient light vector and fog concentration parameters during the acquisition of the multispectral image sequence; The saturation exceedance risk of each spectral band is determined based on the ambient light vector, and the scattering and obscuring degree of each spectral band is determined based on the fog concentration parameter. The effective signal strength value of each spectral band is determined based on the difference between the local contrast and the cross-spectral response, and the environmental interference level of each spectral band is determined based on the saturation exceedance risk and the scattering shielding degree. Calculate the ratio between the effective signal strength value and the environmental interference level value, and determine the penetration reliability of each spectral band based on the ratio. The penetration reliability of each spectral band is arranged in order of wavelength to generate reliable distribution data for that spectral band.

[0013] Optionally, in one possible implementation of the first aspect, extracting the magnitude of the difference between the local contrast and the cross-spectral response of each spectral band in the structure stripping spectral data includes: For any target spectral band in the structure stripping spectral data, the difference in radiance between the target region pixels and the local background pixels within the target spectral band is determined to obtain the local contrast. Select a reference spectral segment adjacent to the target spectral segment, and align the pixel response values ​​of the target spectral segment and the reference spectral segment at the same spatial location; Obtain the response difference sequence between the aligned target spectral band and the reference spectral band; The amplitude of the cross-spectral response difference is determined based on the fluctuation amplitude of the response difference sequence.

[0014] Optionally, in one possible implementation of the first aspect, generating target true spectral data from the structure-exfoliated spectral data using the spectral band reliable distribution data includes: Based on the reliable distribution data of the spectral bands, calculate the statistical mean of the penetration reliability of each spectral band, and select the spectral bands with a penetration reliability greater than the statistical mean as valid spectral bands; Extract the wavelength identifiers and pixel response values ​​belonging to the effective spectral band from the structure stripping spectral data to construct a discrete effective spectral point set; Curves are constructed for the discrete effective spectral point set to establish the correspondence between wavelength and spectral distribution of radiance; Based on the spectral distribution correspondence, the derived values ​​of radiance at interpolation wavelengths not included in the effective spectral band are calculated to obtain the reconstructed spectral curve; The reconstructed spectral curves are aligned with radiance to generate the target's true spectral data.

[0015] Optionally, in one possible implementation of the first aspect, the step of calculating the derived radiance value at the interpolated wavelength not included in the effective spectral band based on the spectral distribution correspondence to obtain the reconstructed spectral curve includes: Determine the wavelength interval between two adjacent data points in the discrete effective spectral point set; When the wavelength interval is greater than the original spectral band wavelength step size of the multispectral image sequence, the interval corresponding to the wavelength interval is determined as the interval to be interpolated; Obtain the radiance and wavelength values ​​of the data points on both sides of the interpolation interval; Based on the radiance and wavelength values ​​of the data points on both sides, the local transitional correspondence between wavelength and radiance within the interpolation interval is determined; Based on the local transition correspondence, calculate the derived radiance values ​​for each interpolation wavelength within the interpolation interval; The derived radiance values ​​are added to the set of discrete effective spectral points to obtain the reconstructed spectral curve.

[0016] Optionally, in one possible implementation of the first aspect, the output of the UAV matching result includes: Obtain a standard UAV spectral template library, which contains a set of standard spectral curves; Determine the spectral similarity between the target real spectral data and each of the standard spectral curves to obtain a similarity set containing multiple spectral similarity values; Extract the maximum and second-largest similarity values ​​from the similarity set; Obtain the motion trajectory features of the candidate multi-band response sequence in consecutive frames; When the difference between the maximum similarity value and the second largest similarity value is greater than the standard deviation of the similarity set, and the motion trajectory features exhibit continuous displacement and non-zero acceleration, a drone confirmation signal is generated. The drone matching result is output based on the drone confirmation signal.

[0017] A second aspect of this application provides a power transmission channel UAV target spectral adaptive matching system, comprising: The structure mask module acquires a multispectral image sequence and channel structure prior data, projects the channel structure prior data onto the multispectral image sequence, and generates a channel structure mask. The spectral stripping module extracts the motion region that deviates from the channel structure mask to obtain candidate multi-band response sequences. Based on the edge neighborhood spectrum in the channel structure mask, the module performs structural reflectance stripping on the candidate multi-band response sequences to obtain structural stripped spectral data. The spectral evaluation module determines the penetration reliability of each spectral band based on the structure stripping spectral data and generates reliable spectral band distribution data. The target matching module uses the reliable distribution data of the spectral band to generate the true spectral data of the target from the spectral data stripped from the structure, and outputs the UAV matching result.

[0018] The adaptive spectral matching method and system for UAV targets in power transmission channels provided in this application have the following advantages: This application does not directly compare a pre-defined fixed spectral band combination with a single target spectral template. Instead, it projects prior channel structure data onto a multispectral image sequence to generate a channel structure mask. Based on the neighborhood spectrum of the mask edge, it performs structural reflection stripping on the candidate multispectral band response sequence to obtain pure structural stripped spectral data. Then, it dynamically selects effective spectral bands based on the penetration reliability of each spectral band and reconstructs the true target spectral data for matching. This eliminates noise interference from invalid spectral bands caused by differences in the penetration ability of different wavelengths of light under dynamic weather conditions such as fog or backlight. It also removes spurious responses caused by local reflections or hot spots on conductor edges and fittings, avoiding misjudging environmentally influenced structural backgrounds as UAV targets. Furthermore, the reliable distribution data of spectral bands generated based on real-time penetration reliability can appropriately reflect the true imaging quality of each spectral band under the current observation environment. This ensures that the reconstruction of the true target spectral data relies only on high-confidence effective spectral bands, thereby preventing erroneous samples from being continuously written into the historical template library and improving the stability and accuracy of matching and identification during long-term continuous monitoring of UAV targets in complex environments of UHV transmission channels. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the adaptive matching method for the UAV target spectrum in power transmission channels provided in this application embodiment; Figure 2 This is an application scenario diagram of the UAV target spectrum adaptive matching method for power transmission channels provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the UAV target spectral adaptive matching system for power transmission channels provided in the embodiments of this application; Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] The technical solutions of this application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] See Figure 1This is a flowchart of the UAV target spectrum adaptive matching method for power transmission channels provided in the embodiments of this application. Figure 1 The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. Steps S1 to S4 are detailed as follows: S1: Acquire multispectral image sequences and channel structure prior data, project the channel structure prior data onto the multispectral image sequences, and generate a channel structure mask; In some embodiments, step S1, which involves acquiring the multispectral image sequence and channel structure prior data, and projecting the channel structure prior data onto the multispectral image sequence to generate a channel structure mask, encompasses three sub-steps, S11 to S13: S11: Obtain the current observation attitude and focal length parameters of the monitoring equipment, and project the three-dimensional structural wireframe in the prior data of the channel structure onto the two-dimensional imaging plane of the multispectral image sequence based on the current observation attitude and focal length parameters to obtain the structural projection contour.

[0023] Specifically, the processing unit reads the current observation attitude from the gimbal control module of the monitoring equipment and the focal length parameters from the lens drive module. Then, it reads the three-dimensional structural wireframe stored in the prior data of the channel structure, uses the pinhole imaging model to project each spatial point in the three-dimensional structural wireframe onto the two-dimensional imaging plane, and connects the projected two-dimensional points to obtain the structural projection contour.

[0024] It should be noted that the current observation attitude encompasses the horizontal deflection angle and vertical pitch angle of the gimbal, the focal length parameter reflects the current optical scaling state of the lens, and the prior data of the channel structure is a set of three-dimensional spatial coordinates of conductors, towers, insulators, and fittings within the transmission channel, obtained in advance through LiDAR scanning or engineering drawing modeling. The three-dimensional structural wireframe is a geometric framework formed by connecting the three-dimensional spatial coordinates, and the structural projection profile is the two-dimensional projection trajectory of the three-dimensional structural wireframe on the camera's imaging plane. Since the monitoring equipment is installed on a tower or specific platform, the observation attitude and focal length will continuously change with the rotation and zooming of the gimbal. Therefore, the projection relationship corresponding to each frame of image is different. Projection calculations must be performed based on the real-time attitude and focal length to ensure that the projected profile accurately coincides with the actual structural position in the current image, providing a geometric reference for generating an accurate structural mask.

[0025] Ideally, by reading the gimbal attitude and lens focal length in real time and performing frame-by-frame projection calculations, it is possible to compensate for minor attitude shifts caused by gimbal rotation, lens zoom, and wind vibration, ensuring that the projection position of the 3D structure wireframe in the 2D image is strictly aligned with the imaging position of the actual conductor and tower, thus avoiding the problem of target mis-elimination caused by incomplete or excessive mask coverage due to projection deviation.

[0026] For example, in a UHV transmission channel monitoring scenario, the monitoring equipment is installed at the crossarm of the tower. The pan-tilt unit has a horizontal deflection angle of 35 degrees north of east and a vertical pitch angle of 12 degrees downward. The focal length is 50 mm. The prior data of the channel structure stores the three-dimensional coordinates of the suspension points at both ends of a conductor. The processing unit uses a pinhole imaging model to project the two suspension points onto an imaging plane with a resolution of 1920×1080, obtains the corresponding two-dimensional pixel coordinates, and connects them to form the structural projection outline of a conductor segment.

[0027] Furthermore, when the monitoring device is equipped with multiple sensors with different field of view, the processing unit reads the intrinsic parameter matrix of each sensor and independently performs projection calculation of the three-dimensional structural wireframe for the imaging plane of each sensor to obtain the structural projection contour corresponding to each sensor.

[0028] S12: Generate an initial structural mask based on the pixel distribution of the structural projection contour in the two-dimensional imaging plane.

[0029] Specifically, the processing unit extracts the spatial depth values ​​corresponding to each structural line segment in the 3D structural wireframe, and layers the structural projection contour according to the spatial depth values ​​to obtain multiple depth-level contours. It determines the local imaging region corresponding to each depth-level contour in the 2D imaging plane, obtains the radiance difference between adjacent pixels in the local imaging region, determines the statistical mean and standard deviation of the radiance difference in the local imaging region, constructs a judgment interval with the sum of the statistical mean and standard deviation as the upper limit and the difference between the statistical mean and standard deviation as the lower limit, determines the pixel position where the absolute value of the radiance difference exceeds the judgment interval as the actual structural boundary, updates the set of projection boundary pixels corresponding to the depth-level contours in the 2D imaging plane to the pixel positions corresponding to the actual structural boundaries to obtain the corrected layer contours, and combines the corrected layer contours corresponding to each depth-level contour to generate the initial structural mask.

[0030] It should be noted that the spatial depth value is the straight-line distance from each point in the structural line segment to the monitoring device; the depth-level contour is a set of multiple projections formed by dividing the structural projection contour according to distance; the local imaging area is the pixel range occupied by the depth-level contour in the image; the actual structural boundary is the set of real structural edge pixels determined by the location of abrupt changes in radiation brightness in the image; the corrected-level contour is the structural contour after correction by actual image data; and the initial structural mask is a binary region marker map formed by combining all corrected-level contours. Since the contour obtained by projecting the 3D structural wireframe is only the geometric theoretical position, the boundaries of wires and fittings in actual imaging will be offset by factors such as illumination direction, surface material, and atmospheric scattering. Simply relying on theoretical projection can easily leave uncovered structural pixels at the mask edge or mistakenly cover the target space. Therefore, it is necessary to perform pixel-level correction on the theoretical projection boundary in combination with the actual radiation brightness distribution in the image so that the initial structural mask can accurately fit the imaging contour of the real structure.

[0031] Ideally, by searching for locations of abrupt changes in radiance near the theoretical projection contour to determine the actual structural boundary, geometric prior knowledge can be organically combined with real image data. This eliminates projection position deviations caused by atmospheric refraction, lens distortion, and differences in surface reflection, reducing the edge accuracy of the initial structural mask from a few pixels of theoretical projection error to the sub-pixel level.

[0032] Specifically, when determining the statistical mean and standard deviation of the radiance difference, the processing unit traverses all adjacent pixel pairs within the local imaging region and calculates the absolute value of the radiance difference between each pair of adjacent pixels. After summing all the absolute values ​​of the difference, it calculates the arithmetic mean to obtain the statistical mean, and calculates the dispersion of all the absolute values ​​of the difference from the statistical mean to obtain the standard deviation. When the absolute value of the radiance difference of a certain pair of adjacent pixels exceeds the judgment interval, it is determined that there is a sudden change in radiance at the location of the current pixel pair, and the pixel in the current pixel pair that is closer to the structure projection outline is marked as the actual structure boundary pixel.

[0033] Ideally, by replacing the fixed threshold judgment with a dynamic judgment interval composed of the statistical mean and standard deviation, the boundary detection logic can automatically adapt to the differences in radiation brightness distribution under different light intensities and weather conditions, ensuring accurate positioning of structural boundaries in various scenarios such as backlight, cloudy days, or smog.

[0034] For example, in a local imaging region, the processing unit traverses 200 pairs of adjacent pixels to obtain the absolute values ​​of the difference in radiance. The arithmetic mean is calculated to obtain a statistical mean of 12, and the dispersion is calculated to obtain a standard deviation of 8. The upper limit of the judgment interval is 20 and the lower limit is 4. When the absolute value of the difference in radiance of an adjacent pixel pair is 35, it exceeds the upper limit of the judgment interval. It is determined that there is a sudden change in radiance at the location of the current pixel pair and it is marked as an actual structural boundary pixel. When the absolute value of the difference in radiance of another adjacent pixel pair is 15, it is within the judgment interval. It is determined that the current pixel pair is located in a uniform region and is not regarded as a structural boundary.

[0035] Furthermore, when the distribution of radiance difference in a local imaging area exhibits a bimodal state, the processing unit sorts the absolute values ​​of the radiance difference by size, takes the absolute values ​​of the radiance difference in the first quarter after sorting, and recalculates the statistical mean and standard deviation. The recalculated statistical mean and standard deviation are used to construct a new judgment interval to determine the actual structural boundary, thus avoiding the uniform noise in the low-contrast area from lowering the overall statistic and causing boundary missed detection.

[0036] S13: Extract the edge pixel set of the initial structural mask, determine the number of tolerance pixels to be extended outward based on the wind deflection amplitude of the transmission line and the calibration error of the monitoring equipment, push the edge pixel set outward along the normal direction by the number of tolerance pixels to generate the structural extension region, and supplement the structural extension region to the outer edge of the initial structural mask to obtain the channel structural mask.

[0037] Specifically, the processing unit extracts the edge pixel set of the initial structural mask, reads the wind deflection amplitude parameter of the transmission line and the calibration error parameter of the monitoring equipment, determines the number of tolerance pixels to be extended outward based on the wind deflection amplitude parameter and the calibration error parameter, calculates the normal direction of each edge pixel in the edge pixel set, pushes each edge pixel outward along the normal direction by the number of tolerance pixels to obtain the extended pixel position set, and uses the extended pixel position set as the structural extension region to supplement the outer edge of the initial structural mask to obtain the channel structural mask.

[0038] It should be noted that the edge pixel set is the sequence of boundary pixels directly adjacent to the outer pixels of the initial structural mask; the wind deflection amplitude parameter is the maximum lateral displacement of the power transmission line from its stationary position under wind force; the calibration error parameter is the residual deviation between the gimbal coordinate system and the real geographic coordinate system during the installation of the monitoring equipment; the tolerance pixel count is the number of pixels after converting the offset distance in physical space to the imaging plane; the normal direction is the direction of the outward normal of the mask boundary at the edge pixels; and the channel structure mask is the final mask after edge expansion of the initial structural mask. Because the power transmission line will exhibit significant lateral sway in strong winds, and residual errors are unavoidable in the installation and calibration of the monitoring equipment, if the mask only covers the theoretical stationary position without leaving an expansion margin, the wind-deflected line will overflow the mask boundary, becoming residual interference and directly affecting the purity of moving target extraction. Therefore, it is necessary to expand the mask edge outward according to the actual wind deflection amplitude and calibration error to ensure that the channel structure mask can completely cover all possible structural positions.

[0039] Preferably, by jointly converting the wind deflection amplitude and calibration error into the number of tolerance pixels of the imaging plane and performing edge expansion along the normal direction, it is possible to avoid over-coverage caused by uniform expansion of the entire image while ensuring that the mask completely covers the conductor under wind deflection conditions. This allows the expansion amount of the mask edge to accurately match the physical offset requirements, thus preserving the effective spatial area available for target detection to the greatest extent.

[0040] Specifically, the process for determining the number of tolerance pixels is as follows: The processing unit reads the wind deflection amplitude parameter calculated from the maximum design wind speed and conductor tension parameters in the transmission line design document, and reads the calibration error parameter measured by the laser calibrator during installation. In the physical space, the larger value between the wind deflection amplitude parameter and the calibration error parameter is selected. Based on the current focal length parameter and the pixel size of the imaging sensor, the larger value is converted from physical distance into the number of pixels to obtain the number of tolerance pixels.

[0041] Ideally, by selecting the larger value of wind deviation and calibration error in physical space and then converting it into the number of pixels, it can be ensured that the expansion amount always covers the most unfavorable offset situation, and avoid the mask still having a coverage gap when extreme wind deviation and calibration deviation occur at the same time.

[0042] For example, in a monitoring scenario of an ultra-high voltage transmission channel, the wind deflection amplitude parameter of the transmission line is 0.3 meters and the calibration error parameter of the monitoring equipment is 0.05 meters. The larger of the two values ​​is 0.3 meters. The current focal length parameter is 50 millimeters and the pixel size of the imaging sensor is 5 micrometers. According to the optical imaging ratio, the physical offset of 0.3 meters is converted into 30 pixels on the imaging plane. The number of tolerance pixels is determined to be 30. The processing unit pushes the edge pixel set of the initial structure mask outward by 30 pixels along the normal direction to generate the structure extension area. The structure extension area is added to the outer edge of the initial structure mask to obtain the channel structure mask.

[0043] Furthermore, when there are multiple layers of conductors at different heights in the transmission channel, the processing unit calculates the number of tolerance pixels corresponding to each layer of conductors based on the height difference of each layer. Since the imaging ratio of the conductors near the ground is larger and the imaging ratio of the conductors far away is smaller, the number of tolerance pixels of each layer of conductors is different. The processing unit performs edge expansion according to the actual imaging ratio of each layer of conductors to generate a non-uniformly expanded channel structure mask.

[0044] S2: Extract the motion region that deviates from the channel structure mask to obtain candidate multi-band response sequences. Perform structural reflection stripping on the candidate multi-band response sequences based on the edge neighborhood spectrum in the channel structure mask to obtain structural stripping spectral data. In some embodiments, step S2 involves extracting the motion region deviating from the channel structure mask to obtain candidate multi-band response sequences. The candidate multi-band response sequences are then subjected to structural reflectance stripping based on the edge neighborhood spectra in the channel structure mask to obtain structurally stripped spectral data. This process encompasses two sub-steps, S21 and S22. S21: Extract the motion region that deviates from the channel structure mask and generate candidate multi-spectral response sequences.

[0045] It should be noted that the UHV transmission channels are densely packed with conductors and fittings. The strong structural background is prone to localized high-brightness reflections under illumination. Relying solely on single-frame image detection can easily lead to the misidentification of stationary towers or insulators as targets. Step S21 separates real moving targets from background artifacts, and preliminarily purifies the candidate target set from both spatial and temporal dimensions, thus solving the problem of large-area false detection caused by fixed structural backgrounds.

[0046] S22: Based on the edge neighborhood spectrum in the channel structure mask, perform structural reflection stripping on the candidate multi-band response sequence to obtain structural stripping spectral data.

[0047] It should be noted that the reflectivity of the wire and the edge of the fitting decreases rapidly with increasing distance. Step S22 accurately estimates the amount of background reflection interference at the location of the moving target by constructing the edge neighborhood band and quantifying the decreasing law of radiation brightness. The estimated interference is removed from the original response of the target, and the pseudo-spectral features brought by the strong structure background are stripped away, thus solving the technical problem of wire reflection obscuring the real material information of the UAV under complex lighting.

[0048] In some embodiments, step S21, which involves extracting the motion region deviating from the channel structure mask and generating candidate multi-spectral response sequences, encompasses four sub-steps from S211 to S214: S211: Determine the difference in radiance at the same pixel coordinates between adjacent frames in the multispectral image sequence to obtain a motion difference image. Remove the pixels covered by the channel structure mask from the motion difference image to obtain the residual motion region.

[0049] Specifically, the processing unit reads data from two adjacent frames in the multispectral image sequence and calculates the difference in radiant energy for pixels with the same row and column indices. It then summarizes the differences to construct a two-dimensional matrix. Subsequently, it reads the channel structure mask and masks the fixed structure positions marked by the mask in the two-dimensional matrix to retain the non-zero difference region and obtain the residual motion region.

[0050] It should be noted that the motion difference image reflects the changes in scene radiation energy under adjacent time slices. The residual motion region is the set of pixels of suspected moving targets retained after removing known fixed structures. Since fixed structures occupy most of the image in the UHV channel, directly finding small targets is computationally expensive. By combining inter-frame difference with prior masks, the region where actual displacement occurs is locked, and the static background and potential moving targets are initially separated. This reduces the computational resource consumption of the full-image search and eliminates artifact interference caused by changes in lighting for fixed facilities such as towers and insulators.

[0051] Specifically, sensor noise interference exists when calculating the difference in radiated energy. Therefore, a noise tolerance parameter is set. When the absolute value of the energy difference between pixels at the same position in adjacent frames is less than the noise tolerance parameter, it is determined that the current pixel has not moved and the difference is forcibly reduced to zero. The noise tolerance parameter is derived from the dark current noise variance in the multispectral camera's factory calibration document. When the camera gain is increased in low-light conditions at night, resulting in amplified noise, the noise tolerance parameter corresponding to the current gain level is automatically retrieved.

[0052] Ideally, matching the noise margin parameter based on the camera gain state can overcome the problem of thermal noise amplification caused by sensor gain increase in nighttime or low-light environments, avoid misjudging the inherent electronic noise of the sensor as real physical moving targets, and maintain the purity of motion detection in low-light scenes.

[0053] For example, in a monitoring scenario of an ultra-high voltage transmission channel, a multispectral camera acquires images at a frequency of 30 frames per second. The difference in radiance at the same pixel coordinate is calculated by extracting the images of the 10th and 11th frames and setting the noise floor tolerance parameter of the current environment to 5. If the radiance of a certain pixel is 120 in the 10th frame and 125 in the 11th frame, the absolute value of the difference is 5, which is equal to the noise floor tolerance parameter. Therefore, the current pixel is determined to be noise and is reset to zero. If the absolute value of the difference of another pixel is 30, which is greater than the noise floor tolerance parameter, it is retained as a valid pixel in the motion difference image. Subsequently, the channel structure mask covers the tower area in the center of the image and removes all valid pixels in the tower area. Finally, the residual motion area is retained in the vegetation area at the edge of the image and the suspected flying object area in the air.

[0054] Furthermore, if the monitoring device is equipped with an event camera or dynamic vision sensor, it can directly read the asynchronous event stream output by the sensor and project it onto a two-dimensional plane to generate an event density map. Combined with the channel structure mask to shield the events in the fixed structure area, the residual motion area can be directly obtained, thus eliminating the need for inter-frame difference calculation.

[0055] S212: Extract the set of spatially adjacent candidate pixels in the residual motion region, obtain the centroid position sequence of the candidate pixel set in multiple consecutive frames of images, and determine the displacement direction between adjacent frames in the centroid position sequence.

[0056] Specifically, the processing unit performs connected component labeling on the residual motion region to aggregate spatially adjacent pixels with similar radiance differences into a candidate pixel set. Then, it calculates the weighted composite value of all pixel coordinates in the candidate pixel set to obtain the centroid coordinates of the current frame. It also tracks the same candidate pixel set in multiple consecutive frames, records the centroid coordinates of each frame to form a centroid position sequence, and finally calculates the vector difference of the centroid coordinates of adjacent frames to obtain the displacement direction.

[0057] It should be noted that the candidate pixel set is a spatially connected pixel block with similar grayscale features within the residual motion region. The centroid position sequence is the geometric center coordinate trajectory of the candidate pixel set in the time dimension. The displacement direction is the vector direction of the line connecting the geometric centers between adjacent time nodes. Since intrusion targets such as drones usually occupy multiple pixels in an image, relying solely on the changes of a single pixel is easily affected by environmental disturbances. However, the centroid trajectory can reflect the physical displacement law of the target, transforming discrete pixel-level changes into a macroscopic motion trajectory at the target level. This provides a data basis for judging the target's motion intention and overcomes pixel-level false detections caused by local occlusion or sensor defects in a single frame image.

[0058] Specifically, the connected component marking process requires setting a spatial adjacency radius parameter and a radiance difference tolerance parameter. The spatial adjacency radius parameter is used to determine whether pixels are adjacent, and the radiance difference tolerance parameter is used to determine whether pixel grayscale is similar. The spatial adjacency radius parameter is set to a distance of 2 pixels, and the radiance difference tolerance parameter is set to 10% of the global radiance standard deviation of the current frame.

[0059] Ideally, automatically relaxing the aggregation conditions of spatial adjacency and grayscale difference in hazy or low-contrast weather can prevent pixels of the same physical target from being incorrectly segmented into multiple isolated fragments due to edge blurring, ensuring the integrity and continuity of the target-level centroid trajectory and improving the anti-interference capability of moving target extraction under adverse weather conditions.

[0060] For example, after identifying a candidate pixel set covering 50 pixels, the sum of the horizontal and vertical coordinates of the 50 pixels is calculated, and the ratio of the sum to 50 is calculated respectively. The centroid coordinates of the current frame are obtained as horizontal coordinate 450 and vertical coordinate 320. In the next frame, the candidate pixel set moves and the centroid coordinates are recalculated to obtain horizontal coordinate 455 and vertical coordinate 322. The vector difference between the centroids of the two frames is calculated to determine the displacement direction as downward to the right. The process is repeated for 10 frames to record 10 centroid coordinates, thus forming a complete centroid position sequence.

[0061] Furthermore, the optical flow method is used to calculate the motion vector field of each pixel in the residual motion region and to perform cluster analysis on the motion vector field. Pixel regions with similar motion vector directions and sizes are directly divided into candidate pixel sets, and the centroid position sequence of each cluster region is calculated.

[0062] S213: Determine the change in displacement direction between consecutive frames and the trajectory state of the centroid position sequence, and filter the set of valid moving pixels.

[0063] Specifically, the processing unit analyzes the angle change of the displacement direction between adjacent frames in the centroid position sequence and calculates the straight-line distance between the start and end points in the centroid position sequence as well as the total path length of the centroid position sequence. Then, it calculates the ratio of the total path length to the straight-line distance and filters the effective set of moving pixels based on the values ​​of the angle and the ratio.

[0064] It should be noted that the effective set of moving pixels is a set of candidate pixels with a clear unidirectional or smooth curve displacement trend and no periodic reciprocating motion characteristics. Since drone intrusions have a clear destination or cruise route, and natural environmental disturbances are affected by wind or water flow, they manifest as periodic back-and-forth movements near the original position. Therefore, distinguishing between drones with autonomous flight capabilities and environmental disturbances such as tree branches swaying in the wind and water ripples can eliminate false alarms caused by vegetation disturbances and water surface reflections within the UHV channel, thus improving the purity of candidate targets.

[0065] Specifically, the directional angle between the displacement directions of adjacent frames is calculated. When the directional angle is less than the preset directional deflection angle for multiple consecutive frames and the ratio of the total path length to the straight-line distance is less than the preset trajectory tortuosity, the candidate pixel set is determined to be a unidirectional smooth motion and is identified as a valid moving pixel set. When the directional angle is frequently greater than the preset directional deflection angle between consecutive frames and the ratio of the total path length to the straight-line distance is greater than the preset trajectory tortuosity, the candidate pixel set is determined to be in a state of back-and-forth movement along the same trajectory and is identified as a pseudo-motion region caused by environmental disturbance and is removed from the residual motion region. When the directional angle is less than the preset directional deflection angle in some frames and greater than the preset directional deflection angle in some frames, and the ratio of the total path length to the straight-line distance is between the two, the candidate pixel set is determined to be in a maneuvering turning state and is retained as a pending moving pixel set to extend the tracking frame number for secondary confirmation. The preset directional deflection angle is set to 45 degrees and the preset trajectory tortuosity is set to 1.5.

[0066] Ideally, the threshold for automatically adjusting the trajectory curvature under extreme weather conditions such as strong winds can be adapted to the physical laws of large-scale vegetation swaying, preventing the misjudgment of thick tree branches or long floating objects affected by strong winds as drones performing maneuvers, thus reducing the false alarm rate caused by complex natural environments.

[0067] For example, tracking a candidate pixel set for 15 frames and calculating that the angle between the displacement directions of adjacent frames is within 10 degrees, while calculating the total path length as 150 pixels and the straight-line distance between the start and end points as 140 pixels, the ratio result is 1.07, which is less than 1.5, thus determining that the candidate pixel set is a unidirectional smooth motion and confirming it as a valid moving pixel set; while another candidate pixel set has an angle between 0 degrees and 170 degrees within 10 frames, with a total path length of 80 and a straight-line distance between the start and end points as 5, the ratio result reaches 16, which is greater than 1.5, thus determining that the other candidate pixel set is a tree branch swaying and eliminating it.

[0068] Furthermore, an acceleration calculation logic is introduced to calculate the rate of change of displacement velocity between adjacent frames in the centroid position sequence. When the rate of change of velocity remains within the non-zero range and the displacement direction does not reverse 180 degrees, it is directly determined as a valid set of moving pixels.

[0069] S214: Take the effective set of moving pixels as an independent moving region, determine the coordinates of the corresponding pixels in each spectral band of the multispectral image sequence for the independent moving region, extract the radiance value at the corresponding pixel coordinates, and generate a candidate multispectral response sequence.

[0070] Specifically, the processing unit projects the bounding box of the effective set of moving pixels onto each spectral channel of the multispectral image sequence and reads the pixel radiance value of each spectral channel within the bounding box. Then, the radiance values ​​are arranged in order of wavelength from shortest to longest to generate a candidate multispectral response sequence.

[0071] It should be noted that the independent motion region is a clean target region verified by spatiotemporal consistency, and the candidate multispectral response sequence is a set of radiation energy responses of the target at different wavelengths. Since multispectral cameras are usually composed of multiple sensors of different bands stitched together or split, there may be slight translation or rotation deviations between images of different bands. Directly extracting the co-position coordinates can force alignment and eliminate the registration error caused by physical hardware, obtain the spectral fingerprint of the target under the absorption and reflection laws of different materials, provide raw data for material identification and background stripping, ensure that the extracted spectral data is strictly aligned in space, and avoid spectral misalignment caused by parallax between different sensors of the multispectral camera.

[0072] Specifically, to address the sub-pixel level registration deviation between images in different spectral bands of the multispectral camera, registration compensation parameters are set using the multispectral calibration matrix provided at the camera's factory. When extracting the radiance values ​​of each spectral band, instead of directly reading the pixel values ​​at integer coordinates, a bilinear interpolation algorithm is used based on the registration compensation parameters to calculate the radiance values ​​at sub-pixel positions in order to eliminate spectral distortion caused by registration deviation.

[0073] Ideally, introducing registration compensation parameters provided by the multispectral calibration matrix for sub-pixel level interpolation calculation can eliminate the slight parallax caused by physical splicing or beam splitting of sensors in different bands, ensuring that the spectral response at the same physical location in different spectral bands has a very high spatial overlap, laying a data foundation for high-precision material identification.

[0074] For example, after determining the center coordinates of the independent motion region as x-coordinate 500 and y-coordinate 400, the radiance values ​​of the visible light blue band, green band, red band, and near-infrared band near the center coordinates are read. Since there is a 2-pixel lateral registration deviation in the near-infrared band, the radiance value of the near-infrared band is read at the position of x-coordinate 502 and y-coordinate 400 according to the registration compensation parameter, and the radiance values ​​of the four bands are combined to generate a candidate multi-spectral response sequence covering the four values.

[0075] Furthermore, if the multispectral image sequence is acquired by a single sensor combined with a rotating filter wheel, the actual position of the target at the time of image capture for each spectral band is calculated based on the time difference of the filter wheel rotation and the displacement velocity of the independent moving area, and the radiance value at the corrected co-position pixel coordinates is extracted.

[0076] In some embodiments, step S22 involves structural reflectance stripping of the candidate multi-band response sequence based on the edge neighborhood spectrum in the channel structure mask to obtain structural stripped spectral data, encompassing four sub-steps from S221 to S224: S221: Starting from the boundary of the channel structure mask, extend the adjacent pixel rows to the outside of the channel structure mask to obtain the edge neighborhood band. Extract the spectral response values ​​of each pixel in the edge neighborhood band in the multispectral image sequence to generate neighborhood spectral distribution data.

[0077] Specifically, the processing unit performs a morphological dilation operation on the channel structure mask and sets the dilation structure element to a circular matrix. Then, it calculates the difference between the dilated region and the original channel structure mask to obtain the edge neighborhood band, and extracts the radiance values ​​of all pixels in each spectral band to generate neighborhood spectral distribution data.

[0078] It should be noted that the edge neighborhood zone is a ring-shaped or strip-shaped area immediately outside the background of a strong structure. The neighborhood spectral distribution data is a set of data reflecting the attenuation law of background reflection in space. Since the reflective intensity of wires or hardware decreases rapidly and non-linearly with increasing distance, only the reflective law of the edge region immediately adjacent to the target has a physical correlation with the reflective law of the target's location. Therefore, collecting actual attenuation samples of background reflection can provide data support for building a background interference subtraction model. By using the background region data immediately adjacent to the target to reflect the intensity of background reflection interference at the target's location, we can avoid over-subtraction or under-subtraction caused by using the global background.

[0079] Specifically, the expansion radius parameter of the morphological dilation operation is set to 3 to 5 pixels, and the pixel row is expanded outward according to the expansion radius parameter. The numerical range of the expansion radius parameter is statistically derived based on the typical resolution of the UHV channel monitoring camera and the actual diffusion width of the conductor edge halo on the sensor.

[0080] Ideally, automatically expanding the search range of the edge neighborhood zone under hazy or strong light scattering weather conditions can ensure that the collected neighborhood spectral distribution data completely covers the structural reflective halo area that becomes larger due to increased light scattering, prevent interference subtraction residue caused by incomplete background sampling, and improve the completeness of spectral stripping under complex weather conditions.

[0081] For example, an expansion operation is performed on the channel structure mask and the expansion radius parameter is set to 4 pixels so that the expanded region covers the wire and the area outside the wire with a width of 4 pixels. The difference between the expanded region and the original wire mask is calculated to obtain an edge neighborhood band with a width of 4 pixels. Then, the radiance values ​​of all pixels in the edge neighborhood band in the near-infrared band are extracted to generate neighborhood spectral distribution data covering hundreds of values.

[0082] Furthermore, instead of using a fixed morphological dilation, the normal direction of the channel structure mask boundary is calculated and a specific number of pixels are sampled outward at equal intervals along the normal direction. The sampled points are then connected to form an edge neighborhood band.

[0083] S222: Divide the edge neighborhood into pixel sub-bands arranged along the normal direction of the channel structure mask boundary, obtain the radiance value of the center pixel in each pixel sub-band under a single spectral band, and determine the negative correlation between radiance and normal distance extension.

[0084] Specifically, the processing unit calculates the shortest distance from each pixel in the edge neighborhood to the boundary of the channel structure mask and divides the edge neighborhood into multiple distance gradients based on the shortest distance to form multiple pixel sub-bands. Then, it calculates the median of the radiance values ​​of all pixels in each pixel sub-band as the representative radiance value of the current pixel sub-band. Finally, it constructs a curve with the normal distance as the independent variable and the representative radiance value as the dependent variable to obtain a negative correlation relationship.

[0085] It should be noted that a pixel sub-band is a set of pixels within the edge neighborhood that are equidistant from the mask boundary. The negative correlation is a mathematical mapping that describes the decrease in radiance with increasing distance. In a small local area, the reflection of wires or hardware changes little in the direction parallel to the boundary but changes drastically in the direction perpendicular to the boundary normal. Dividing the sub-band along the normal direction can extract the attenuation gradient of the reflection, quantify the specific law of background reflection attenuation with distance, and provide a basis for calculating the amount of reflection interference at the target position. It simplifies the complex two-dimensional spatial reflection distribution into a one-dimensional distance attenuation model, reduces the computational complexity of background subtraction, and preserves the physical characteristics of reflection attenuation.

[0086] Specifically, the analysis examines the trend of the representative radiance value as the normal distance increases. When the representative radiance value shows a monotonically decreasing trend with increasing normal distance and the decrease rate gradually slows down, the current edge neighborhood is determined to be a pure structural reflective region, and an exponential decay model is used to construct a negative correlation. When the representative radiance value shows a monotonically decreasing trend with increasing normal distance and the decrease rate remains constant, the current edge neighborhood is determined to be a uniform scattering region, and a linear decay model is used to construct a negative correlation. When the representative radiance value shows a fluctuating state or a non-monotonic change state with increasing normal distance, the current edge neighborhood is determined to be interfered with by other independent light sources or moving objects. The data of the current edge neighborhood is discarded, and the search range is expanded to find a pure background region to reconstruct the edge neighborhood. The tolerance parameter for determining whether the decrease rate is constant is set to 5%, that is, when the difference in the decrease rate of adjacent sub-bands is less than 5%, the decrease rate is considered constant.

[0087] Ideally, the introduction of constant radiance decay rate verification and non-monotonic fluctuation interception logic can automatically identify and eliminate transient light source interference such as vehicle high beams sweeping across the conductor at night or birds skimming over the edge of the conductor, ensuring that the constructed background subtraction model only reflects the inherent physical reflection law of the conductor or fitting, thereby guaranteeing the purity of the structural stripping spectral data.

[0088] For example, the edge neighborhood band with a width of 4 pixels is divided into 4 pixel sub-bands and the representative radiance values ​​of the 4 sub-bands in the near-infrared band are calculated to be 800, 600, 450 and 400 respectively. It is found that the radiance decreases monotonically with increasing distance and the rate of decrease gradually slows down. Therefore, an exponential decay model is used to construct a negative correlation relationship.

[0089] Furthermore, instead of calculating the median, the average band weight of the radiance value within each pixel sub-band is calculated. The weight is set inversely based on the distance from the pixel to the mask boundary, with the weight increasing as the distance increases. The aforementioned band weight calculation emphasizes the reflective characteristics closer to the boundary.

[0090] S223: Determine the radiance drop amplitude between adjacent pixel sub-bands based on the negative correlation correspondence, combine the radiance drop amplitude of each single spectral band to generate a radiance decrease distribution state, and construct edge radiance subtraction reference data based on the radiance decrease distribution state.

[0091] Specifically, the processing unit calculates the difference between the dependent variables corresponding to adjacent independent variable nodes in the negative correlation relationship to obtain the radiance reduction amplitude of a single spectral band. Then, it arranges the radiance reduction amplitudes of all spectral bands in wavelength order to generate a radiance reduction distribution state, and discretizes the negative correlation relationship to generate a lookup table indexed by the normal distance and containing the radiance deduction value to obtain edge radiance deduction reference data.

[0092] It should be noted that the radiance drop amplitude is the difference in reflected light intensity between adjacent distance gradients, the radiance decrease distribution is a global description of the attenuation law of all spectral bands, and the edge radiance subtraction reference data is a lookup table used to query the background interference at a specific distance. In video stream processing, each pixel needs to undergo background subtraction. If an exponential or linear model is called for calculation every time, it will consume a lot of processor resources. Converting the continuous attenuation model into a discrete data structure that is easy for computers to query quickly can improve the execution efficiency of background subtraction. By constructing a lookup table, the subtraction operation only requires simple index lookup and difference calculation, avoiding complex real-time calculations to meet the real-time processing requirements of UAV monitoring equipment.

[0093] Specifically, the sampling step size parameter of the discretization sampling is set to 1 pixel distance so that the lookup table covers the deduction value corresponding to each integer pixel distance. For moving targets with non-integer pixel distances, a linear interpolation algorithm is used and the accurate deduction value of the current non-integer distance is calculated based on the deduction values ​​of two adjacent integer distances in the lookup table.

[0094] Ideally, transforming the continuous attenuation model into a discretized lookup table structure can improve computational efficiency with a smaller memory cost, thereby reducing the time consumed by the processing unit for background subtraction in a single operation when processing high-resolution multispectral video streams, which meets the computing power requirements of edge computing equipment for real-time alarms in UHV transmission channels.

[0095] For example, the radiance deduction values ​​at distances of 1, 2, 3 and 4 pixels from the mask boundary are calculated to be 200, 150, 100 and 80 respectively based on the exponential decay model of the near-infrared band, and 200, 150, 100 and 80 are stored in the near-infrared band lookup table of the edge radiance deduction reference data.

[0096] Furthermore, instead of constructing a one-dimensional lookup table, the negative correlation correspondence is projected outward along the channel structure mask boundary to generate a two-dimensional radiance subtraction matrix with the same size as the original image. During subtraction, the matrix value of the coordinates of the moving target is directly read for subtraction to save the distance calculation step.

[0097] S224: Calculate the normal distance of each moving pixel in the candidate multi-band response sequence from the boundary of the channel structure mask. Match the corresponding radiance subtraction value in the edge radiance subtraction reference data according to the normal distance. Subtract the radiance subtraction value from the corresponding moving pixel in the candidate multi-band response sequence to obtain the structure stripping spectral data.

[0098] Specifically, the processing unit calculates the shortest distance from each pixel in the independent motion region to the boundary of the channel structure mask to obtain the normal distance. Then, using the normal distance as an index, it queries the corresponding radiance subtraction value in the edge radiance subtraction reference data and obtains the difference between the original radiance value and the radiance subtraction value of the corresponding moving pixel in each spectral band in the candidate multi-spectral response sequence to obtain the structure stripping spectral data.

[0099] It should be noted that the normal distance is the shortest vertical distance from the moving pixel to the nearest mask boundary, and the structural stripping spectral data is the pure target spectrum after removing background reflection interference. Since the amount of background reflection interference on the target pixel depends on the physical distance from the reflective light source, pixel-level interference stripping can be achieved by calculating the normal distance and matching the corresponding subtraction value. This eliminates the contamination of the target spectrum by strong structural backgrounds and restores the true material reflection law of the target. It solves the problem of conductor reflections in UHV channels obscuring the true spectrum of UAVs, enabling the spectral matching algorithm to perform comparisons based on pure data to reduce the false alarm rate.

[0100] Specifically, it is determined whether the normal distance of the moving pixel exceeds the valid query range of the edge radiation subtraction reference data. When the normal distance is less than or equal to the maximum index distance of the edge radiation subtraction reference data, the corresponding radiance subtraction value is directly matched in the lookup table and the difference is calculated. When the normal distance is greater than the maximum index distance of the edge radiation subtraction reference data, it is determined that the background reflection interference of the current moving pixel has been attenuated to a negligible level. The radiance subtraction value is set to zero and the original radiance value is retained as the structure stripping spectral data. When the normal distance is less than zero, that is, the moving pixel is located inside the channel structure mask, it is determined that the current pixel is a pixel of the fixed structure itself. The current pixel is removed from the candidate multi-band response sequence and does not participate in the matching. The maximum index distance of the edge radiation subtraction reference data is set to the outer boundary distance of the edge neighborhood band, which is usually 4 or 5 pixels.

[0101] Ideally, setting a maximum index distance and introducing a smooth transition branch when the normal distance exceeds the limit can handle the spectral abrupt change problem when the UAV flies at high speed over the edge of the guide line from a strong interference area into a clean airspace. This ensures that the generated structure stripping spectral data maintains a continuous and stable state in time sequence, avoiding matching failure caused by spectral feature breaks.

[0102] For example, the normal distance of a drone pixel from the boundary of the wire mask is calculated to be 2 pixels. The radiation subtraction value corresponding to a distance of 2 pixels in the near-infrared band is found to be 150 in the edge radiation subtraction reference data. The original radiation value of the drone pixel in the near-infrared band is 600. The difference between 600 and 150 is 450. 450 is used as the structural stripping spectrum data of the drone pixel in the near-infrared band. For another drone pixel with a normal distance of 6 pixels, since it exceeds the maximum index distance of 5 pixels, the radiation subtraction value of the other drone pixel is set to zero and the original radiation value is retained.

[0103] Furthermore, instead of directly performing the difference calculation, the ratio of the radiance deduction value to the original radiance value is obtained. When the ratio is greater than the preset interference ratio parameter, it is determined that the current spectral band data has been completely submerged by background reflection, and the current spectral band data is directly marked as invalid and not included in the matching.

[0104] S3: Determine the penetration reliability of each spectral band based on the structural stripping spectral data, and generate reliable spectral band distribution data; In some embodiments, step S3, which involves determining the penetration reliability of each spectral band based on the structural stripping spectral data and generating reliable spectral band distribution data, encompasses three sub-steps, S31 to S33: S31: Extract the local contrast and cross-band response difference magnitude of each spectral band in the structure stripping spectral data.

[0105] Specifically, the processing unit traverses each spectral band in the structure-exposed spectral data, extracts the target region pixels and the adjacent local background pixels for any target spectral band, calculates the absolute value of the difference in radiance between the target region pixels and the local background pixels to obtain the local contrast, then selects a reference spectral band with adjacent wavelengths, spatially aligns the pixel response values ​​of the target spectral band and the reference spectral band in the same two-dimensional coordinates, calculates the response difference between the aligned target spectral band and the reference spectral band pixel by pixel to form a response difference sequence, and calculates the difference between the maximum and minimum values ​​in the response difference sequence to obtain the cross-spectral band response difference amplitude.

[0106] It should be noted that local contrast reflects the difference in reflected energy between the target material and its surrounding environment at a single wavelength, while the amplitude of the cross-spectral response difference characterizes the degree of continuous and drastic change in the reflectance spectrum of the target material at adjacent wavelengths. Since the surface coating of a drone possesses unique absorption and reflection peaks at its inherent wavelengths, relying solely on the absolute brightness of a single spectral band is easily affected by overall illuminance. However, by combining local contrast with the response gradient of adjacent spectral bands, the inherent spectral fingerprint characteristics of the target material can be captured, providing a core basis for assessing whether the current spectral band retains true material information.

[0107] Specifically, when calculating the absolute difference between the radiance of the target region pixels and the local background pixels, the processing unit calculates the arithmetic mean of the radiance of all pixels in the target region to obtain the average radiance of the target region, and calculates the arithmetic mean of the radiance of all pixels in the local background region to obtain the average radiance of the background region. The absolute difference between the average radiance of the target region and the average radiance of the background region is then used as the local contrast ratio. When calculating the response difference between the aligned target spectral segment and the reference spectral segment pixel pixel by pixel, the processing unit uses a bilinear interpolation algorithm to align the coordinates of the reference spectral segment pixels that have sub-pixel level translational deviations, ensuring that the target spectral segment pixels and the reference spectral segment pixels participating in the comparison point to the same physical spatial location.

[0108] Preferably, by combining spatial contrast differences under a single wavelength with spectral gradient changes between adjacent wavelengths, false responses caused by sudden changes in global illumination or overall sensor gain drift can be filtered out, ensuring that the extracted spectral features reflect the physical material properties of the UAV itself and improving the robustness of material recognition under complex lighting conditions.

[0109] For example, when processing the near-infrared target spectrum, the average radiance of the target region pixels is extracted as 450 and the average radiance of the local background pixels is 200. The absolute value of the difference is 250, which is the local contrast. The short-wave infrared with the adjacent wavelength is selected as the reference spectrum. After alignment, the difference is calculated pixel by pixel to form a response difference sequence covering 100 values. The maximum value in the sequence is 85 and the minimum value is 15. The difference of 70 is the cross-spectral response difference amplitude.

[0110] Furthermore, when the number of effective pixels in the target area is reduced to less than a preset threshold due to partial occlusion by wires or tree canopies, the processing unit automatically expands the search radius of the local background pixels and uses the spectral response in the unoccluded state in historical frames to fill the missing data in the current frame, ensuring the continuity of the calculation of the difference between local contrast and cross-spectral response.

[0111] S32: Obtain the ambient light vector and fog concentration parameters during the acquisition of multispectral image sequences, determine the saturation exceedance risk of each spectral band based on the ambient light vector, and determine the scattering and occlusion degree of each spectral band based on the fog concentration parameters.

[0112] Specifically, the processing unit reads the ambient light vector and fog concentration parameters from the meteorological and light sensors of the monitoring equipment. The ambient light vector includes the solar altitude angle, azimuth angle, and direct radiation intensity, while the fog concentration parameter reflects the density of suspended particles in the atmosphere. The processing unit compares the ambient light vector with the upper limit of the spectral response of each spectral band to determine the risk of saturation exceeding the limit, and matches the fog concentration parameter with the wavelength penetration physical model of each spectral band to determine the degree of scattering and shielding.

[0113] It should be noted that the ambient light vector describes the direction and energy intensity of the light source in the current scene, the fog concentration parameter quantifies the turbidity of the atmospheric medium, the saturation exceedance risk characterizes whether the sensor has reached its photoelectric conversion limit under strong light and lost details in dark areas, and the scattering obscuration degree reflects the proportion of energy attenuation caused by light being scattered by particles when passing through fog. Because the Rayleigh and Mie scattering effects of different wavelengths of light in fog vary greatly, short-wavelength light is easily scattered and obscured while long-wavelength light has stronger penetrating power. Therefore, it is necessary to combine real-time meteorological and lighting data to quantify the environmental physical interference experienced by each spectral band, providing an objective basis for eliminating failed spectral bands.

[0114] Specifically, when determining the risk of saturation exceeding the threshold, the processing unit checks whether the direct radiation intensity is greater than a preset radiation threshold and whether the solar altitude angle is within the lens's field of view. If the direct radiation intensity is greater than the preset radiation threshold and the solar altitude angle is within the lens's field of view, the saturation exceeding risk for the corresponding short-wavelength spectral band is determined to be high-risk; otherwise, it is determined to be low-risk. When determining the degree of scattering obstruction, the processing unit checks whether the fog concentration parameter is greater than a preset concentration threshold. If the fog concentration parameter is greater than the preset concentration threshold, the scattering obstruction degree for each spectral band is calculated based on the Rayleigh scattering model. The shorter the wavelength, the greater the scattering obstruction value. If the fog concentration parameter is less than or equal to the preset concentration threshold, the scattering obstruction degree for each spectral band is determined to be zero. The preset radiation threshold and preset concentration threshold are derived from historical meteorological statistics of the area where the monitoring equipment is located, combined with the sensor's saturation upper limit calibration.

[0115] Ideally, by introducing the illumination vector and fog concentration parameters from the real physical world and combining them with the wavelength physical properties of each spectral band for joint measurement, it is possible to quantify the non-uniform interference caused by complex meteorological conditions to each independent spectral band, thus avoiding the problem of invalid data contaminating the matching results caused by treating all spectral bands equally in traditional methods.

[0116] For example, in a scene with strong backlight at noon and light haze, the ambient light vector shows that the direct radiation intensity is extremely high and the light source is located at the edge of the lens field of view. The processing unit determines that the saturation of the visible light blue band and green band exceeds the risk level, which is a high-risk state. At the same time, the fog concentration parameter shows that the density of suspended particles is relatively large. Based on the Mie scattering model, the processing unit determines that the scattering and obstruction of the visible light blue band reaches 80% and the scattering and obstruction of the near-infrared band is 15%.

[0117] Furthermore, when the monitoring equipment is not equipped with physical meteorological sensors, the processing unit directly extracts the radiance distribution of the sky region and the contrast attenuation gradient of the distant region from the multispectral image sequence, and uses the image inversion algorithm to derive the equivalent ambient light vector and fog concentration parameters, ensuring that environmental interference quantification can still be performed even with simplified hardware.

[0118] S33: Determine the effective signal strength value of each spectral band based on the difference between local contrast and cross-spectral response, and determine the environmental interference level of each spectral band based on the saturation exceedance risk and scattering shielding degree. Calculate the ratio between the effective signal strength value and the environmental interference level value, determine the penetration reliability of each spectral band based on the ratio, and generate reliable spectral band distribution data by arranging the penetration reliability of each spectral band in order of wavelength.

[0119] Specifically, the processing unit combines the local contrast and the amplitude of the cross-spectral response difference to obtain the effective signal strength value of each spectral band, combines the saturation exceedance risk and the degree of scattering obstruction to obtain the environmental interference level value of each spectral band, calculates the ratio of the effective signal strength value to the environmental interference level value, determines the penetration reliability of each spectral band based on the ratio result, and finally arranges the penetration reliability of each spectral band in order of wavelength from shortest to longest to generate reliable spectral band distribution data.

[0120] It should be noted that the effective signal strength value comprehensively characterizes the true reflected energy and spectral gradient characteristics retained by the target material in the current spectral band; the environmental interference level value quantifies the combined destructive effect of overexposure and fog scattering on the current spectral band; the penetration reliability reflects the confidence level of the current spectral band data for target material matching; and the spectral band reliability distribution data is an ordered set of penetration reliability values ​​for all spectral bands. Because the destructive effects of fog and strong light on different wavelengths within the UHV channel are highly non-uniform, calculating the ratio of effective signal strength to interference level allows us to transform abstract image quality into a quantitative indicator that can be directly used for matching and allocation, ensuring that the matching algorithm relies only on high-confidence spectral data.

[0121] Specifically, when calculating the ratio, the processing unit checks if the environmental interference level is zero. If the environmental interference level is greater than zero, the processing unit directly calculates the ratio of the effective signal strength to the environmental interference level as the penetration reliability. If the environmental interference level is zero, the processing unit determines that the current spectral band is not affected by any environmental interference and sets the penetration reliability to the theoretical maximum value. If the environmental interference level exceeds the preset interference limit threshold, the processing unit determines that the current spectral band data is completely invalid and sets the penetration reliability to zero. The preset interference limit threshold is calibrated by the sensor's noise floor level under extreme conditions of dense fog and strong backlight.

[0122] Ideally, by setting extreme condition branches with environmental interference levels of zero and exceeding limits, it is possible to handle extreme conditions where the sensor is in an ideal dark room environment or completely obscured by dense fog, preventing calculation crashes or incorrect participation of failed spectral segments in matching due to a denominator of zero, and ensuring the numerical stability and logical rigor of the entire spectral evaluation process.

[0123] For example, for the near-infrared band, the calculated effective signal strength value is 320 and the environmental interference level value is 40. The ratio of the effective signal strength value to the environmental interference level value is 8, and the penetration reliability of the near-infrared band is 8. For the visible light blue band, the effective signal strength value is 50 and the environmental interference level value is 250. The ratio is 0.2, and the penetration reliability is 0.2. The generated reliable distribution data sequence of the spectral bands arranged from shortest to longest wavelengths shows a distribution state with extremely low values ​​in the visible light band and extremely high values ​​in the infrared band.

[0124] Furthermore, after generating reliable spectral distribution data, the processing unit calculates the statistical variance of the penetration reliability of all spectral bands. When the statistical variance exceeds a preset variance threshold, it determines that there is severe selective spectral attenuation in the current scene. The processing unit automatically triggers the exposure parameter reconstruction command of the multispectral camera and reduces the exposure time of the high-reliability spectral bands to prevent saturation in the next acquisition cycle, thereby improving the overall spectral acquisition quality. The preset variance threshold is derived from the statistical analysis of historical spectral fluctuation variance during alternations between clear and hazy days.

[0125] S4: Generate the target's true spectral data from the structure-exposed spectral data using reliable spectral distribution data, and output the UAV matching results.

[0126] In some embodiments, step S4, which involves generating the target's true spectral data from the structure stripping spectral data using reliable spectral distribution data and outputting the UAV matching result, encompasses two sub-steps, S41 and S42: S41: Generate the target's true spectral data from the structural stripping spectral data using reliable spectral distribution data.

[0127] It should be noted that after the aforementioned steps of removing background interference and evaluating the reliability of each spectral band, the remaining spectral data still has wavelength sampling fragmentation caused by the removal of some spectral bands due to environmental interference. Step S41 restores the continuous and complete material reflection law of the UAV target by screening highly reliable spectral bands and reconstructing the response values ​​of the missing wavelength range, thereby solving the problem of incomplete target spectral features caused by the failure of some spectral bands in the UHV channel, which leads to mismatch.

[0128] S42: Determine the similarity between the target's true spectral data and the standard spectral curve, and output the UAV matching result based on the motion trajectory features.

[0129] It should be noted that kites, balloons, or birds in the UHV channel may have a very high spectral overlap with UAVs in certain spectral bands. Simply relying on the highest similarity can easily lead to false alarms. Step S42 calculates the difference between the best and second-best results and compares it with the standard deviation of the overall distribution. At the same time, it superimposes trajectory continuity verification to lock the UAV target from the dual dimensions of spectral discrimination and physical motion law, thereby reducing the false alarm rate caused by similar interference objects.

[0130] In some embodiments, the step S41, which generates the target true spectral data from the structure stripping spectral data using spectral band reliable distribution data, encompasses three sub-steps from S411 to S413: S411: Calculate the statistical mean of the penetration reliability of each spectral band based on the reliable distribution data of the spectral bands, select the spectral bands with a penetration reliability greater than the statistical mean as the effective spectral bands, and extract the wavelength identifiers and pixel response values ​​of the effective spectral bands from the structure stripping spectral data to construct a discrete effective spectral point set.

[0131] Specifically, the processing unit reads the penetration reliability of each wavelength channel recorded in the spectral band reliability distribution data, calculates the arithmetic mean of the penetration reliability of all wavelength channels to obtain the statistical mean, traverses all wavelength channels and marks the wavelength channels with a penetration reliability greater than the statistical mean as valid spectral bands, and extracts the wavelength identifiers and pixel response values ​​of the valid spectral bands from the structure stripping spectral data to construct a discrete set of valid spectral points.

[0132] It should be noted that the statistical mean reflects the central level of the overall quality of all spectral bands under the current observation environment. The effective spectral band is the set of wavelength channels that are less affected by environmental interference and retain true material information. The discrete effective spectral point set is a two-dimensional coordinate sequence composed of the wavelength values ​​and corresponding radiance of the effective spectral bands. Since multispectral cameras will inevitably have some spectral bands that are completely ineffective under fog or strong light, directly using the incomplete discrete points for template comparison will lead to the loss of feature dimensions. By calculating the statistical mean and dynamically selecting high-quality anchor points, we can ensure that the spectral data input to the matching module has a very high physical confidence level.

[0133] Ideally, by taking the arithmetic mean of the penetration reliability as a dynamic screening benchmark, we can get rid of the rigid mode of relying on manually fixing the reliable bands in traditional methods. This allows the selection range of effective spectral bands to automatically shrink or expand with the current illumination and fog concentration, ensuring that the underlying data on which the reconstructed curve depends always has extremely high purity.

[0134] For example, the spectral band reliability distribution data records the penetration reliability of the visible light blue band, green band, red band, and near-infrared band as 0.2, 0.4, 0.8, and 0.9, respectively. The processing unit calculates the arithmetic mean and obtains a statistical mean of 0.575. After traversing the spectrum, the near-infrared band and the red band are marked as valid spectral bands, and the wavelength identifiers and pixel response values ​​corresponding to these two bands are extracted to construct a discrete set of valid spectral points.

[0135] S412: Construct curves for the discrete effective spectral point set to establish the correspondence between wavelength and radiance spectral distribution, determine the wavelength interval between two adjacent data points in the discrete effective spectral point set, judge the relationship between the wavelength interval and the original spectral band wavelength step size of the multispectral image sequence to determine the interpolation interval, obtain the radiance value and wavelength value of the data points on both sides of the interpolation interval to determine the local transition correspondence, calculate the radiance derivation value of each interpolation wavelength to supplement the discrete effective spectral point set to obtain the reconstructed spectral curve.

[0136] Specifically, the processing unit constructs curves from the discrete effective spectral point set to establish the spectral distribution correspondence between wavelength and radiance, determines the wavelength interval between two adjacent data points in the discrete effective spectral point set, and judges the relationship between the wavelength interval and the original spectral band wavelength step size of the multispectral image sequence. When the wavelength interval is equal to the original spectral band wavelength step size, it is determined that there is no data loss between the current adjacent data points and the original state is maintained. When the wavelength interval is less than the original spectral band wavelength step size, it is determined that there is spectral aliasing in the sensor and a filtering and smoothing operation is triggered. When the wavelength interval is greater than the original spectral band wavelength step size, the interval corresponding to the wavelength interval is determined as the interpolation interval. The radiance values ​​and wavelength values ​​of the data points on both sides of the interpolation interval are obtained. Based on the radiance values ​​and wavelength values ​​of the data points on both sides, the local transition correspondence between wavelength and radiance in the interpolation interval is determined. The radiance derived value of each interpolated wavelength in the interpolation interval is calculated according to the local transition correspondence. The radiance derived value is added to the discrete effective spectral point set to obtain the reconstructed spectral curve.

[0137] It should be noted that the wavelength interval is the physical difference between the center wavelengths of two adjacent effective spectral bands, the original spectral band wavelength step size is the fixed physical spacing between the center wavelengths of adjacent filters in the multispectral camera hardware design, the interpolation interval is the wavelength blank band left due to spectral band failure or low signal-to-noise ratio, and the local transition correspondence is the local energy gradient law derived from the anchor points at both ends of the blank band. Since the spectral reflection peaks of the UAV surface coating usually have a specific physical half-width at half-maximum, if the missing interval is not properly filled, the position of the characteristic peak during template comparison will shift. By determining the relationship between the wavelength interval and the hardware step size and performing targeted interpolation, spectral breaks caused by environmental interference can be repaired, ensuring that the reconstructed curve truly reflects the physical reflection peak position of the target material.

[0138] Preferably, by introducing full-coverage conditional branches with wavelength intervals equal to, less than, and greater than the hardware step size, it is possible to identify three distinct physical conditions: normal sensor sampling, spectral aliasing, and data loss. This avoids unnecessary interpolation calculations during normal sensor sampling, which could lead to spectral smoothing distortion. At the same time, it can trigger filtering operations in a timely manner when spectral aliasing occurs, ensuring the fidelity of the reconstructed spectral curve under various hardware conditions.

[0139] For example, the wavelength step of the original spectral band of the multispectral camera is set to 20 nanometers. The wavelengths of two adjacent data points in the discrete effective spectral point set are 800 nanometers and 860 nanometers, respectively, with a wavelength interval of 60 nanometers, which is greater than 20 nanometers. The processing unit determines that the interval between 800 nanometers and 860 nanometers is the interpolation interval, obtains the radiance value of 400 at 800 nanometers and the radiance value of 500 at 860 nanometers, constructs a linear gradient law based on the values ​​at both ends as a local transition correspondence, and calculates the derived values ​​of radiance at 820 nanometers and 840 nanometers as 433 and 466, respectively. The derived values ​​are added to the point set to complete the curve reconstruction.

[0140] Furthermore, when the span of the interpolation interval exceeds three original spectral band wavelength steps, the processing unit determines that the current missing interval is too large and the anchor points at both ends cannot accurately reflect the local gradual change pattern. It abandons the linear gradual change derivation and retrieves the prior morphological curve of the corresponding wavelength interval from the standard UAV spectral template library. The prior morphological curve is scaled according to the values ​​of the anchor points at both ends and then filled into the interpolation interval to prevent the characteristic peak position from being seriously shifted due to long-distance blind interpolation.

[0141] S413: Align the reconstructed spectral curve with radiance to generate the target's true spectral data.

[0142] Specifically, the processing unit reads the radiative response reference curves for each spectral band in the factory calibration document of the multispectral camera, compares the radiance value of each wavelength channel in the reconstructed spectral curve with the reference response value of the corresponding wavelength channel in the radiative response reference curve, calculates the ratio between the two to obtain the radiative correction factor, and uses the radiative correction factor to align the radiance of the reconstructed spectral curve to generate the target's true spectral data.

[0143] It should be noted that the radiation response baseline curve reflects the difference in photoelectric conversion efficiency of the sensor at different wavelengths, the radiation correction factor is used to eliminate spectral distortions caused by the hardware itself, and the target true spectral data is a standardized spectral vector after eliminating the differences in device response. Multispectral cameras have inherent differences in the quantum efficiency of sensors across different bands. Directly outputting the raw radiance will lead to an overestimation of energy in the short-wavelength band and an underestimation of energy in the long-wavelength band. By introducing the baseline curve from the factory calibration document for channel-by-channel alignment, the spectral distortion caused by hardware physical differences can be eliminated, providing a standardized data foundation for subsequent high-precision template matching.

[0144] Ideally, by calling the radiation response reference curve in the factory calibration document to perform channel-by-channel proportional correction, the quantum efficiency gap between the sensors of the multispectral camera can be smoothed out, ensuring that the generated true spectral data of the target only reflects the objective physical reflection law of the UAV material, and improving the generalization ability of cross-device and cross-scene spectral matching.

[0145] In some embodiments, step S42, which determines the similarity between the target's true spectral data and the standard spectral curve and combines this with motion trajectory features to output the UAV matching result, encompasses three sub-steps: S421 to S423. S421: Obtain a standard UAV spectral template library, determine the spectral similarity between the target's real spectral data and each standard spectral curve, obtain a similarity set containing multiple spectral similarity values, and extract the maximum and second-largest similarity values ​​from the similarity set.

[0146] Specifically, the processing unit loads a standard UAV spectral template library containing standard spectral curves of various UAV models from non-volatile memory, and uses a cosine similarity algorithm to calculate the cosine value of the spatial angle between the target's real spectral data and each standard spectral curve. The calculated cosine values ​​are mapped to similarity values ​​and collected into a similarity set. Then, all values ​​in the similarity set are sorted in descending order, and the value at the top is extracted as the maximum similarity value and the value at the second top is extracted as the second maximum similarity value.

[0147] It should be noted that the standard UAV spectral template library is a set of reference data on the reflectance of various UAV fuselage materials, collected in advance in a darkroom under standard light sources. Spectral similarity reflects the overlap between the target's reconstructed curve and the reference curve in terms of shape and amplitude. The similarity set is the response distribution generated after comparing the target with all templates in the library. The difference between the maximum and second-largest similarity values ​​represents the degree of advantage of the best matching result over the second-best matching result. Since floating objects in the UHV channel may exhibit extremely high spectral overlap with UAVs in certain spectral bands, relying solely on the highest similarity can easily lead to false alarms. Extracting the maximum and second-largest values ​​provides core data support for subsequent calculations of discrimination.

[0148] Specifically, when calculating the cosine of the spatial angle, the processing unit calculates the inner product of the target's true spectral data vector and the standard spectral curve vector, and then calculates the product of the magnitudes of the two vectors. The ratio of the inner product to the magnitude product is used to obtain the cosine of the spatial angle. The ratio is 1 when the target's true spectral data vector and a certain standard spectral curve vector are in the same direction, and 0 when the two vectors are orthogonal.

[0149] Preferably, by using a cosine similarity algorithm to calculate the cosine value of the spatial angle, the influence of the scaling of the overall radiance amplitude caused by changes in observation distance or local shadows can be eliminated, so that the similarity calculation focuses only on the morphological fluctuation characteristics of the spectral curve, which greatly improves the robustness of template matching under different lighting conditions.

[0150] For example, the standard UAV spectral template library contains standard spectral curves of 10 different UAV models. The processing unit calculates the cosine of the spatial angle between the target's real spectral data and the 10 standard spectral curves one by one. After mapping, a similarity set containing 10 values ​​is obtained. After sorting the similarity set in descending order, the value at the top, 0.92, is extracted as the maximum similarity value, and the value at the second top, 0.85, is extracted as the second maximum similarity value.

[0151] S422: Obtain the motion trajectory features of the candidate multi-spectral response sequence in consecutive frames; when the difference between the maximum similarity value and the second largest similarity value is greater than the standard deviation of the similarity set, and the motion trajectory features show a state of continuous displacement and non-zero acceleration, generate a UAV confirmation signal.

[0152] Specifically, the processing unit acquires the motion trajectory features of the candidate multi-spectral response sequence in consecutive frames, calculates the dispersion of all values ​​in the similarity set to obtain the standard deviation, calculates the difference between the largest and second largest similarity values, and determines the relationship between the difference and the standard deviation. When the difference is less than or equal to the standard deviation, it is determined that the current target spectral features lack discriminative power and a suspected interference signal is generated. When the difference is greater than the standard deviation and the motion trajectory features show continuous displacement and non-zero acceleration, a UAV confirmation signal is generated. When the difference is greater than the standard deviation but the motion trajectory features show discontinuous displacement or zero acceleration, a suspected stationary object signal is generated.

[0153] It should be noted that the motion trajectory features record the spatial displacement and velocity changes of the target over time, while the standard deviation reflects the dispersion of each value in the similarity set. By calculating the difference between the optimal and suboptimal results and comparing it with the standard deviation of the overall distribution, while simultaneously superimposing trajectory continuity verification, it is possible to lock onto the UAV target from both spectral discrimination and physical motion laws.

[0154] Preferably, by setting a three-state full-coverage discrimination logic—where the difference result is less than or equal to the standard deviation, the difference result is greater than the standard deviation and the trajectory is continuous, and the difference result is greater than the standard deviation but the trajectory is interrupted—it is possible to separate floating objects with similar spectra but inconsistent motion patterns from stationary reflective points. This ensures that the final output UAV confirmation signal is supported by both physical and spectral evidence, providing a decision-making basis for power channel airspace risk warnings.

[0155] For example, the similarity set includes 10 comparison values. The largest similarity value of 0.92 corresponds to a specific multi-rotor UAV template, and the second largest similarity value of 0.85 corresponds to a specific fixed-wing UAV template. The difference between the two is 0.07, and the standard deviation of the similarity set is 0.04. Since the difference of 0.07 is greater than the standard deviation of 0.04, and the motion trajectory features show that the target has continuous displacement and non-zero acceleration in 15 consecutive frames, the processing unit generates a UAV confirmation signal. If the second largest similarity value is 0.89, the difference between the two is 0.03, which is less than the standard deviation of 0.04. The processing unit determines that the current target may be a plastic floating object with a similar surface coating and generates a suspected interference signal.

[0156] S423: Outputs drone matching results based on drone confirmation signals.

[0157] Specifically, the processing unit outputs the drone matching result based on the drone confirmation signal, and after generating the drone confirmation signal, it further extracts the wavelength channels in the target's real spectral data whose radiance derivation value exceeds the preset proportion parameter and marks them as high-risk sensitive spectral bands. The wavelength identifiers of the high-risk sensitive spectral bands are packaged with the drone matching result and output to the alarm terminal.

[0158] It should be noted that the preset percentage parameter is set to 30%. The aforementioned value is derived from the contribution of interpolation intervals in historical false alarm samples. The high-risk sensitive spectral bands are wavelength regions in the reconstruction curve that rely on mathematical derivation rather than actual sensor sampling. This indicates to maintenance personnel that the current matching results are highly dependent on specific reconstruction intervals, facilitating a focus on the potential impact of environmental interference on specific wavelengths during manual review.

[0159] Ideally, by identifying and marking high-risk sensitive spectral bands with a high proportion of derived radiance values, spectral confidence distribution maps can be provided to maintenance personnel, enabling the manual review process to accurately focus on the derived range that is susceptible to environmental interference, thereby improving the review efficiency and accuracy of UHV channel airspace supervision.

[0160] See Figure 2 This is an application scenario diagram of this application. The multispectral monitoring device 101 is positioned facing the power transmission line 102 and its surrounding low-altitude airspace to collect multispectral image sequences generated when the intruding drone 103 approaches the power transmission line 102. The intruding drone 103 can be a small flying target observed from a distance. When it flies into the airspace adjacent to the power transmission line 104 in foggy or backlit conditions and is interfered with by reflections from the power transmission line edge or hot spots on hardware, the channel structure mask and the edge neighborhood band 105 cover the background of the power transmission line 102 with a strong structure and its reflective attenuation area. This is used to represent the background pseudo-response stripping range determined by the method of this application and the target true spectrum matching area of ​​the adaptive reconstruction.

[0161] See Figure 3This is a schematic diagram of the structure of the UAV target spectral adaptive matching system for power transmission channels provided in this application embodiment, including: The structure mask module acquires multispectral image sequences and channel structure prior data, projects the channel structure prior data onto the multispectral image sequences, and generates a channel structure mask. The spectral stripping module extracts the motion region that deviates from the channel structure mask to obtain candidate multi-band response sequences. Based on the edge neighborhood spectrum in the channel structure mask, the candidate multi-band response sequences are stripped of structural reflection to obtain structural stripped spectral data. The spectral evaluation module determines the penetration reliability of each spectral band based on the structural stripping spectral data and generates reliable spectral band distribution data. The target matching module generates true target spectral data from the structure-exposed spectral data using reliable spectral distribution data and outputs the UAV matching results. Figure 3 The system of the illustrated embodiment can be used to perform corresponding operations. Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0162] See Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device 40 includes: a processor 41, a memory 42, and a computer program; wherein, The memory 42 is used to store computer programs, and the memory may also be flash memory. Computer programs may be, for example, application programs or functional modules that implement the methods described above.

[0163] The processor 41 is used to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0164] Alternatively, the memory 42 can be either standalone or integrated with the processor 41.

[0165] When the memory 42 is a device independent of the processor 41, the device may also include: Bus 43 is used to connect memory 42 and processor 41.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. An adaptive spectral matching method for UAV targets in power transmission channels, characterized in that, include: Acquire multispectral image sequences and channel structure prior data, and project the channel structure prior data onto the multispectral image sequences to generate a channel structure mask; The motion region deviating from the channel structure mask is extracted to obtain candidate multi-band response sequences. The candidate multi-band response sequences are then subjected to structural reflectance stripping based on the edge neighborhood spectrum in the channel structure mask to obtain structural stripping spectral data. Based on the structure stripping spectral data, the penetration reliability of each spectral band is determined, and reliable spectral band distribution data is generated; The target's true spectral data is generated from the structure-extracted spectral data using the reliable distribution data of the spectral band, and the UAV matching result is output.

2. The method according to claim 1, characterized in that, The step of projecting the prior data of the channel structure onto the multispectral image sequence to generate a channel structure mask includes: Acquire the current observation attitude and focal length parameters of the monitoring equipment; Based on the current observation posture and the focal length parameters, the three-dimensional structure wireframe in the prior data of the channel structure is projected onto the two-dimensional imaging plane of the multispectral image sequence to obtain the structure projection profile; An initial structural mask is generated based on the pixel distribution of the structural projection contour in the two-dimensional imaging plane. Extract the edge pixel set of the initial structure mask, and determine the number of tolerance pixels to be extended outward based on the wind deflection amplitude of the transmission line and the calibration error of the monitoring equipment; The number of tolerance pixels is shifted outward along the normal direction of the edge pixel set to generate a structural extension region; The extended region of the structure is added to the outer edge of the initial structure mask to obtain the channel structure mask.

3. The method according to claim 2, characterized in that, The step of generating an initial structural mask based on the pixel distribution of the structural projection contour in the two-dimensional imaging plane includes: Extract the spatial depth value corresponding to each structural line segment in the three-dimensional structural wireframe; The structural projection contour is layered according to the spatial depth value to obtain multiple depth-level contours, and the local imaging region corresponding to any of the depth-level contours is determined in the two-dimensional imaging plane. Obtain the radiance difference between adjacent pixels within the local imaging region; Determine the statistical mean and standard deviation of the radiance difference within the local imaging region, and determine the pixel position corresponding to the radiance difference that exceeds the range of the statistical mean plus or minus the standard deviation as the actual structural boundary; The set of projected boundary pixels corresponding to the depth level contour in the two-dimensional imaging plane is updated to the pixel position corresponding to the actual structural boundary to obtain the corrected level contour. The initial structure mask is generated by combining the modified layer contours corresponding to each depth layer contour.

4. The method according to claim 1, characterized in that, The extraction of motion regions deviating from the channel structure mask to obtain candidate multi-band response sequences includes: The difference in radiance between adjacent frames in the multispectral image sequence at the same pixel coordinates is determined to obtain a motion difference image; The pixels covered by the channel structure mask are removed from the motion difference image to obtain the residual motion region; Extract the set of spatially adjacent moving pixels with the same direction of motion from the residual motion region to obtain the independent motion region; The coordinates of the independent motion region in each spectral band of the multispectral image sequence are determined, the radiance value at the coordinates of the coordinates is extracted, and the candidate multispectral response sequence is generated.

5. The method according to claim 4, characterized in that, The step of extracting a set of spatially adjacent motion pixels with the same direction of motion from the residual motion region to obtain an independent motion region includes: Extract the set of spatially adjacent candidate pixels in the residual motion region, and obtain the centroid position sequence of the candidate pixel set in multiple consecutive frames of images; Determine the displacement direction between adjacent frames in the centroid position sequence; When the displacement direction does not alternately reverse between consecutive frames, and the centroid position sequence does not exhibit a state of back-and-forth movement along the same trajectory, the candidate pixel set is determined to be a valid set of moving pixels. The set of effective moving pixels is taken as the independent motion region.

6. The method according to claim 1, characterized in that, The step of performing structural reflectance stripping on the candidate multi-band response sequence based on the edge neighborhood spectrum in the channel structure mask to obtain structurally stripped spectral data includes: Starting from the boundary of the channel structure mask, extend at least three adjacent pixel rows outward from the channel structure mask to obtain the edge neighborhood band; Extract the spectral response values ​​of each pixel within the edge neighborhood band in the multispectral image sequence to generate neighborhood spectral distribution data; The decreasing radiance distribution of the edge neighborhood band along the direction away from the boundary of the channel structure mask is determined based on the neighborhood spectral distribution data. Edge radiation subtraction reference data is constructed based on the described decreasing radiance distribution. Calculate the normal distance of each moving pixel in the candidate multi-band response sequence from the boundary of the channel structure mask; match the corresponding radiance subtraction value in the edge radiance subtraction reference data according to the normal distance, and subtract the radiance subtraction value from the corresponding moving pixel in the candidate multi-band response sequence to obtain the structure stripping spectral data.

7. The method according to claim 6, characterized in that, The step of determining the decreasing radiance distribution of the edge neighborhood band along the direction away from the channel structure mask boundary based on the neighborhood spectral distribution data includes: The edge neighborhood band is divided into a group of pixel sub-bands arranged along the normal direction of the channel structure mask boundary; For any of the aforementioned pixel sub-bands, obtain the radiance value of the center pixel within the pixel sub-band in a single spectral band; Based on the radiance value of each pixel sub-band and the normal distance of the pixel sub-band from the boundary of the channel structure mask, a negative correlation relationship is determined as radiance extends with normal distance. The radiance drop amplitude between adjacent pixel sub-bands is determined based on the negative correlation correspondence. The radiance decrease amplitude of each individual spectral band is combined to generate the radiance decrease distribution state.

8. The method according to claim 1, characterized in that, The step of determining the penetration reliability of each spectral band based on the structural stripping spectral data and generating reliable spectral band distribution data includes: Extract the local contrast and cross-band response difference magnitude of each spectral band in the structure stripping spectral data; Obtain the ambient light vector and fog concentration parameters during the acquisition of the multispectral image sequence; The saturation exceedance risk of each spectral band is determined based on the ambient light vector, and the scattering and obscuring degree of each spectral band is determined based on the fog concentration parameter. The effective signal strength value of each spectral band is determined based on the difference between the local contrast and the cross-spectral response, and the environmental interference level of each spectral band is determined based on the saturation exceedance risk and the scattering shielding degree. Calculate the ratio between the effective signal strength value and the environmental interference level value, and determine the penetration reliability of each spectral band based on the ratio. The penetration reliability of each spectral band is arranged in order of wavelength to generate reliable distribution data for that spectral band.

9. The method according to claim 8, characterized in that, The extraction of the local contrast and cross-band response difference magnitude of each spectral band in the structure stripping spectral data includes: For any target spectral band in the structure stripping spectral data, the difference in radiance between the target region pixels and the local background pixels within the target spectral band is determined to obtain the local contrast. Select a reference spectral segment adjacent to the target spectral segment, and align the pixel response values ​​of the target spectral segment and the reference spectral segment at the same spatial location; Obtain the response difference sequence between the aligned target spectral band and the reference spectral band; The amplitude of the cross-spectral response difference is determined based on the fluctuation amplitude of the response difference sequence.

10. The method according to claim 1, characterized in that, The step of generating target true spectral data from the structure-exfoliated spectral data using the reliable distribution data of the spectral band includes: Based on the reliable distribution data of the spectral bands, calculate the statistical mean of the penetration reliability of each spectral band, and select the spectral bands with a penetration reliability greater than the statistical mean as valid spectral bands; Extract the wavelength identifiers and pixel response values ​​belonging to the effective spectral band from the structure stripping spectral data to construct a discrete effective spectral point set; Curves are constructed for the discrete effective spectral point set to establish the correspondence between wavelength and spectral distribution of radiance; Based on the spectral distribution correspondence, the derived values ​​of radiance at interpolation wavelengths not included in the effective spectral band are calculated to obtain the reconstructed spectral curve; The reconstructed spectral curves are aligned with radiance to generate the target's true spectral data.

11. The method according to claim 10, characterized in that, The step of calculating the derived radiance values ​​at interpolated wavelengths not included in the effective spectral band based on the spectral distribution correspondence, to obtain the reconstructed spectral curve, includes: Determine the wavelength interval between two adjacent data points in the discrete effective spectral point set; When the wavelength interval is greater than the original spectral band wavelength step size of the multispectral image sequence, the interval corresponding to the wavelength interval is determined as the interval to be interpolated; Obtain the radiance and wavelength values ​​of the data points on both sides of the interpolation interval; Based on the radiance and wavelength values ​​of the data points on both sides, the local transitional correspondence between wavelength and radiance within the interpolation interval is determined; Based on the local transition correspondence, calculate the derived radiance values ​​for each interpolation wavelength within the interpolation interval; The derived radiance values ​​are added to the set of discrete effective spectral points to obtain the reconstructed spectral curve.

12. The method according to claim 1, characterized in that, The output drone matching results include: Obtain a standard UAV spectral template library, which contains a set of standard spectral curves; Determine the spectral similarity between the target real spectral data and each of the standard spectral curves to obtain a similarity set containing multiple spectral similarity values; Extract the maximum and second-largest similarity values ​​from the similarity set; Obtain the motion trajectory features of the candidate multi-band response sequence in consecutive frames; When the difference between the maximum similarity value and the second largest similarity value is greater than the standard deviation of the similarity set, and the motion trajectory features exhibit continuous displacement and non-zero acceleration, a drone confirmation signal is generated. The drone matching result is output based on the drone confirmation signal.

13. A target spectral adaptive matching system for UAVs operating on power transmission channels, employing the target spectral adaptive matching method for UAVs operating on power transmission channels as described in any one of claims 1 to 12, characterized in that, include: The structure mask module acquires a multispectral image sequence and channel structure prior data, projects the channel structure prior data onto the multispectral image sequence, and generates a channel structure mask. The spectral stripping module extracts the motion region that deviates from the channel structure mask to obtain candidate multi-band response sequences. Based on the edge neighborhood spectrum in the channel structure mask, the module performs structural reflectance stripping on the candidate multi-band response sequences to obtain structural stripped spectral data. The spectral evaluation module determines the penetration reliability of each spectral band based on the structure stripping spectral data and generates reliable spectral band distribution data. The target matching module uses the reliable distribution data of the spectral band to generate the true spectral data of the target from the spectral data stripped from the structure, and outputs the UAV matching result.

Images