Airport bird identification and tracking method and system based on visual radar fusion

Abstract

The invention discloses an airport bird identification and tracking method and system based on visual radar fusion, and the method comprises the steps: carrying out the three-dimensional positioning and morphological analysis of an airport airspace target through the cooperative work of a binocular stereo camera and a millimeter-wave radar; a background difference technology is adopted to extract candidate targets, and multi-dimensional descriptors are constructed through visual features such as morphological asymmetry and wing flapping frequency; performing distance and speed verification on a visual identification result by using micro-Doppler features of radar echoes, and determining a bird identification target through fusion error evaluation; an inter-frame image matching and shielding detection technology is implemented to generate a continuous trajectory sequence, and trajectory fluctuation parameters and periodic features are extracted to reflect flight dynamics; according to the method, identification inconsistency parameters are obtained through consistency checking of multi-frame identification results, bird target identification tracking results are output according to quality grading, and reliable information support is provided for airport airspace safety early warning and bird repelling decision making.

Description

Technical Field

[0001] This invention relates to the field of airspace safety monitoring technology, and in particular to a method and system for airport bird identification and tracking based on visual radar fusion. Background Technology

[0002] Biological activity in airport airspace remains a significant threat to aviation safety. Collisions between aircraft and biological organisms during takeoff and landing can cause serious consequences, including engine damage and structural breakage, threatening flight safety and resulting in economic losses. Statistics show that such incidents are on the rise globally, making the effective monitoring and identification of biological activity in airspace an urgent need for airport safety management.

[0003] Existing airspace monitoring technologies primarily rely on single sensors for target detection, which has numerous limitations. Optical imaging-based methods are significantly affected by ambient light and weather conditions, making stable operation difficult at night or in foggy or hazy weather, and their ability to identify the shape of small, distant targets is limited. While radio frequency (RF) detection-based methods are not constrained by lighting conditions, they lack information on target appearance, making it difficult to distinguish between biological species and non-biological interference, resulting in a high false alarm rate. More importantly, existing methods mostly focus on instantaneous target detection, lacking continuous tracking of target trajectories and in-depth analysis of behavioral characteristics, making it impossible to predict target flight trends and threat levels, leading to delayed early warning responses. Summary of the Invention

[0004] This invention discloses an airport bird identification and tracking method and system based on visual radar fusion. By coordinating the processing of three-dimensional visual images and radar echo data, the morphological and dynamic features of the target are extracted, and reliable identification of bird targets is achieved through multimodal feature fusion. Furthermore, continuous tracking of the target is achieved through inter-frame correlation and trajectory analysis, and finally, a hierarchical identification result is formed through multi-frame consistency verification, providing accurate information support for airport airspace safety early warning.

[0005] The first aspect of this invention proposes a method for airport bird identification and tracking based on visual radar fusion, comprising the following steps: Acquire 3D visual image data and millimeter-wave radar echo data of the airport airspace, and perform background subtraction on the 3D visual image data to determine the outline of candidate targets; Connected component segmentation is performed on the candidate target contour to determine independent target regions. Morphological analysis is performed on the independent target regions to obtain morphological asymmetry parameters. Visual features are extracted from the independent target regions to construct visual descriptors. The morphological asymmetry parameter and the visual descriptor are fused to determine the preliminary recognition result. The millimeter-wave radar echo data is used to perform distance and velocity verification on the preliminary recognition result to generate fusion error parameters. Based on the fusion error parameters and the morphological asymmetry parameter, a credibility assessment is performed to determine the bird recognition target. Inter-frame image matching is performed on the bird identification target to generate a target trajectory sequence. The continuity of the target trajectory sequence is verified to obtain trajectory fluctuation parameters and extract trajectory periodic features. Dynamic feature data is determined based on the trajectory fluctuation parameters and the trajectory periodic features. The dynamic feature data and the bird identification target are combined to generate a multi-frame identification result sequence. A consistency check is performed on the multi-frame identification result sequence to obtain the identification inconsistency parameter. Based on the identification inconsistency parameter, the bird target identification and tracking results are output in a graded manner.

[0006] A second aspect of this invention proposes an airport bird identification and tracking system based on visual radar fusion, comprising: The data acquisition module is used to acquire 3D visual image data and millimeter-wave radar echo data of the airport airspace, and to perform background subtraction on the 3D visual image data to determine the outline of candidate targets. The feature extraction module is used to perform connected component segmentation on the candidate target contour to determine independent target regions, perform morphological analysis on the independent target regions to obtain morphological asymmetry parameters, and extract visual features from the independent target regions to construct visual descriptors. The target recognition module is used to perform feature fusion of the morphological asymmetry parameter and the visual descriptor to determine the preliminary recognition result, use the millimeter-wave radar echo data to perform distance and velocity verification on the preliminary recognition result to generate fusion error parameters, and perform credibility assessment based on the fusion error parameters and the morphological asymmetry parameter to determine the bird recognition target; The trajectory generation module is used to perform inter-frame image matching to generate a target trajectory sequence for the bird identification target, perform continuity verification on the target trajectory sequence to obtain trajectory fluctuation parameters and extract trajectory periodic features, and determine dynamic feature data based on the trajectory fluctuation parameters and the trajectory periodic features. The result output module is used to combine the dynamic feature data with the bird identification target to generate a multi-frame identification result sequence, perform a consistency check on the multi-frame identification result sequence to obtain an identification inconsistency parameter, and output the bird target identification and tracking results in a graded manner based on the identification inconsistency parameter.

[0007] The beneficial effects of this invention are reflected in the following aspects: First, by extracting candidate target contours through depth-layer differential technology of three-dimensional visual images and performing morphological analysis on independent target regions to obtain asymmetry parameters, and simultaneously extracting wing flapping frequency from continuous frame area changes to construct visual descriptors, a comprehensive characterization of the static morphology and dynamic flapping features of bird targets is achieved, enhancing the ability of visual features to distinguish bird species. Second, by employing the micro-Doppler features of millimeter-wave radar echoes to perform distance and velocity verification on the visual recognition results, flapping pattern matching is used to form flapping matching degree, and credibility is evaluated based on fusion error parameters and morphological asymmetry parameters. Multimodal feature fusion enhances the accuracy of bird target recognition and effectively reduces the false recognition rate of a single sensor. Finally, by establishing target tracking association through inter-frame image matching, performing continuity verification on trajectory sequences to obtain fluctuation parameters and extracting periodic features, and combining dynamic feature data to generate a multi-frame recognition result sequence, a consistency check is performed to obtain recognition inconsistency parameters and output quality grading, achieving continuous tracking and recognition quality assessment of bird targets, providing reliable decision support for airspace threat early warning. Attached Figure Description

[0008] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0009] Figure 1 This is a flowchart illustrating an airport bird identification and tracking method based on visual radar fusion according to the present invention.

[0010] Figure 2 This is a structural block diagram of an airport bird identification and tracking system based on visual radar fusion according to the present invention. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0013] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0014] The technical solutions of the embodiments of this application will be described below.

[0015] like Figure 1 As shown, this embodiment of the invention provides a method for airport bird identification and tracking based on visual radar fusion, including the following steps S110-S150: Step S110: Obtain 3D visual image data and millimeter-wave radar echo data of the airport airspace, and perform background subtraction on the 3D visual image data to determine the outline of the candidate target.

[0016] Specifically, 3D visual image data and millimeter-wave radar echo data of the airport airspace are acquired. Binocular stereo cameras are installed on monitoring towers on both sides of the airport runway, at a height of 15 meters, with a pitch angle tilted downwards by 30 degrees to cover an airspace of 500 meters above the runway. 3D visual image data is acquired simultaneously through the left and right lenses of the stereo cameras at a frame rate of 30fps and a resolution of 1920×1080 pixels. The stereo camera baseline is 0.8 meters, the camera focal length is 16mm, the stereo matching accuracy is 0.1 meters, and the depth measurement range covers an airspace interval of 50-500 meters. The 3D visual image data includes both RGB color values ​​and depth values ​​for each pixel. The 3D visual image data of the airport environment includes static backgrounds such as the runway surface, buildings, and vehicles on the apron, as well as dynamic targets such as flying birds, aircraft, and drones. For example, in a monitoring image at the runway starting point, the foreground shows taxiing aircraft (depth 150 meters), the midground shows the terminal building (depth 200-250 meters), and the background is the sky (depth > 400 meters or invalid). When a bird flies from left to right, its depth value changes from 180 meters to 220 meters; the depth difference with the background makes it easy to detect. Millimeter-wave radar is installed at a monitoring station on the extended centerline of the runway. The radar operates at 77 GHz, has a detection range of 1000 meters, an angular resolution of 1 degree, and a velocity resolution of 0.1 m / s. Millimeter-wave radar echo data records four parameters of the target: range, azimuth, pitch angle, and radial velocity. Each scan cycle outputs multiple target point cloud data. Millimeter-wave radar echo data is sensitive to bird flight speed; typical bird flight speeds are 8-15 m / s. The radar can effectively distinguish birds from a stationary background, but it cannot provide detailed information about the target's shape. The fusion of millimeter-wave radar echo data and 3D visual image data can achieve complementary advantages.

[0017] In some embodiments, the step of determining candidate target contours by background subtraction of the 3D visual image data includes: extracting depth information from the 3D visual image data to generate a depth map; dividing the depth map into depth layers to form a near-field layer and a far-field layer; performing differential detection in the near-field layer and the far-field layer to generate layered differential results; and performing depth boundary fusion based on the layered differential results to determine candidate target contours.

[0018] Depth information is extracted from 3D visual image data to generate a depth map. 3D visual image data is simultaneously acquired by the left and right lenses of a binocular camera. The images acquired by the left and right lenses are stereo matched to obtain a disparity map. Each pixel in the 3D visual image data is converted into a depth value through stereo matching. The disparity value d of each pixel in the disparity map is converted into a depth value Z using the formula Z = f × B / d, where f is the camera focal length (16mm, converted to 0.016 meters), B is the baseline (0.8 meters), d is the disparity value in pixels, and the depth value Z is in meters. Disparity and depth are inversely proportional. The depth map stores the depth information of the entire image in a two-dimensional matrix format. The matrix size is consistent with the image resolution (1920 × 1080). The depth value range in the depth map is 50-500 meters; pixels outside this range are marked as invalid depth. In airport scenes, the depth values ​​of buildings are typically in the range of 100-200 meters, the depth values ​​of flying birds are in the range of 150-400 meters, and the depth values ​​of the distant sky background are close to 500 meters or are invalid values. Noise points in the depth map are suppressed using median filtering with a 5×5 pixel filter window. The filtered depth map retains the depth contour of the main target while removing isolated noise. The depth measurement error increases with distance: ±0.05 meters at 50 meters, ±0.3 meters at 200 meters, and ±0.8 meters at 400 meters.

[0019] The depth map is divided into near-field and far-field layers. The depth values ​​range from 50 to 500 meters, divided into two layers with a depth threshold of 200 meters. Pixels with a depth less than 200 meters are assigned to the near-field layer, while those with a depth of 200 meters or more are assigned to the far-field layer. Pixels with invalid depth values ​​are treated as background in the far-field layer. The near-field layer covers the airport runway, taxiway, apron, and surrounding buildings. Targets in this layer are relatively large and have clear textures, resulting in high depth measurement accuracy. Birds flying in the near-field layer are typically low-flying during takeoff and landing or taking off after roosting. The far-field layer covers the mid-to-high altitude airspace above the runway and the distant skyline. Targets in this layer are smaller and more susceptible to background interference, leading to larger depth measurement errors. Birds flying in the far-field layer are typically cruising or crossing long distances. The 200-meter depth threshold was chosen because the depth resolution of a binocular camera drops to 1 meter at this distance. Beyond this distance, the reliability of depth information decreases. For example, when a bird flies at 180 meters, the depth measurement accuracy is ±0.25 meters, accurately distinguishing the bird from a building 190 meters away. However, at 210 meters, the accuracy drops to ±0.5 meters. Therefore, a larger differential threshold is used to accommodate the increased error. Layered processing allows for differentiated detection strategies based on different depth characteristics. Pixel labels for the near-field and far-field layers are denoted as 0 and 1, respectively, and the label map and depth map have the same size.

[0020] Differential detection is performed in the near and far layers to generate layered differential results. The background model is constructed using the median depth of 50 consecutive frames. Independent background models are established for the near and far layers. The near layer background model includes the depth distribution of static structures such as buildings, ground, and helipads, while the far layer background model includes the depth distribution of distant backgrounds such as the sky and distant mountains. The depth map of the current frame is differentially analyzed with the corresponding background models in both the near and far layers. The difference is the current depth minus the background depth. Pixels whose absolute difference exceeds a threshold T are identified as foreground pixels. The differential threshold T_near for the near-field layer is set to 0.5 meters. This threshold is greater than the near-distance depth error of ±0.3 meters to avoid false detections, while being less than the 0.3-0.5 meter depth span of a typical bird's body length to detect bird targets. For example, a pigeon flying over a helipad at a height of 120 meters has a body thickness of approximately 0.3 meters and a wingspan of 0.8 meters. The depth difference between the pigeon and the background buildings is approximately 5 meters. When the differential threshold is set to 0.5 meters, although the 0.3-meter depth change produced by the pigeon's body is less than the threshold, the depth gradient produced by its wingspan exceeds the threshold and can be detected. The depth difference between flying birds and background buildings in the near-field layer is typically 2-10 meters, satisfying the differential threshold. The differential threshold T_far for the far-field layer is set to 1.5 meters. This threshold is suitable for the characteristic of a far-distance depth error of ±0.8 meters. The depth difference between flying birds and the sky background in the far-field layer can reach 50-100 meters, providing sufficient sensitivity for differential detection. The foreground mask generated by the background difference is superimposed on the original depth map to form a layered difference result. The layered difference result retains the depth value and position information of the foreground pixels, while the depth value of the background pixels is set to zero.

[0021] Candidate target contours are determined by depth boundary fusion based on the hierarchical difference results. In the hierarchical difference results, foreground pixels in the near and far layers may belong to different parts of the same target near the depth transition zone. Boundary fusion merges the hierarchical foreground pixels from the hierarchical difference results into a complete target through spatial proximity determination. Adjacent pixels in the hierarchical difference results with a depth difference less than 3 meters and a spatial distance less than 10 pixels are considered the same target. A connected component algorithm is applied to extract each connected region from the fused foreground mask. The bounding rectangle of each connected region is extracted to form the candidate target contour. The candidate target contour includes the pixel coordinate sequence of the contour, the coordinates (x, y) of the top-left corner of the bounding rectangle, its width and height dimensions (w, h), and the center position of the bounding rectangle. The area range of candidate target contours is set to 50-5000 pixels to filter out excessively small noise points and excessively large false positive regions. For example, a sparrow at a distance of 200 meters with a projected area of ​​about 80 pixels passes the screening, while a baggage cart on the tarmac 5 meters away with a projected area of ​​8000 pixels is excluded, ensuring that only potential bird targets are retained. Connected regions with an area less than 50 pixels are identified as noise points and removed, while connected regions with an area greater than 5000 pixels are identified as aircraft or large objects and excluded. In airport scenes, the typical bird's projected area at a distance of 200 meters is 200-800 pixels, and the candidate target contours are used for subsequent connected component segmentation and target recognition.

[0022] Step S120: Perform connected component segmentation on the candidate target contour to determine independent target regions, perform morphological analysis on the independent target regions to obtain morphological asymmetry parameters, and extract visual features from the independent target regions to construct visual descriptors.

[0023] Specifically, connected component segmentation is performed on candidate target contours to determine independent target regions. Multiple contours within a candidate target contour may spatially overlap or be distributed in close proximity. Connected component segmentation splits overlapping contours into multiple non-interfering target regions. Candidate target contours with a minimum bounding rectangle distance of less than 20 pixels are identified as potentially overlapping targets. The contour pixel labels of potential overlapping targets are located in the same mask image, with the mask image size matching the original image. Pixels within the candidate target contour are assigned a value of 255, while background pixels are assigned a value of 0. An eight-neighbor connected component labeling algorithm is applied to the mask image to extract each connected region. The labels of connected regions are numbered sequentially starting from 1, with each label corresponding to an independent target region. The pixel coordinates of independent target regions are extracted to form a region point set. The centroid position and area of ​​the point set are recorded in the attribute structure of the independent target region. The average depth value of the independent target region is extracted from the original depth image according to the region pixel coordinates. The depth value is recorded in the attribute structure of the independent target region for depth consistency verification in subsequent inter-frame matching. In airport airspace, a single bird corresponds to one independent target area, while a flock of birds corresponds to multiple independent target areas. In typical scenarios, the number of detected independent target areas ranges from 5 to 20. The area distribution of independent target areas reflects the distance and size of the birds; the area of ​​large birds at close range, such as herons, can reach 800-1200 pixels, while the area of ​​small birds at distant range, such as sparrows, is only 50-150 pixels. The boundaries of independent target areas after segmentation of overlapping targets may be incomplete. Missing boundary areas are repaired using morphological closing operations with a kernel size of 5×5 pixels, resulting in smoother and more continuous contours for the repaired independent target areas. The RGB color information of independent target areas is extracted from the original 3D visual image data. The RGB color distribution of pixels within the area is statistically analyzed using a color histogram and then dimensionality-reduced to represent a 12-dimensional color feature vector. The color histogram reflects the distribution of the dominant color tones of the bird's feathers. Bird feather colors typically exhibit natural tones such as gray, brown, and white, and these color features aid in species identification.

[0024] In some embodiments, the step of performing morphological analysis on the independent target region to obtain morphological asymmetry parameters includes: extracting the target contour boundary in the independent target region to generate a set of boundary points; fitting the boundary point set to the principal axis direction to determine the target principal axis; measuring the area ratio of the left and right sides along the target principal axis to form lateral asymmetry; and determining the morphological asymmetry parameters based on the aspect ratio characteristics of the lateral asymmetry and the target principal axis.

[0025] In an independent target region, the target contour boundary is extracted to generate a boundary point set. Pixels with a value of 255 in the mask image of the independent target region constitute the foreground region, and the boundary between the foreground region and the background is defined as the target contour boundary. Boundary point extraction of the independent target region is implemented using a contour tracking algorithm. The algorithm starts from the top-left pixel of the foreground region and traverses clockwise along the boundary until it returns to the starting point. The pixel coordinates along the traversal path are recorded sequentially to form the boundary point set. The number of points in the boundary point set is equal to the perimeter of the independent target region's contour in pixels. A typical bird target's boundary point set contains 100-400 points; the number of points reflects the complexity of the target contour. The spacing between adjacent points in the boundary point set is 1 pixel or... Each pixel corresponds to a four- or eight-neighbor movement, and the continuity of the point sequence ensures the integrity of the contour. The jagged edges of the contour boundary are smoothed using polygonal approximation, with an approximation accuracy set to 1.5 pixels. The approximated boundary point set retains the main shape of the contour while removing detail fluctuations. In airport scenes, the contour boundaries of birds are dynamically affected by wing flapping; when wings are spread, the boundary point set appears wide and flat, while when wings are folded, it appears long and thin. This temporal change in contour shape contains flight dynamics information. The origin of the boundary point set is the top-left corner of the image, and the coordinate unit is pixels. The boundary point set is converted to relative coordinates with the centroid as the origin for easier principal axis analysis.

[0026] For example, the step of fitting the principal axis direction of the boundary point set to determine the target principal axis includes: extracting centroid coordinates from the boundary point set to generate a target center point; constructing a point distribution matrix based on the target center point and the boundary point set; performing eigenvalue decomposition on the point distribution matrix to obtain the principal eigenvector; and determining the target principal axis based on the principal eigenvector.

[0027] The centroid coordinates are extracted from the boundary point set to generate the target center point. The boundary point set contains a sequence of pixel coordinates of N points, where the i-th point is denoted as P_i(x_i, y_i), and i ranges from 0 to N-1. The centroid coordinates are calculated by averaging the coordinates of all points in the boundary point set. The centroid C(x_c, y_c) is defined as the geometric center of the boundary point set, representing the equilibrium position of the contour. For example, for a bird contour containing 200 boundary points, the centroid coordinates are located at C(850, 420). The coordinates of the target center point are usually located inside or near the edge of an independent target region. For convex targets, such as a bird in flight with outstretched wings, the target center point must be inside the region. For concave targets, such as a bird with folded wings, the target center point may fall in the concave area. The target center point serves as the reference origin for morphological analysis. The positional relationship of all points in the boundary point set relative to the target center point reflects the shape distribution of the target. A uniform point distribution indicates a regular shape. In airport scenes, when birds are flying, the target center point corresponds to the approximate center of the bird's body. When the wings are spread, the target center point is biased towards the middle of the body, and when the wings are folded, the target center point may be biased towards the head or tail. The inter-frame displacement of the target center point reflects the bird's flight speed and direction.

[0028] A point distribution matrix is ​​constructed based on the target center point and the set of boundary points. The coordinate offsets of each point in the boundary point set relative to the target center point form offset vectors. The offset vector of the i-th point in the boundary point set is d_i = P_i - C = (x_i - x_c, y_i - y_c). The set of offset vectors constitutes the geometric description of the point distribution. The distribution characteristics of the offset vectors of the boundary point set reflect the main direction and scale of the target shape. The point distribution matrix is ​​defined as the covariance matrix of the offset vectors. The covariance matrix M is a 2×2 matrix, with diagonal elements representing the variance of the offsets in the x and y directions, and off-diagonal elements representing the covariance of the x and y offsets. The diagonal elements of the covariance matrix reflect the dispersion of points in the x and y directions, while the off-diagonal elements reflect the correlation in the x and y directions. An off-diagonal element of 0 indicates that the x and y directions are independent, and a larger off-diagonal element indicates that the point distribution has a tilted main direction. The point distribution matrix is ​​a symmetric positive definite matrix. Symmetry ensures that the eigenvalues ​​are real numbers, and positive definiteness ensures that the eigenvalues ​​are positive. In airport scenarios, the eigenvalues ​​of the point distribution matrix of birds vary considerably. The largest eigenvalue corresponds to a wing spread direction that is typically 3-5 times that of the secondary axis direction. These eigenvalue differences reflect the anisotropy of the shape.

[0029] The principal eigenvectors are obtained by eigenvalue decomposition of the point distribution matrix. The eigenvalues ​​of the point distribution matrix M are obtained by solving its characteristic equation, which expands to a quadratic equation, yielding two eigenvalues ​​λ1 and λ2. The eigenvalues ​​are sorted by size such that λ1 ≥ λ2. λ1 is defined as the principal direction corresponding to the largest eigenvalue, and λ2 is defined as the secondary direction corresponding to the smallest eigenvalue. The eigenvector corresponding to the largest eigenvalue λ1 is obtained by solving a system of linear equations. The eigenvector is normalized to unit length to form the principal eigenvector. The direction angle of the principal eigenvector is determined by the arctangent function, ranging from -180 degrees to 180 degrees, representing the angle between the principal eigenvector and the x-axis. The eigenvector corresponding to the smallest eigenvalue λ2 is perpendicular to the principal eigenvector. The secondary eigenvector is obtained by rotating the principal eigenvector by 90 degrees. The magnitude of the eigenvalue reflects the dispersion of the points in the corresponding direction, and the square root of the eigenvalue is approximately equal to the semi-axis length of the point distribution in that direction. In airport scenarios, when birds spread their wings, λ1 is approximately 4-9 times λ2. For example, when a pigeon spreads its wings to fly, λ1=1.8 and λ2=0.3, with a ratio of about 6, indicating that the dispersion of the wing spread direction is 6 times that of the body width direction. When the wings are folded, λ1 is approximately 9-25 times λ2, reflecting the transformation of the shape from wide and flat to slender.

[0030] The principal axis of the target is determined based on the principal eigenvector. The principal eigenvector defines the direction of the principal axis, which is a straight line passing through the center point of the target and oriented along the principal eigenvector. The projection length of each point in the boundary point set onto the direction of the principal eigenvector is obtained through the vector dot product. The maximum and minimum values ​​of the projection length determine the endpoints of the principal axis. The start and end points of the principal axis are determined by the extreme values ​​of the projection. The principal axis length L_major is the distance between the two endpoints, reflecting the maximum span of the target in the principal direction. When birds are in flight with their wings spread, the principal axis length is 30-60 pixels, corresponding to a wingspan of 1.5-3 meters at a distance of 200 meters. When gliding with their wings folded, the principal axis length is 15-30 pixels, corresponding to a body length of 0.8-1.5 meters. The direction angle of the principal axis is equal to the direction angle of the principal eigenvector, which describes the attitude of the target relative to the image coordinate system. The direction angle of birds flying horizontally is close to 0 degrees or 180 degrees, the direction angle deviates to -90 degrees when diving, and the direction angle deviates to 90 degrees when climbing. The principal axis of the target serves as the baseline axis for morphological analysis; lateral area division and aspect ratio measurement are both based on the principal axis. The secondary axis passes through the target's center point and is perpendicular to the principal axis. The direction of the secondary axis corresponds to the minimum eigenvalue, and its length L_minor is determined using a similar method. The principal and secondary axes constitute the target's local coordinate system, which simplifies the extraction and description of morphological features.

[0031] Lateral asymmetry is determined by measuring the area ratio of the left and right sides along the principal axis of the target. A straight line passing through the centroid C and perpendicular to the principal axis divides the independent target region into a right-side region along the positive direction of the principal axis and a left-side region along the negative direction. Pixels in the right-side region are identified by the sign of their x-coordinates in the target principal axis coordinate system; pixels with x-coordinates greater than 0 are assigned to the right side, and pixels with x-coordinates less than 0 are assigned to the left side. The number of pixels in the right-side region is denoted as N_right, and the number of pixels in the left-side region is denoted as N_left. The lateral area ratio R_area is defined as N_right / N_left, and R_area reflects the symmetry of the area on both sides. When birds fly symmetrically, the left and right wings are spread to the same degree, and R_area is close to 1.0, typically ranging from 0.95 to 1.05. When birds turn, the inner wing contracts and the outer wing expands, and R_area deviates from 1.0. For example, when a swallow makes a sharp turn, the inner wing contracts to 70% of its full extent, while the outer wing remains at 90% of its full extent, and the lateral area ratio R_area drops to 0.66, reflecting a significant left-right asymmetry. When turning right, R_area can reach 1.3-1.5, while when turning left, R_area drops to 0.65-0.77. Lateral asymmetry is defined as |1-R_area|, with a value ranging from 0 to positive infinity. The larger the value, the greater the left-right asymmetry. Lateral asymmetry less than 0.1 corresponds to symmetrical flight, 0.1-0.3 corresponds to normal flight, 0.3-0.5 corresponds to maneuvering, and greater than 0.5 corresponds to violent maneuvers or abnormal flight patterns. Birds in airport airspace typically perform turning maneuvers when avoiding obstacles or changing flight direction; abrupt changes in lateral asymmetry can identify these maneuvers.

[0032] The morphological asymmetry parameter is determined based on lateral asymmetry and the aspect ratio of the target's principal axis. Lateral asymmetry reflects left-right symmetry, while the aspect ratio R_aspect reflects elongation; together, they describe the target's morphological characteristics. The aspect ratio R_aspect is defined as the ratio of the principal axis length to the secondary axis length, i.e., R_aspect = L_major / L_minor, where L_major is the principal axis length and L_minor is the secondary axis length. The morphological asymmetry parameter integrates the two features using a weighted combination, with weight allocation considering the feature's discriminative ability and stability. Lateral asymmetry is sensitive to steering maneuvers but significantly affected by noise, while the aspect ratio is sensitive to flight conditions and has good measurement stability; the weight allocation is 0.6 for lateral asymmetry and 0.4 for the aspect ratio deviation term. The aspect ratio deviation term is defined as max(0, R_aspect - 2.5), which is 0 when R_aspect is less than 2.5 and increases linearly when it is greater than 2.5, reflecting an abnormally elongated morphology. The formula for the morphological asymmetry parameter is A_asym = 0.6 × |1 - R_area| + 0.4 × max(0, R_aspect - 2.5), where A_asym is the morphological asymmetry parameter, R_area is the lateral area ratio, and R_aspect is the aspect ratio. The value of A_asym ranges from 0 to approximately 2.0. A_asym less than 0.3 corresponds to normal symmetrical flight, such as cruising with wings outstretched; A_asym between 0.3 and 0.8 corresponds to slight asymmetry, such as slow turns or slight wing folding; A_asym between 0.8 and 1.5 corresponds to significant asymmetry, such as rapid turns or significant wing folding gliding; and A_asym greater than 1.5 corresponds to extreme asymmetry or non-bird targets, such as the slender fuselage of aircraft or drones. In airport scenarios, the morphological asymmetry parameter of birds is mainly concentrated in the range of 0.2-1.2, which covers most flight states.

[0033] In some embodiments, the step of extracting visual features from the independent target region to construct a visual descriptor includes: extracting continuous frame area changes in the independent target region to generate an area change sequence; performing periodic analysis on the area change sequence to form a flapping periodic feature; performing frequency quantization based on the flapping periodic feature to generate a wing flapping frequency; and constructing a visual descriptor based on the wing flapping frequency.

[0034] An area change sequence is generated by extracting continuous frame area changes within an independent target region. The area of ​​an independent target region in the current frame t is denoted as A(t), and the area equals the number of pixels within the independent target region, with the unit of area being pixels. The correspondence of the same target in consecutive frames is established using a target matching algorithm. The matching criteria include three indicators: centroid distance, color similarity, and depth consistency. Successful matching of independent target regions indicates that the preceding and following frames contain the same bird target. The area sequence of successfully matched targets in N consecutive frames forms an area change sequence. The area change sequence is denoted as {A(t), A(t+1), ..., A(t+N-1)}, with the sequence length N set to 15 frames corresponding to a 500ms time window. This window length can capture 1-3 complete wing flapping cycles. During bird flight, wing flapping causes periodic changes in the projected area in the observation direction; the projected area increases during wing spread and decreases during wing folding. The fluctuation amplitude of the area change sequence reflects the wing flapping amplitude. In airport scenes, the amplitude of the area variation sequence for large birds such as herons can reach 300-600 pixels, for medium-sized birds such as pigeons it is 150-300 pixels, and for small birds such as sparrows it is 50-150 pixels. The area variation sequence is affected by the viewing angle; the area change is most significant when viewed from the side, and smaller when viewed from the front or overhead. The viewing angle factor is estimated using the azimuth angle of the line connecting the target and the camera. The area variation sequence is normalized to a standard sequence with a mean of 0 and a standard deviation of 1. Normalization eliminates scale differences between targets at different distances.

[0035] Periodic analysis is performed on the area change sequence to identify the periodic characteristics of the pulsation. The normalized area change sequence is denoted as s(t), and the periodic analysis identifies the periodic component of the sequence through the autocorrelation function. The autocorrelation function is expressed by the formula... The formula is defined as follows: R(τ) represents the lag time, N represents the sequence length, and the summation range is from t=0 to N-τ-1. R(τ) reflects the similarity between the sequence and its own lag time τ. For periodic sequences, R(τ) peaks when τ equals a multiple of the period. R(0) represents the energy of the sequence, which is always equal to the normalized variance of 1. R(τ) gradually decreases as τ increases, and the τ value corresponding to the first significant positive peak is the flapping period T. The criteria for determining a significant positive peak are R(τ) > 0.3 and R(τ) being a local maximum. A threshold of 0.3 ensures that the detected peak is sufficiently significant. In airport scenes, the flapping period T of birds is usually 3-15 frames, corresponding to a time span of 100-500ms. Large birds have longer periods, while small birds have shorter periods. The flapping periodicity characteristics include three parameters: period length T, autocorrelation peak value R(T), and periodic stability. The magnitude of R(T) reflects the strength of periodicity. For example, when a heron flaps its wings at a frequency of 2Hz, the autocorrelation function shows a significant peak value R(15) = 0.72 at τ = 15 frames (corresponding to 0.5 seconds), indicating that the flapping periodicity is strong and stable. An R(T) close to 1 indicates extremely strong periodicity, while an R(T) less than 0.5 indicates weak periodicity or irregular flapping. Periodic stability is assessed by detecting the consistency of the lengths of multiple periods. A ratio of the standard deviation to the mean of three consecutive period lengths less than 0.15 is considered stable flapping, while a ratio greater than 0.3 is considered unstable flapping.

[0036] Wing flapping frequencies are generated using frequency quantization based on flapping period characteristics. The period length T in the flapping period characteristics is in frames; to convert it to time, it needs to be multiplied by the frame interval of 33ms. The time length of the flapping period characteristics is used for frequency conversion. The wing flapping frequency is defined as the reciprocal of the flapping period, in Hz. Typical bird wing flapping frequencies range from 1-15Hz. Large birds such as herons and cranes have frequencies of 1-3Hz corresponding to slow flapping; medium-sized birds such as pigeons and crows have frequencies of 3-6Hz corresponding to medium-speed flapping; and small birds such as sparrows and swallows have frequencies of 6-15Hz corresponding to fast flapping. Wing flapping frequency is an important feature dimension for distinguishing bird species. Different species have relatively little overlap in frequency ranges, resulting in high identification accuracy. Common bird species in airport airspace include pigeons (f=4-5Hz), crows (f=3-4Hz), sparrows (f=10-12Hz), and swallows (f=12-14Hz). Combining wing flapping frequency with body size can effectively identify these species. The measurement uncertainty of wing flapping frequency stems from periodic detection errors and frame rate limitations. A periodic detection error of approximately ±0.5 frames corresponds to a frequency error of ±10%, and the 30fps frame rate limits the frequency resolution to 2Hz. The temporal stability of wing flapping frequency is assessed through frequency consistency across multiple consecutive time windows. Frequency fluctuations within ±1Hz are considered stable flight, while fluctuations greater than ±2Hz indicate flight state transitions or high measurement uncertainty. During gliding flight, birds do not flap their wings; the area change sequence is non-periodic, and the autocorrelation function shows no significant peak. In this case, the wing flapping frequency is assigned a value of 0 to indicate gliding.

[0037] A visual descriptor is constructed based on wing flapping frequency. Wing flapping frequency is a core feature for distinguishing bird species. Large birds have a frequency of 1-3 Hz, medium-sized birds 3-6 Hz, and small birds 6-15 Hz, with minimal overlap in frequency ranges among different species. A dynamic feature set of flapping is expanded around wing flapping frequency, including eight dimensions: frequency, flapping amplitude, periodic stability, frequency time-varying rate, flapping phase, frequency fluctuation, amplitude variation coefficient, and phase continuity. A wing flapping frequency of 0 corresponds to gliding; in this case, the flapping amplitude is set to 0, and periodic stability is marked as gliding mode. Gliding mode identification is based on a wing flapping frequency threshold. The flapping amplitude is defined as the peak-to-peak value of the area change sequence. Combined with wing flapping frequency, it distinguishes flapping modes: high frequency and low amplitude correspond to rapid flapping, while low frequency and high amplitude correspond to slow gliding. Periodic stability reflects the consistency of the flapping period; stable frequency and stable period correspond to cruising flight, while frequency fluctuation and periodic instability correspond to maneuvering flight. The frequency time-varying rate is defined as the rate of change of wing flapping frequency within a continuous time window, reflecting the speed of flight state transitions. The flapping phase is defined as the relative position of the current moment within the flapping cycle, with phase 0 corresponding to the wing-spreading apex and phase 0.5 corresponding to the wing-folding apex. The visual descriptor supplements the flapping dynamic features with color distribution and morphological texture features to form a 32-dimensional feature vector. The color features include a 12-dimensional representation of the RGB color histogram of the independent target region, and the morphological features include 12 dimensions such as morphological asymmetry parameters and aspect ratio. The 32-dimensional features of the visual descriptor are normalized by subtracting the mean from each dimension and dividing by the standard deviation. The normalized feature values ​​are distributed in the range of -3 to 3.

[0038] Step S130: The morphological asymmetry parameter and the visual descriptor are fused to determine the preliminary recognition result. The distance and velocity of the preliminary recognition result are verified using millimeter-wave radar echo data to generate fusion error parameters. The credibility is evaluated based on the fusion error parameters and the morphological asymmetry parameter to determine the bird recognition target.

[0039] Specifically, the morphological asymmetry parameter and the visual descriptor are fused to determine the preliminary identification results. The morphological asymmetry parameter reflects the static morphological characteristics of the target. The numerical range of the morphological asymmetry parameter is 0 to approximately 2.0. A value less than 0.3 indicates normal symmetrical flight, 0.3-0.8 indicates slight asymmetry, 0.8-1.5 indicates significant asymmetry, and a value greater than 1.5 indicates extreme asymmetry or non-avian targets such as airplanes or drones. The visual descriptor reflects the dynamic flapping and appearance characteristics of the target. The 32-dimensional feature vector of the visual descriptor includes 8-dimensional flapping dynamic features such as flapping frequency, flapping amplitude, and periodic stability, 12-dimensional color features, and 12-dimensional morphological features. Among these, flapping frequency is the core dimension for distinguishing bird species, and the frequency ranges of birds of different body sizes overlap little. The two types of features complement each other to describe the comprehensive visual characteristics of birds. Feature fusion constructs a comprehensive feature vector through weighted concatenation. The morphological asymmetry parameter is a 1-dimensional scalar feature, and the visual descriptor is a 32-dimensional vector feature. After concatenation, a 33-dimensional fused feature vector is formed. In airport scenarios, bird species identification primarily relies on a combination of flapping frequency and body size. Pigeons flap at 4-5 Hz and are medium-sized, while sparrows flap at 10-12 Hz and are small; the two are significantly distinguishable by their flapping frequency. Feature vectors are fused and input into a classifier for pattern recognition. The classifier is pre-trained to obtain the decision boundary. The training process includes three stages: feature normalization, kernel function selection, and parameter optimization. Normalization eliminates dimensional differences across dimensions, the kernel function maps features to a high-dimensional space for non-linear classification, and parameter optimization uses cross-validation to determine the regularization coefficient and kernel parameters. The classifier outputs a confidence score for each category, ranging from 0 to 1. The category with the highest score is used as the initial category label. The initial identification result includes both a category label and a confidence score. The category label indicates the identified bird species, while the confidence score reflects the reliability of visual feature recognition. High confidence indicates a high match between the visual features and the category template, while low confidence indicates that the visual features are blurry or similar to multiple categories.

[0040] In some embodiments, the step of using the millimeter-wave radar echo data to perform distance and velocity verification on the preliminary identification result to generate fusion error parameters includes: performing time-frequency analysis on the millimeter-wave radar echo data to generate micro-Doppler features; performing wing flapping pattern matching on the micro-Doppler features to form a flapping matching degree; performing consistency verification between the flapping matching degree and the preliminary identification result to generate a verification deviation; and determining the fusion error parameters based on the verification deviation.

[0041] Micro-Doppler features are generated through time-frequency analysis of millimeter-wave radar echo data. Millimeter-wave radar echo data records the amplitude and phase information of the echo signal for each detection cycle. The phase change rate of the echo signal reflects the radial velocity of the target, and periodic phase modulation reflects the target's micro-motion characteristics. Time-frequency analysis decomposes the millimeter-wave radar echo data into a two-dimensional time-frequency spectrum using short-time Fourier transform. The horizontal axis of the spectrum represents time, and the vertical axis represents Doppler frequency. Color or grayscale indicates energy intensity. The periodic motion generated by bird wing flapping in millimeter-wave radar echo data is manifested as modulation sidebands around the main Doppler frequency in the spectrum. The main Doppler frequency corresponds to the bird's flight speed, and the modulation sideband frequencies correspond to the wing flapping frequency. Micro-Doppler features are extracted from the time-frequency spectrum and include three parameters: main Doppler frequency, modulation depth, and modulation period. The main Doppler frequency is determined by the energy center of the spectrum, which corresponds to the target's average radial velocity. Modulation depth is defined as the offset of the sideband frequency relative to the main frequency. It reflects the speed amplitude of wing flapping and is positively correlated with wing length and flapping speed. Modulation period is defined as the time interval between repeated sideband occurrences. The reciprocal of the modulation period corresponds to the flapping frequency. Micro-Doppler characteristic flapping frequency measurement has stability and anti-interference capability.

[0042] For example, the step of matching wing flapping patterns to form a flapping matching degree based on the micro-Doppler features includes: extracting modulation depth and modulation period based on the micro-Doppler features to form flapping dynamic parameters; constraining the flapping dynamic parameters by body size range to generate a body size confidence interval; searching a bird species flapping template library within the body size confidence interval to obtain a candidate flapping pattern set; and calculating the similarity between the micro-Doppler features and the candidate flapping pattern set to form a flapping matching degree.

[0043] Modulation depth and modulation period are extracted based on micro-Doppler features to form flapping dynamics parameters. The frequency shift of the modulation sideband in the time-spectrum diagram of the micro-Doppler feature is defined as the modulation depth. The modulation depth of the micro-Doppler feature is measured from the center of the main Doppler frequency to the peak of the upper or lower sideband. Theoretically, the frequency shifts of the upper and lower sidebands are symmetrical, but in actual measurements, deviations may occur due to noise and target attitude. The physical meaning of modulation depth is the amplitude of the radial velocity change caused by wing flapping. In the micro-Doppler feature, when the wing flaps towards the radar, the radial velocity increases, producing a positive frequency shift; when the wing flaps away from the radar, the radial velocity decreases, producing a negative frequency shift. The amplitude of the frequency shift reflects the flapping velocity. Modulation depth is converted into velocity amplitude through the relationship between Doppler frequency and radar wavelength, with the conversion formula v_mod=Δf×λ / 2, where Δf is the modulation depth in Hz, λ is the radar wavelength in meters, and v_mod is the velocity amplitude in m / s. Significant differences in modulation depth exist among birds of different sizes in airport scenarios. Large birds, such as herons, have long wings and large flapping amplitudes, resulting in high modulation depth. Medium-sized birds, such as pigeons, have moderate modulation depth, while small birds, such as sparrows, have short wings, resulting in low modulation depth. The modulation period is defined as the time interval between the repetition of sidebands in the time-spectrum graph. The modulation period is extracted by the time series of sideband peaks, and the time difference between adjacent peaks is one modulation period. The reciprocal of the modulation period is the flapping frequency, which is a key parameter for bird species identification. Different bird species have relatively little overlap in their flapping frequency ranges, with a clear pattern of low flapping frequencies in large birds and high flapping frequencies in small birds. The flapping dynamics parameter integrates the two core parameters of modulation depth and modulation period to form a two-dimensional feature vector (v_mod, T_cycle). This vector describes the dynamic characteristics of wing flapping, including both amplitude and rhythm information.

[0044] Body size confidence intervals are generated by constraining the flapping dynamics parameters to a range. A correlation exists between bird body size and flapping dynamics parameters, stemming from fundamental principles of flight dynamics: large birds have long wings and large flapping amplitudes but low frequencies, while small birds have short wings and small flapping amplitudes but high frequencies. The modulation depth of the flapping dynamics parameters reflects body size; a large modulation depth indicates long wings and a large body, while a small modulation depth indicates short wings and a small body. The body size range is defined by modulation depth thresholds: modulation depths exceeding the upper threshold correspond to large birds such as herons and cranes; modulation depths in the middle range correspond to medium-sized birds such as pigeons and crows; and modulation depths below the lower threshold correspond to small birds such as sparrows and swallows. A body size confidence interval is defined as the body size range inferred based on modulation depth, containing two endpoints: a lower bound and an upper bound. The endpoint values ​​are estimates of body length or wingspan. The confidence level of the body size confidence interval is determined by measurement accuracy; higher modulation depth measurement accuracy results in higher confidence levels, while higher measurement noise results in lower confidence levels. In airport scenarios, radar detection range affects the measurement accuracy of modulation depth. The greater the detection range, the lower the echo signal-to-noise ratio and the larger the measurement error; conversely, shorter detection ranges result in higher measurement accuracy. The body size confidence interval is used to screen candidate bird species. Birds whose nominal body size falls within the confidence interval are retained as candidates, while those whose nominal body size exceeds the interval are excluded.

[0045] Candidate flapping pattern sets are obtained by searching a bird species flapping template library within the body size confidence interval. The bird species flapping template library contains standard flapping patterns of common birds found at airports. For each bird species, the template records features such as body size range, flapping frequency range, and modulation depth range. The template search is primarily constrained by the body size confidence interval; a template is included in the candidate set when its body size range overlaps with the body size confidence interval. Different bird species exhibit distinguishable differences in their flapping patterns. Pigeons and crows have similar body size ranges but significantly different flapping frequencies; pigeons have a higher flapping frequency, while crows have a lower flapping frequency, allowing them to be distinguished by frequency characteristics. Sparrows and swallows are both small birds, but swallows have a slightly higher flapping frequency than sparrows. They also differ in their flight patterns, with swallows more often employing an alternating gliding and flapping flight method. The candidate flapping pattern set includes all templates that pass the body size screening; the body size constraint significantly reduces the number of candidates and improves matching efficiency. The modulation period of the flapping dynamics parameters and the frequency range of each template in the candidate flapping pattern set are used for secondary screening. Templates whose modulation period corresponds to a frequency within the template's frequency range are retained, while templates with mismatched frequencies are discarded. The candidate flapping pattern set is further reduced after this secondary screening, significantly improving the matching accuracy. When the candidate flapping pattern set is empty, it indicates that the micro-Doppler features do not match any known bird species, and the target is determined to be an unknown bird species or a non-avian target.

[0046] The similarity between micro-Doppler features and the candidate flapping pattern set is calculated to form the flapping match score. When the candidate flapping pattern set is empty, the flapping match score is set to 0 and the best-matching bird species is labeled "unknown," skipping the similarity calculation. Micro-Doppler features contain two-dimensional feature vectors: modulation depth and modulation period. Each template in the candidate flapping pattern set is a standard two-dimensional feature vector. Similarity is measured using Euclidean distance in the feature space. Euclidean distance measures the difference between the feature vector and the template; a smaller distance indicates a closer match. Feature vectors are normalized before distance calculation to eliminate dimensional differences and ensure balanced contributions from each dimension. The similarity score is obtained from the distance using the formula S=exp(-d² / σ²), with a value ranging from 0 to 1. Each template in the candidate flapping pattern set receives a similarity score, and the template with the highest similarity score corresponds to the flapping match score value. The flapping match score ranges from 0 to 1; a larger value indicates a better match between the micro-Doppler features and the bird species template. In the airport scene, the micro-Doppler feature modulation depth of 45Hz and modulation period of 0.22 seconds showed a high degree of match with the pigeon template, with a flapping motion matching degree of 0.85. The flapping motion matching degree was accompanied by a bird species label of the best matching template, which was inferred from radar data. This label was compared with the species label of visual recognition to achieve multimodal verification.

[0047] A consistency check is performed based on the flapping match degree and the preliminary identification results to generate a check deviation. The species label of the preliminary identification result is compared with the best-matching bird species label of the flapping match degree. If the labels of the preliminary identification result and the flapping match degree match, the species is considered consistent; if the labels do not match, the species is considered conflicting. A smaller check deviation is set when the species are consistent, reflecting a match between visual and radar identification results; a larger check deviation is set when the species conflict occurs, reflecting a discrepancy in identification results. The confidence score of the preliminary identification result is numerically compared with the flapping match degree. If the values ​​of the two are close, it indicates that the visual and radar identification are equally reliable; a large difference indicates that one identification is more reliable and the other has higher uncertainty. The absolute value of the confidence score difference is used as the confidence score deviation. The check deviation combines the species consistency deviation and the confidence score consistency deviation, and is determined by a weighted sum. A larger deviation value is assigned when there is a species conflict to emphasize the importance of species identification; a smaller deviation value is assigned when the species are consistent to indicate a basic match. The confidence score deviation directly uses a normalized value to reflect the degree of difference in confidence scores. The verification bias D_check = w_species × D_species + w_confidence × D_confidence, where D_check is the verification bias, D_species is the species consistency bias, D_confidence is the confidence difference bias, w_species is the weighting coefficient of the species bias, and w_confidence is the weighting coefficient of the confidence bias. The verification bias ranges from 0 to 1. A smaller value indicates greater consistency between visual and radar detection. A verification bias close to 0 indicates high consistency and a highly reliable identification result, while a verification bias close to 1 indicates a serious conflict and an unreliable identification result. In airport scenarios, the verification bias for bird targets is typically small, indicating that visual and radar detection are basically consistent, while the verification bias for aircraft targets is typically large, indicating a significant conflict between visual and radar detection.

[0048] The fusion error parameter is determined based on the verification deviation. The verification deviation reflects the degree of consistency between visual and radar recognition. This consistency is converted into an error parameter through a mapping function of the deviation value. The fusion error parameter is defined as an indicator reflecting the fusion quality. High consistency indicates a small error and good fusion quality, while low consistency indicates a large error and poor fusion quality. When the verification deviation is small, visual and radar recognition are highly consistent, and a small fusion error parameter indicates a very small error and highly reliable recognition results. When the verification deviation is moderate, consistency is good, and a moderate fusion error parameter indicates a moderate error and reliable recognition results. When the verification deviation is large, consistency is poor or there are conflicts, and a large fusion error parameter indicates a very large error and unreliable recognition results requiring further verification. The mapping formula for the fusion error parameter is E_fusion=min(1.0,k×D_check), where E_fusion is the fusion error parameter, D_check is the verification deviation, and k is the mapping coefficient. A mapping coefficient greater than 1 causes the error parameter to grow faster than the deviation, reflecting sensitivity to conflicts. The value range of the fusion error parameter is strictly limited to 0-1. The closer the value is to 0, the better the fusion quality; the closer the value is to 1, the worse the fusion quality. In airport scenarios, the fusion error parameter for bird targets is typically small, and within this range, visual and radar identification is generally consistent and reliable. For non-bird targets such as aircraft, drones, and balloons, the fusion error parameter is typically large, and within this range, significant discrepancies between visual and radar identification indicate that the target is not avian.

[0049] In some embodiments, determining the bird identification target based on the credibility assessment of the fusion error parameter and the morphological asymmetry parameter includes: establishing bird and non-bird hypotheses based on the fusion error parameter and the morphological asymmetry parameter to generate a multi-class hypothesis set; accumulating multi-frame evidence for the multi-class hypothesis set to form a hypothesis support sequence; implementing competitive elimination in the hypothesis support sequence to generate a dominant hypothesis; and determining the bird identification target based on the dominant hypothesis.

[0050] Multiple hypothesis sets are generated by establishing bird and non-bird hypotheses based on fusion error parameters and morphological asymmetry parameters. The fusion error parameter reflects the consistency between visual and radar data, while the morphological asymmetry parameter reflects the target's morphological characteristics. These two parameters are combined to determine the target type. The initial confidence level of the bird hypothesis is set according to the fusion error parameter. A smaller fusion error parameter indicates better consistency between visual and radar data, making the bird hypothesis more credible and thus resulting in a higher initial confidence level. Conversely, a larger fusion error parameter indicates poor consistency, leading to a lower initial confidence level for the bird hypothesis. When the candidate flapping pattern set is empty or the flapping pattern matching is too low, the initial confidence level of the non-bird hypothesis directly increases to above 0.7. At this point, the non-bird hypothesis dominates, indicating that the micro-Doppler features do not match any known bird species. The initial confidence level of the non-bird hypothesis is set to 1 minus the confidence level of the bird hypothesis. The complementary sum of the confidence levels of the two hypotheses (to 1) ensures probability normalization. The morphological asymmetry parameter serves as a hypothesis correction factor to adjust the initial confidence level. Within the normal bird morphological range, the parameter maintains the confidence level of the bird hypothesis. A parameter that is too small suggests an overly symmetrical shape, potentially indicating a stationary or spherical object, thus reducing the confidence level of the bird hypothesis. A parameter that is too large suggests an overly elongated shape, potentially indicating an airplane or rod-like object, also reducing the confidence level of the bird hypothesis. In airport scenes, the morphological asymmetry parameter of airplanes typically far exceeds the range for birds. Airplanes have elongated fuselages and high aspect ratios when wings are spread, effectively eliminating interference from airplanes. The morphological asymmetry parameter of drones partially overlaps with that of birds. Small multi-rotor drones have nearly symmetrical shapes, while fixed-wing drones tend to be elongated, requiring consideration of other features such as flight speed and flapping frequency for comprehensive judgment. The multi-class hypothesis set contains two elements: bird and non-bird hypotheses. Each hypothesis includes a confidence level attribute. The confidence level distribution of the multi-class hypothesis set reflects the current frame's tendency to determine the target type. Large differences in confidence levels indicate high certainty in the judgment, while similar confidence levels indicate high uncertainty in the judgment.

[0051] A hypothesis support sequence is formed by accumulating evidence across multiple frames for various hypothesis sets. The hypothesis support of the current frame's various hypothesis sets and the previous frame's hypothesis support is updated through temporal fusion. The fusion strategy for the various hypothesis sets employs an exponentially weighted moving average method. The current frame's confidence and historical support are combined with certain weights. The current frame weight reflects the contribution of the latest observation, while the historical support weight reflects the smoothing effect of historical information. The update formula for the bird hypothesis support is H_bird(t) = α × P_bird(t) + (1-α) × H_bird(t-1), where H_bird(t) is the bird hypothesis support in the current frame t, P_bird(t) is the bird hypothesis confidence in the current frame t derived from the various hypothesis sets, H_bird(t-1) is the bird hypothesis support in the previous frame, α is the current frame weight coefficient, and t is the time index. The support for non-bird hypotheses is updated similarly, and the two types of support are normalized to ensure a sum of 1. The hypothesis support sequence records the support evolution trajectory across multiple consecutive frames, and the trend of the sequence reflects the direction of change in hypothesis strength. A sustained increase in support for the bird hypothesis indicates accumulated evidence supporting the bird determination, while a sustained decrease in support suggests evidence favoring non-avian targets. The fluctuation range of the support sequence reflects the stability of the determination; small fluctuations indicate stable determination, while large fluctuations indicate unstable determination or changes in target status. In airport scenarios, the support for the bird hypothesis for bird targets typically fluctuates stably within a high range, while the support for the non-avian hypothesis for aircraft targets also remains stable within a high range. The time it takes for the hypothesis support sequence to reach a steady state depends on the stability of the weighting parameters and target features; the support at the steady state serves as the final determination criterion.

[0052] A competitive elimination process is implemented within the hypothesis support sequence to generate a dominant hypothesis. The support of the bird hypothesis and the non-bird hypothesis compete to determine the final judgment; the hypothesis with higher support is the dominant hypothesis, and the hypothesis with lower support is the suboptimal hypothesis. The competitive elimination is determined by the support difference. The difference between the support of the bird hypothesis and the non-bird hypothesis in the hypothesis support sequence reflects the relative strength of the two hypotheses. A significantly positive difference indicates that the bird hypothesis is dominant, a significantly negative difference indicates that the non-bird hypothesis is dominant, and a difference close to zero indicates that the two hypotheses are evenly matched and the judgment is uncertain. Determining a dominant hypothesis requires sustained dominance rather than momentary dominance; the dominant hypothesis is confirmed only if the judgment condition is met for multiple consecutive frames. This persistence requirement avoids misjudgments caused by fluctuations in a single frame. In the airport scene, after bird targets are identified stably, the support of the bird hypothesis is significantly higher than that of the non-bird hypothesis, and the support difference meets the elimination condition, making the bird hypothesis the dominant hypothesis. For aircraft targets, the support of the non-bird hypothesis is significantly higher than that of the bird hypothesis, and the support difference meets the elimination condition, indicating that the target is not avian. For uncertain targets, a support difference that hovers within a small range is insufficient to form a dominant hypothesis. Such targets are marked as pending confirmation, requiring longer observation times or human intervention. Once a dominant hypothesis is established, it exhibits inertia; support fluctuations in subsequent frames do not immediately alter the dominant determination. Multiple consecutive frames of contradictory evidence are needed to overturn the dominant hypothesis. The confidence level of a dominant hypothesis is defined as its support value; higher confidence indicates a more reliable determination. The dominant hypothesis is accompanied by a hypothesis type label (birds or non-birds), which is output as the target type determination result.

[0053] Bird targets are identified based on a dominant hypothesis. The type label of the dominant hypothesis directly determines whether the target is a bird, and the type label is derived from the competition results of the hypothesis support sequence. When the dominant hypothesis of bird wins, the target is identified as a bird, based on the fact that the support of the bird hypothesis is significantly higher than that of the non-bird hypothesis and remains stable. When the dominant hypothesis of non-bird wins, the target is identified as a non-bird, indicating that the target characteristics are significantly inconsistent with known bird patterns. In an uncertain state, the target is identified as a target to be confirmed, indicating that the support of the two hypotheses is close and cannot form a dominant one. The specific species of the bird target is determined by combining the preliminary visual recognition results with the best matching species of radar flapping, and a multimodal fusion strategy is used to improve the recognition accuracy. When the two species are consistent, that species is directly used as the final species label; when the two species conflict, the species label with higher confidence is selected as the final result. The data structure of bird target identification includes six core fields: type determination, species label, confidence, location coordinates, appearance features, and timestamp. The type determination result is either bird or non-bird, and the species label is the specific bird species. The stability of the recognition is verified during multi-frame consistency checks. Confidence score is a quantitative indicator of recognition reliability, derived from the support level of the dominant hypothesis. Position coordinates represent the target's three-dimensional spatial location, obtained by combining the image coordinates of the centroid of the independent target region with the depth value of the depth map through camera calibration parameters, and represented in the camera coordinate system. Position coordinates provide a basis for position prediction during inter-frame matching. Appearance features are a core dimension subset of the visual descriptor, which, along with velocity information, constitutes the target state description during inter-frame matching. A timestamp records the judgment time, ensuring temporal alignment of multi-frame data during trajectory association. In airport scenes, the output frequency for bird target recognition is consistent with the video frame rate, outputting a recognition result once per frame.

[0054] Step S140: Perform inter-frame image matching to generate target trajectory sequence for bird identification target, perform continuity verification on target trajectory sequence to obtain trajectory fluctuation parameters and extract trajectory periodic features, and determine dynamic feature data based on trajectory fluctuation parameters and trajectory periodic features.

[0055] In some embodiments, the step of performing inter-frame image matching to generate a target trajectory sequence for the bird identification target includes: extracting the target position and appearance features of the current frame based on the bird identification target to generate a target state vector; performing feature similarity matching in adjacent frames based on the target state vector to form an inter-frame association; performing occlusion detection on the inter-frame association to identify trajectory interruption points and generate occlusion markers; and performing trajectory interpolation and identity preservation based on the occlusion markers to generate a target trajectory sequence.

[0056] Based on bird identification targets, the target's position and appearance features in the current frame are extracted to generate a target state vector. The spatial position information of the bird identification target includes three-dimensional coordinates (x, y, z), with the origin of the coordinate system located at the optical center of the stereo camera. The position coordinates of the i-th target in the current frame are denoted as P_i, obtained through stereo vision or radar ranging. Stereo vision calculates depth by the disparity between the left and right camera images; disparity is inversely proportional to depth, and stereo vision provides high accuracy at close range. Radar ranging calculates distance by the round-trip time of electromagnetic waves, providing stable measurements at long range. The fusion of stereo vision and radar ranging employs a weighted average strategy, with weights dynamically adjusted according to the measurement distance. The appearance features of the bird identification target are extracted from the visual descriptor, which contains a 32-dimensional feature vector. The subset of appearance features that contributes the most to target identification is selected based on the features' discriminative power and stability. Discriminative power is measured by the ratio of inter-class variance to intra-class variance, and stability is measured by the feature variation coefficient of consecutive frames. Beating frequency, beating amplitude, dominant color tone, and aspect ratio are commonly selected as appearance features. These features remain relatively stable over a short period and exhibit significant differences between different targets. The appearance feature vector is denoted as F_i, and the feature vector is normalized to the unit norm to eliminate differences in feature magnitude. The velocity estimation for bird targets is obtained by the difference between the current frame position and the previous frame position. The velocity vector V_i is calculated using the position difference and the frame interval. The target state vector integrates position, appearance, and velocity information to form a comprehensive state description S_i={P_i,F_i,V_i}. The state vector contains complete information about the target's spatial position, visual appearance, and motion state. These three types of information complement each other during inter-frame matching, providing multi-dimensional support.

[0057] Inter-frame associations are formed by feature similarity matching between adjacent frames based on target state vectors. The target state vector set in the current frame t is paired with the state vector set in the previous frame t-1 for similarity measurement. The similarity measurement of target state vectors is comprehensively evaluated from three dimensions: position, appearance, and velocity. Position similarity is measured by the distance between the predicted position and the actual position. The predicted position in the current frame is formed by adding the target's position in the previous frame to the velocity estimate. The predicted position of the target state vector is based on the assumption of uniform motion, assuming that the target maintains uniform linear motion within a short time interval. The predicted position provides the expected position of the target in the current frame, and the distance between the predicted position and the actual position in the current frame is used as the positional dissimilarity; the smaller the distance, the closer the positions. Appearance similarity is measured by the cosine similarity of the feature vectors. Cosine similarity calculates the cosine value of the angle between two vectors; the closer the cosine value is to 1, the more similar the appearance. Velocity similarity is comprehensively evaluated by the consistency of the direction and magnitude of the velocity vectors. Directional consistency is measured by the angle between the velocity vectors, and magnitude consistency is measured by the difference in velocity magnitude. The comprehensive similarity is determined by a weighted sum, with the weight allocation reflecting the contribution of each dimension to the matching. Inter-frame association is established using the Hungarian algorithm for globally optimal matching. The algorithm takes a similarity matrix as input and outputs a one-to-one match that maximizes the total similarity. The Hungarian algorithm finds the optimal matching scheme globally, avoiding local suboptimal solutions. Successfully matched target pairs establish inter-frame associations, which are recorded as association pairs (i,j). In airport scenes, the inter-frame association similarity of stably tracked targets is usually high, decreasing with occlusion or rapid maneuvers. Targets in the current frame without established associations are considered new targets, while targets in the previous frame without established associations are considered targets that may disappear or be occluded.

[0058] Occlusion detection is implemented for inter-frame correlation to identify trajectory interruption points and generate occlusion markers. Inter-frame correlation interruptions may arise from the actual disappearance of the target or temporary occlusion. Occlusion detection distinguishes between these two cases to prevent premature trajectory termination. Targets present in the previous frame but not matched in the current frame are defined as potential occluders. The predicted position of potential occluders in interrupted inter-frame correlation is inferred from the motion trend of the previous frame. The depth information of the predicted position is compared with the depth map of the current frame, achieved using depth maps obtained from binocular stereo vision or radar. A depth value close to the target's historical depth indicates no occlusion; a significantly smaller depth value indicates the presence of a foreground occluder, suggesting a closer object at the predicted position. Foreground occluders may include other birds, aircraft, building edges, etc. The occlusion determination criteria for inter-frame correlation combine depth deviation and occlusion area area; a large depth deviation and a high proportion of the occluded area indicate occlusion. Occlusion markers record the starting frame, duration, and type of occlusion. Occlusion types include partial occlusion and complete occlusion. In partial occlusion, the target remains partially visible; in complete occlusion, the target is completely invisible. In airport scenes, occlusion events mainly occur when multiple bird targets occlude each other or when birds fly over building edges. For example, when two pigeons fly at the same height, the pigeon in front occludes the pigeon behind for 2-3 frames, or a bird flying over the edge of the terminal building is briefly occluded by the building outline for 1 frame. The occlusion marker includes the target state vector before and after occlusion, used for trajectory interpolation during occlusion and target re-identification after occlusion ends.

[0059] Based on occlusion markers, trajectory interpolation is performed to complete the target trajectory sequence and maintain its identity. During the occlusion period indicated by the occlusion markers, there is no actual observation data; trajectory interpolation fills in missing frames using motion model prediction. The uniform linear motion model is suitable for short-term occlusion; the position during occlusion is estimated by adding the average velocity and time difference to the occlusion start position. The uniform model assumes the target maintains uniform linear flight during occlusion. The uniform acceleration motion model is suitable for occlusion during maneuvering flight; the position estimation includes an acceleration term, which is estimated from the velocity changes in previous frames. The confidence of the interpolated position decreases with increasing occlusion duration; the longer the occlusion time, the larger the interpolation error. Identity maintenance is achieved through target re-identification after occlusion ends. At the end of occlusion, the new target in the current frame is matched with the target state vector of the target before occlusion, based on a comprehensive similarity including position, appearance, and velocity. If the match is successful, the target is identified as the same target, the original trajectory ID remains unchanged, and tracking continues, maintaining trajectory continuity. If the match fails, the target is identified as a different target, the original trajectory is terminated, and a new trajectory is started. The target trajectory sequence integrates all information from the original observations, interpolation completion, and identity preservation to form a time-continuous trajectory data structure. The position sequence of the target trajectory sequence includes the observation position and the interpolated position, and records the spatial coordinates of the target in each frame for subsequent continuity verification.

[0060] The continuity of the target trajectory sequence is verified to obtain trajectory fluctuation parameters and extract trajectory periodic features. The temporal variation of position coordinates in the target trajectory sequence reflects the smoothness and periodicity of bird flight paths. The trajectory position sequence of the target trajectory sequence is fitted with a smooth curve using cubic spline interpolation, and the interpolated curve represents the ideal flight path. The deviation between the actual position and the interpolated curve is defined as the position residual. The root mean square value of the position residual is defined as the trajectory fluctuation parameter, expressed by the formula RMS = The trajectory is defined as follows: P_i is the actual position, P_fit_i is the fitted position, and k is the number of trajectory points. The trajectory fluctuation parameter reflects the stability of the flight path; small fluctuations indicate a smooth trajectory, while large fluctuations indicate maneuvering flight. In the airport scenario, the trajectory fluctuation parameter of a pigeon flying in a straight line is approximately 0.4 meters, increasing to 0.9 meters during a sharp turn. The trajectory fluctuation parameter also includes horizontal and vertical fluctuations; horizontal fluctuations originate from adjustments in flight direction, while vertical fluctuations originate from changes in altitude. Bird flight trajectories exhibit periodic undulations driven by wing flapping. The vertical coordinates of the target trajectory sequence are extracted to form an altitude time series. The autocorrelation function is applied to identify the periodicity of the altitude sequence; the lag time corresponding to the autocorrelation peak is the fluctuation period T_z. The trajectory periodicity feature includes three parameters: fluctuation period, fluctuation amplitude, and periodic stability. The fluctuation amplitude is the standard deviation of the altitude sequence, and the periodic stability is the coefficient of variation of the continuous period length. In the airport scenario, the fluctuation period of a heron is approximately 0.4 seconds, with an amplitude of approximately 0.3 meters, while the fluctuation period of a sparrow is approximately 0.09 seconds, with an amplitude of approximately 0.06 meters.

[0061] Dynamic feature data is determined based on trajectory fluctuation parameters and trajectory periodicity characteristics. The dynamic feature data is constructed through feature vector concatenation. The trajectory fluctuation parameters provide three dimensions: root mean square (RMS) value, horizontal fluctuation, and vertical fluctuation. The RMS value reflects overall path stability, horizontal fluctuation reflects the frequency of directional adjustments, and vertical fluctuation reflects the amplitude of altitude changes. The trajectory periodicity characteristics provide three dimensions: fluctuation period, fluctuation amplitude, and periodic stability. The fluctuation period reflects wing-flapping rhythm, fluctuation amplitude reflects wing-flapping drive intensity, and periodic stability reflects flight regularity. These are concatenated to form a 6-dimensional basic feature vector. The basic feature vector describes the geometric and rhythmic characteristics of the flight path. Geometric characteristics reflect path smoothness and spatial distribution, while rhythmic characteristics reflect the periodic regularity of wing-flapping drive. These two types of characteristics complement each other in describing flight dynamics. The dynamic feature data also includes the trajectory's kinematic parameters. The average speed, maximum speed, and rate of change of speed supplement the description of motion characteristics. The average speed reflects the cruising speed level, the maximum speed reflects acceleration capability, and the rate of change of speed reflects maneuverability. The kinematic parameters describe the target's motion capabilities. Kinematic parameters, combined with fluctuation and periodic features, form a complete dynamic feature data vector. Each dimension of the vector corresponds to an aspect of flight dynamics, and the multi-dimensional features comprehensively describe the complex characteristics of flight dynamics. Each dimension of the feature vector is normalized by subtracting the mean from each dimension and dividing by the standard deviation. Normalization eliminates dimensional differences, ensuring a balanced contribution from each dimension. The normalized feature values ​​are distributed within the range of a standard normal distribution.

[0062] Step S150: Combine dynamic feature data with bird identification target to generate multi-frame identification result sequence, perform consistency check on multi-frame identification result sequence to obtain identification inconsistency parameter, and output bird target identification and tracking results according to the identification inconsistency parameter.

[0063] Specifically, a multi-frame recognition result sequence is generated by combining dynamic feature data with bird identification targets. The dynamic feature data includes a 9-dimensional feature vector comprising trajectory fluctuation parameters, trajectory periodicity features, and kinematic parameters. The bird identification target includes type determination, species label, and confidence level. The type determination distinguishes between avian and non-avian targets, the species label indicates the specific bird species (e.g., pigeon, sparrow, swallow), and the confidence level reflects the recognition reliability (range 0-1). The two types of data are fused to form the comprehensive recognition information for each frame. A single-frame recognition result includes a species label, confidence level, and dynamic feature data. The species label and confidence level are derived from the bird identification target, while the dynamic feature data reflects flight dynamics. Multiple consecutive frames of recognition results are arranged chronologically to form a multi-frame recognition result sequence. The sequence length is set to cover multiple complete flapping cycles of the bird, providing sufficient temporal samples for consistency verification. The multi-frame recognition result sequence is denoted as {R(t), R(t+1), ..., R(t+N-1)}, where R(t_i) is the recognition result of the i-th frame, and N is the sequence length. The consistency of the category labels in the multi-frame recognition result sequence reflects the stability of the recognition. Category labels remaining unchanged across multiple frames indicates stable and reliable recognition, while frequent switching of category labels indicates unstable recognition and potential misidentification. The temporal variation of confidence reflects the fluctuation in recognition quality. A consistently high confidence level indicates good recognition quality, while low confidence or significant fluctuations indicate poor recognition quality.

[0064] Consistency testing is performed on multi-frame recognition result sequences to obtain recognition inconsistency parameters. The consistency test of multi-frame recognition result sequences is comprehensively evaluated through three dimensions: category stability, confidence fluctuation, and dynamic feature continuity. Category stability testing statistically analyzes the frequency of occurrence of various category labels in the sequence; the category with the highest frequency is defined as the dominant category, and the proportion of the dominant category reflects category stability. Category inconsistency is defined as 1 minus the proportion of the dominant category; the lower the inconsistency, the more consistent the categories. Confidence fluctuation testing assesses the confidence level using the standard deviation of the confidence sequence; the standard deviation reflects the degree of confidence fluctuation. Confidence fluctuation is defined as the normalized value of the standard deviation, with normalization eliminating the influence of the mean level. Dynamic feature continuity testing assesses the inter-frame differences in dynamic feature data; the Euclidean distance between dynamic feature data in adjacent frames is used as the inter-frame jump variable. The mean of the jump variable sequence reflects the average rate of change of the dynamic feature data, and the proportion of frames with larger jump variables reflects the frequency of abrupt changes. Dynamic feature discontinuity is defined as a weighted combination of the frequency of abrupt changes and the average jump variable. The inconsistency parameter I_inconsist is determined by combining three indicators using the formula I_inconsist = w_species × I_species + w_confidence × I_confidence + w_dynamic × I_dynamic, where I_species represents the species inconsistency, I_confidence represents the confidence fluctuation, I_dynamic represents the dynamic feature discontinuity, and w is the weighting coefficient for each indicator. The inconsistency parameter ranges from 0 to 1; a smaller value indicates more consistent and reliable identification, while a larger value indicates less consistent and reliable identification. In an airport scenario, when pigeons fly steadily, the species labels remain consistent, the confidence level is stable, and the dynamic feature data changes smoothly, resulting in a small inconsistency parameter. When the same pigeon turns rapidly, the dynamic feature data changes drastically, and the confidence level may fluctuate, leading to an increased inconsistency parameter, but the species labels remain consistent.

[0065] Bird target identification and tracking results are graded based on the identification inconsistency parameter. The results are categorized into three levels: high confidence, medium confidence, and low confidence. Results with a smaller identification inconsistency parameter are rated as high confidence, those with a medium confidence parameter as medium confidence, and those with a larger confidence parameter as low confidence. The grading threshold is determined by statistically analyzing the distribution of identification inconsistency parameters in the sample. High confidence results are directly output for early warning and bird deterrence decisions; medium confidence results are output with an additional warning label; and low confidence results are marked as pending confirmation and require manual review. The data structure for bird target identification and tracking results includes six fields: species label, confidence level, 3D position, flight speed, trajectory prediction, and timestamp. The species label indicates the specific bird species, the confidence level indicates the identification quality, the 3D position indicates the target's spatial location, the flight speed indicates the target's speed, the trajectory prediction indicates the target's future position, and the timestamp indicates the identification time. Trajectory prediction is obtained through linear extrapolation of the current position and speed. The prediction time is calculated by generating multiple prediction points from several future times. The predicted trajectory is used to assess whether the bird is flying towards the runway or entering hazardous airspace. The distance between the predicted trajectory and the runway centerline determines the threat level; smaller distances indicate a threatening target. In airport scenarios, birds flying towards the runway trigger a high-priority warning. The bird target recognition and tracking results output a warning signal to activate bird deterrence equipment or notify the control tower to delay takeoffs and landings. Bird target recognition and tracking results are output in real time, with the output frequency matching the video frame rate. Each frame contains complete information on all currently detected bird targets. A tiered output strategy for bird target recognition and tracking results ensures that high-quality recognition results are used for critical decision-making, avoiding false alarms caused by low-quality recognition results.

[0066] To implement the visual-radar fusion-based airport bird identification and tracking method corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects. See also Figure 2 , Figure 2 This paper presents a structural block diagram of an airport bird identification and tracking system 200 based on visual radar fusion, according to an embodiment of this application, including: Data acquisition module 201 is used to acquire 3D visual image data and millimeter-wave radar echo data of the airport airspace, and to perform background subtraction on the 3D visual image data to determine the outline of candidate targets. Feature extraction module 202 is used to perform connected component segmentation on the candidate target contour to determine independent target regions, perform morphological analysis on the independent target regions to obtain morphological asymmetry parameters, and extract visual features from the independent target regions to construct visual descriptors; The target recognition module 203 is used to perform feature fusion of the morphological asymmetry parameter and the visual descriptor to determine a preliminary recognition result, use the millimeter-wave radar echo data to perform distance and velocity verification on the preliminary recognition result to generate fusion error parameters, and perform credibility evaluation based on the fusion error parameters and the morphological asymmetry parameter to determine the bird recognition target; The trajectory generation module 204 is used to perform inter-frame image matching to generate a target trajectory sequence for the bird identification target, perform continuity verification on the target trajectory sequence to obtain trajectory fluctuation parameters and extract trajectory periodic features, and determine dynamic feature data based on the trajectory fluctuation parameters and the trajectory periodic features. The result output module 205 is used to combine the dynamic feature data with the bird identification target to generate a multi-frame identification result sequence, perform a consistency check on the multi-frame identification result sequence to obtain an identification inconsistency parameter, and output the bird target identification and tracking results in a graded manner according to the identification inconsistency parameter.

[0067] The aforementioned airport bird identification and tracking system 200 based on visual radar fusion can implement the airport bird identification and tracking method based on visual radar fusion described in the above-described method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0068] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A method for airport bird identification and tracking based on visual radar fusion, characterized in that, include: Acquire 3D visual image data and millimeter-wave radar echo data of the airport airspace, and perform background subtraction on the 3D visual image data to determine the outline of candidate targets; Connected component segmentation is performed on the candidate target contour to determine independent target regions. Morphological analysis is performed on the independent target regions to obtain morphological asymmetry parameters. Visual features are extracted from the independent target regions to construct visual descriptors. The morphological asymmetry parameter and the visual descriptor are fused to determine the preliminary recognition result. The millimeter-wave radar echo data is used to perform distance and velocity verification on the preliminary recognition result to generate fusion error parameters. Based on the fusion error parameters and the morphological asymmetry parameter, a credibility assessment is performed to determine the bird recognition target. Inter-frame image matching is performed on the bird identification target to generate a target trajectory sequence. The continuity of the target trajectory sequence is verified to obtain trajectory fluctuation parameters and extract trajectory periodic features. Dynamic feature data is determined based on the trajectory fluctuation parameters and the trajectory periodic features. The dynamic feature data and the bird identification target are combined to generate a multi-frame identification result sequence. A consistency check is performed on the multi-frame identification result sequence to obtain the identification inconsistency parameter. Based on the identification inconsistency parameter, the bird target identification and tracking results are output in a graded manner.

2. The method according to claim 1, characterized in that, The step of determining candidate target contours by background subtraction of the 3D visual image data includes: Depth information is extracted from the 3D visual image data to generate a depth map; Based on the depth map, a depth hierarchy is formed to create a near-field layer and a far-field layer; Differential detection is performed in the near-field layer and the far-field layer to generate layered differential results; Based on the layered difference results, deep boundary fusion is performed to determine the contours of candidate targets.

3. The method according to claim 1, characterized in that, The step of performing morphological analysis on the independent target region to obtain morphological asymmetry parameters includes: Extract the target contour boundary from the independent target region to generate a boundary point set; The target principal axis is determined by fitting the principal axis direction to the set of boundary points; The lateral asymmetry is determined by measuring the area ratio of the left and right sides along the principal axis of the target. The morphological asymmetry parameters are determined based on the lateral asymmetry and the aspect ratio characteristics of the target principal axis.

4. The method according to claim 1, characterized in that, The step of extracting visual features from the independent target region to construct a visual descriptor includes: Extract continuous frame area changes from the independent target region to generate an area change sequence; Periodic analysis is performed on the area change sequence to form the pulsation periodic characteristics; The wing flapping frequency is generated by frequency quantization based on the flapping period characteristics. A visual descriptor is constructed based on the wing flapping frequency.

5. The method according to claim 1, characterized in that, The step of using the millimeter-wave radar echo data to perform distance and velocity verification on the preliminary identification result to generate fusion error parameters includes: Micro-Doppler features are generated by performing time-frequency analysis on the millimeter-wave radar echo data; The micro-Doppler features are used to perform wing flapping pattern matching to form flapping matching degree; Based on the beat matching degree and the preliminary identification result, a consistency check is performed to generate a check deviation. The fusion error parameters are determined based on the aforementioned verification deviation.

6. The method according to claim 1, characterized in that, The process of determining bird identification targets based on the credibility assessment using the fusion error parameter and the morphological asymmetry parameter includes: Based on the fusion error parameter and the morphological asymmetry parameter, bird and non-bird hypotheses are established to generate multiple hypothesis sets; The various hypothesis sets are accumulated using multiple frames of evidence to form a hypothesis support sequence; The competitive elimination hypothesis is implemented in the hypothesis support sequence; Based on the aforementioned advantage assumption, a category determination is made to identify bird targets.

7. The method according to claim 1, characterized in that, The step of generating a target trajectory sequence by performing inter-frame image matching on the bird identification target includes: Based on the bird identification target, the target position and appearance features of the current frame are extracted to generate a target state vector; Based on the target state vector, feature similarity matching is performed in adjacent frames to form inter-frame association; The inter-frame association is subjected to occlusion detection to identify trajectory interruption points and generate occlusion markers; Based on the occlusion markers, trajectory interpolation and identity preservation are performed to generate a target trajectory sequence.

8. The method according to claim 3, characterized in that, The step of fitting the principal axis direction of the boundary point set to determine the target principal axis includes: Extract the centroid coordinates from the boundary point set to generate the target center point; Construct a point distribution matrix based on the target center point and the boundary point set; The principal eigenvectors are obtained by performing eigenvalue decomposition on the point distribution matrix. The target principal axis is determined based on the principal feature vector.

9. The method according to claim 5, characterized in that, The step of performing wing flapping pattern matching on the micro-Doppler features to form a flapping matching degree includes: Based on the micro-Doppler features, modulation depth and modulation period are extracted to form beat dynamic parameters; The body size range is constrained by the flapping dynamics parameters to generate body size confidence intervals; Within the body size confidence interval, a candidate flapping pattern set is obtained by searching the bird species flapping template library. The similarity between the micro-Doppler features and the candidate beat pattern set is calculated to form the beat matching degree.

10. An airport bird identification and tracking system based on visual radar fusion, characterized in that, include: The data acquisition module is used to acquire 3D visual image data and millimeter-wave radar echo data of the airport airspace, and to perform background subtraction on the 3D visual image data to determine the outline of candidate targets. The feature extraction module is used to perform connected component segmentation on the candidate target contour to determine independent target regions, perform morphological analysis on the independent target regions to obtain morphological asymmetry parameters, and extract visual features from the independent target regions to construct visual descriptors. The target recognition module is used to perform feature fusion of the morphological asymmetry parameter and the visual descriptor to determine the preliminary recognition result, use the millimeter-wave radar echo data to perform distance and velocity verification on the preliminary recognition result to generate fusion error parameters, and perform credibility assessment based on the fusion error parameters and the morphological asymmetry parameter to determine the bird recognition target; The trajectory generation module is used to perform inter-frame image matching to generate a target trajectory sequence for the bird identification target, perform continuity verification on the target trajectory sequence to obtain trajectory fluctuation parameters and extract trajectory periodic features, and determine dynamic feature data based on the trajectory fluctuation parameters and the trajectory periodic features. The result output module is used to combine the dynamic feature data with the bird identification target to generate a multi-frame identification result sequence, perform a consistency check on the multi-frame identification result sequence to obtain an identification inconsistency parameter, and output the bird target identification and tracking results in a graded manner based on the identification inconsistency parameter.

Images