A target paper recognition method based on point cloud data

Abstract

The present invention discloses a target paper recognition method based on point cloud data, which belongs to the field of target paper recognition methods, comprising the following steps: step 1, collecting original point cloud data of the target range environment by laser radar; step 2, anti-rain and fog interference processing; step 3, high wind sway compensation; step 4, basic surface reconstruction, performing random sampling consistency plane fitting; step 5, wrinkle area enhanced reconstruction, specifically: detecting the curvature change rate on the basic surface is greater than 0.05mm ‑2 Mark the area to be enhanced and pass it to the sub-surface processing flow; Step 6, bullet hole recognition, specifically: calculate the local curvature of the point cloud, and filter the curvature value greater than 0.25mm ‑1 The invention can reduce the interference of rain, fog and strong wind, and improve the recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the field of target paper recognition methods, and in particular to a target paper recognition method based on point cloud data. Background Art

[0002] Target paper is a target board used in shooting training or competitions, typically featuring a specific pattern (such as multiple concentric rings) and a scoring area. Target paper recognition typically involves automatically detecting the impact point on the target and calculating the score through image processing or computer vision techniques. This is commonly used in shooting training and sports competitions. Its technical implementation may involve algorithms such as image binarization, edge detection, color clustering, and target location. The core goal is to achieve fast and accurate score determination.

[0003] The existing patent 201811414037.3 is a real-time target paper positioning and ring line recognition method, including a real-time target paper positioning interception and target surface ring line detection and recognition method, which can obtain target surface information in real time, accurately locate the target surface position under various backgrounds, intercept target surface information, have a fast calculation speed, and can accurately identify the target surface ring line, providing a prerequisite for automatic target reporting, and setting a re-inspection step to prevent calculation errors, further improving the accuracy of target surface ring line recognition.

[0004] The existing technology has the following problems: the recognition process is easily interfered by rain, fog and strong wind, which leads to insufficient recognition accuracy. Therefore, those skilled in the art provide a target paper recognition method based on point cloud data to solve the problems raised in the above background technology. Summary of the Invention

[0005] The purpose of the present invention is to provide a target paper recognition method based on point cloud data, which can reduce the interference of rain, fog and strong wind, improve the recognition accuracy, and solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A target paper recognition method based on point cloud data comprises the following steps:

[0008] Step 1: Collect the original point cloud data of the shooting range environment through LiDAR;

[0009] Step 2: Anti-rain and fog interference processing, specifically:

[0010] Based on the meteorological sensor, the real-time rainfall intensity ρ (unit: mm / h) is obtained and the dynamic intensity filter threshold is calculated:

[0011] ;

[0012] Among them, α, β, γ are adjustable parameters, α∈[0.8,1.2], β∈[-0.15,-0.05], γ∈[15,25];

[0013] Eliminate reflections with intensity lower than T intensity Point cloud;

[0014] Step 3: Compensation for strong wind sway, specifically:

[0015] S301: Obtain the angular velocity ω and acceleration a of the target paper support through an inertial measurement unit;

[0016] S302: Construct a motion compensation model to calculate the 3D compensation vector:

[0017] ;

[0018] Where dt refers to the time differential variable, which is discretized into the scanning time interval Δt in actual processing; Δt refers to the time difference between adjacent laser radar frame scans (unit: seconds); The linear velocity vector derived from the angular velocity ω by the Rodriguez formula;

[0019] S303: Applying reverse compensation to the point cloud;

[0020] Step 4: Reconstruct the basic surface and perform random sampling consistency plane fitting, specifically: (1) randomly select three points to generate candidate planes; (2) calculate the distance from all points to the plane; (3) filter out inliers with a distance less than 1 mm; (4) verify that the angle difference between the inlier normal vectors is less than 5°; (5) iteratively select the plane with the most inliers as the fitting result;

[0021] Step 5: Enhanced reconstruction of wrinkle areas, specifically: Detecting the curvature change rate on the base surface greater than 0.05mm -2 Mark the area as the area to be enhanced and pass it to the sub-surface processing flow;

[0022] Step 6: Bullet hole identification: Calculate the local curvature of the point cloud and select the curvature value greater than 0.25mm -1 The mutation point of the bullet hole is fitted using the least square method;

[0023] Step 7: Output the target paper attitude parameters and bullet hole ring value.

[0024] As a further solution of the present invention: in step 2, when the rainfall intensity ρ>10mm / h, multi-frame point cloud fusion is performed, specifically:

[0025] Get 5 consecutive frames of point cloud data;

[0026] Calculate the fusion weight based on the intensity variance of each frame point cloud;

[0027] Generate fused point cloud by weighted averaging.

[0028] As a further solution of the present invention: in step three, when the wind speed is greater than 6m / s, the laser radar scanning frequency is increased to 30Hz; and a Kalman filter is used to smooth the inertial measurement unit data.

[0029] As a further solution of the present invention: in step 5, the sub-surface processing process includes:

[0030] Region segmentation: The area to be enhanced is segmented based on the criterion that the angle difference between the normal vectors is less than 15°;

[0031] Piecewise fitting: bilinear surface equation is used to fit each sub-region independently;

[0032] Surface splicing: Connect adjacent subsurfaces through non-uniform rational B-splines and ensure the continuity of first-order derivatives.

[0033] As a further solution of the present invention, the bullet hole identification in step 6 includes:

[0034] The curvature mutation points were input into a pre-trained 3D convolutional neural network for verification: the network input was a 16×16×16 voxel grid; the feature extractor consisted of three layers of 3×3×3 convolutional kernels;

[0035] When the neural network output probability value is greater than 0.9, the curvature detection result is covered.

[0036] As a further solution of the present invention: after step 4, the following is further included:

[0037] Establish a local coordinate system with the normal of the fitting plane as the Z axis;

[0038] Detect the four boundary lines of the target paper through Hough transform;

[0039] The arithmetic mean of the coordinates of the boundary line intersection points is taken to set the origin of the coordinate system.

[0040] As a further solution of the present invention, point cloud encryption is performed on the wrinkle area in step 5, specifically:

[0041] When the curvature change rate is greater than 0.05mm -2 The area generates a spherical domain with a radius of 5mm;

[0042] Insert new points in the sphere to make the density reach 3 times the baseline value;

[0043] The surface is reconstructed using the moving least squares method with Gaussian kernel function.

[0044] As a further solution of the present invention: a verification mechanism is also included:

[0045] When the bullet hole fitting residual is greater than 0.3 mm, the visible light camera is started to collect images;

[0046] Coordinate matching is achieved by minimizing the projection error from point cloud to image.

[0047] As a further solution of the present invention: the verification mechanism further includes the following collaborative verification process:

[0048] (a) When the bullet hole fitting residual error is greater than 0.3 mm, a visible light camera deployed coaxially with the lidar is synchronously triggered to capture a local image of the target area. The motorized gimbal is then controlled to adjust the camera to two different observation positions with angles ranging from 15° to 45° relative to the normal of the target plane, acquiring two sets of high-resolution images.

[0049] (b) Preprocessing the captured image, including paper edge detection based on adaptive threshold, Gaussian pyramid noise reduction, and lens distortion correction, to extract the contour feature point set of the suspected bullet hole area in the image;

[0050] (c) Projecting the candidate bullet hole coordinates in the point cloud data into the image coordinate system, and calculating the Euclidean distance deviation between the projected feature points and the actual image contour feature points by establishing a feature matching cost function between the 3D point cloud and the 2D image;

[0051] (d) The Levenberg-Marquardt nonlinear optimization algorithm is used to iteratively adjust the three-dimensional coordinates of the bullet holes until the root mean square error of all projection deviation values ​​is less than 0.15 pixels, and the optimized coordinates are used to overwrite the original curvature detection results;

[0052] (e) If the projection deviation is still greater than the threshold after three iterations, the point is determined to be a false bullet hole and an alarm is triggered. At the same time, the spatial location and environmental parameters of the abnormal point are recorded for manual review.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. This application has all-weather anti-interference capabilities: it innovatively combines meteorological sensors to dynamically calculate the intensity filter threshold (the formula involves rainfall intensity ρ and adjustable parameters α, β, γ), and activates a multi-frame weighted fusion mechanism during heavy rainfall (ρ>10mm / h), effectively eliminating point cloud scattering noise caused by rain and fog; at the same time, it uses an inertial measurement unit to construct a three-dimensional compensation vector (based on the Rodriguez motion model of angular velocity ω and acceleration a), and cooperates with wind speed-triggered scanning frequency adaptive increase (increased to 30Hz when >6m / s) and Kalman filtering to effectively solve the problem of target paper shaking in strong winds, thereby improving the final recognition accuracy.

[0055] 2. This application adopts a layered surface processing strategy, firstly realizing basic reconstruction through RANSAC plane fitting (inside point screening threshold <1mm), and then focusing on the wrinkle area (curvature change rate >0.05mm). -2 ) implements regional segmentation and hyperbolic surface patch fitting, supplemented by NURBS surface splicing to ensure continuity, and finally through point cloud encryption (local density increased to 3 times the baseline value) and moving least squares reconstruction, the reconstruction error of the complex deformed target surface is controlled within 0.3mm.

[0056] 3. This application has passed the curvature mutation test (> 0.25mm -1 ) and the 3D convolutional neural network collaborative mechanism, using a 16×16×16 voxel grid and a three-layer convolution kernel structure to verify bullet hole features (confidence threshold > 0.9), significantly reducing the false alarm rate; supplemented by residual-driven verification (visible light projection matching is activated when > 0.3mm), ensuring that the three-dimensional coordinate positioning accuracy of the bullet hole meets the ring value judgment requirements.

[0057] 4. This application achieves stable output of target paper attitude parameters through the establishment of a local coordinate system (the plane normal is the Z axis), Hough transform boundary detection and intersection average positioning, providing a spatial reference for automatic target reporting. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 The figure is a flow chart of a target paper recognition method based on point cloud data. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0060] As mentioned in the background technology of this application, research has found that the existing recognition process is easily interfered by rain, fog and strong winds, which leads to insufficient recognition accuracy and certain defects.

[0061] In order to solve the above-mentioned defects, the present application discloses a target paper recognition method based on point cloud data, which can reduce the interference of rain, fog and strong wind and improve the recognition accuracy.

[0062] The following will describe in detail how the solution of this application solves the above technical problems with reference to the accompanying drawings.

[0063] See also Figure 1 In an embodiment of the present invention, a target paper recognition method based on point cloud data includes the following steps:

[0064] Step 1: Collect the original point cloud data of the shooting range environment through LiDAR;

[0065] Step 2: Anti-rain and fog interference processing, specifically:

[0066] Based on the meteorological sensor, the real-time rainfall intensity ρ (unit: mm / h) is obtained and the dynamic intensity filter threshold is calculated:

[0067] ;

[0068] Among them, α, β, γ are adjustable parameters, α∈[0.8,1.2], β∈[-0.15,-0.05], γ∈[15,25];

[0069] Eliminate reflections with intensity lower than T intensity Point cloud;

[0070] Step 3: Compensation for strong wind sway, specifically:

[0071] S301: Obtain the angular velocity ω and acceleration a of the target paper support through an inertial measurement unit;

[0072] S302: Construct a motion compensation model to calculate the 3D compensation vector:

[0073] ;

[0074] Where dt refers to the time differential variable, which is discretized into the scanning time interval Δt in actual processing; Δt refers to the time difference between adjacent laser radar frame scans (unit: seconds); The linear velocity vector derived from the angular velocity ω by the Rodriguez formula;

[0075] S303: Applying reverse compensation to the point cloud;

[0076] Step 4: Reconstruct the basic surface and perform random sampling consistency plane fitting, specifically: (1) randomly select three points to generate candidate planes; (2) calculate the distance from all points to the plane; (3) filter out inliers with a distance less than 1 mm; (4) verify that the angle difference between the inlier normal vectors is less than 5°; (5) iteratively select the plane with the most inliers as the fitting result;

[0077] Step 5: Enhanced reconstruction of wrinkle areas, specifically: Detecting the curvature change rate on the base surface greater than 0.05mm -2 Mark the area as the area to be enhanced and pass it to the sub-surface processing flow;

[0078] Step 6: Bullet hole identification: Calculate the local curvature of the point cloud and select the curvature value greater than 0.25mm -1 The mutation point of the bullet hole is fitted using the least square method;

[0079] Step 7: Output the target paper attitude parameters and bullet hole ring value.

[0080] This application can achieve all-weather target paper recognition: data quality is guaranteed through dynamic anti-interference (rain and fog) and motion compensation (strong wind); layered reconstruction strategy (base plane + fold enhancement) improves recognition accuracy under complex deformation; curvature mutation detection accurately locates bullet holes.

[0081] In this embodiment, in step 2, when rainfall intensity ρ > 10 mm / h, multi-frame point cloud fusion is performed. Specifically, the following steps are performed: 5 consecutive frames of point cloud data are collected; fusion weights are calculated based on the variance of the point cloud intensity for each frame; and a fused point cloud is generated by weighted averaging the weights. This configuration can address the point cloud sparseness caused by heavy rain. Multi-frame fusion enhances the effective signal, suppresses random raindrop noise, and improves point cloud integrity.

[0082] In this embodiment, in step 3, when the wind speed exceeds 6 m / s, the lidar scanning frequency is increased to 30 Hz, and a Kalman filter is used to smooth the IMU data. This setting can cope with the high-frequency oscillation of the target paper caused by strong winds. The high scanning frequency captures instantaneous displacement, and the Kalman filter smoothes the IMU data, improving motion compensation accuracy.

[0083] In this embodiment, step 5 of the subsurface processing process includes: region segmentation: using the criterion that the angle difference between normal vectors is less than 15° to segment the region to be enhanced; piecewise fitting: using bilinear surface equations to independently fit each subregion; and surface splicing: connecting adjacent subsurfaces using non-uniform rational B-splines while ensuring first-order derivative continuity. This setup can address the problem of surface breaks caused by wrinkles in the target paper. Piecewise fitting preserves local features, while NURBS splicing ensures global smoothness, improving the quality of bullet hole edge reconstruction.

[0084] In this embodiment, bullet hole recognition in step six involves inputting curvature mutation points into a pre-trained 3D convolutional neural network for verification. The network input is a 16×16×16 voxel grid; a feature extractor is implemented using three layers of 3×3×3 convolutional kernels; and curvature detection results are overwritten when the neural network output probability value is greater than 0.9. This setup reduces the false detection rate of foreign objects (such as stains), utilizes the CNN to learn the 3D morphological features of bullet holes, verifies the curvature screening results, and improves recognition robustness.

[0085] In this embodiment, step 4 is followed by establishing a local coordinate system with the normal of the fitted plane as the Z axis; detecting the four boundary lines of the target paper using a Hough transform; and taking the arithmetic mean of the coordinates of the boundary line intersections to set the coordinate system origin. This setup provides a reference framework for calculating the ring value of bullet holes, accurately aligning the target paper plane with the boundary, and ensuring that the ring value determination is consistent with the actual position.

[0086] In this embodiment, point cloud encryption is performed on the wrinkle area in step 5, specifically: when the curvature change rate is greater than 0.05mm, -2 A 5mm radius sphere is generated for the area. New points are inserted within the sphere to triple the baseline density. The surface is reconstructed using a moving least squares method with a Gaussian kernel function. This setting addresses fitting distortion caused by sparse point clouds in wrinkled areas, improves surface reconstruction resolution through local densification, and preserves geometric features through Gaussian kernel weighting.

[0087] This embodiment also includes a verification mechanism: when the bullet hole fitting residual exceeds 0.3mm, the visible light camera is activated to capture images; coordinate matching is achieved by minimizing the projection error from the point cloud to the image. This setting verifies the bullet hole position across modalities, correcting point cloud fitting deviations using visual information, and ensuring output reliability under extreme deformations.

[0088] In this embodiment, the verification mechanism further includes the following collaborative verification process: (a) when the bullet hole fitting residual is greater than 0.3 mm, the visible light camera deployed at the coaxial position of the laser radar is synchronously triggered to collect a local image of the target paper area, and the electric pan-tilt platform is controlled to adjust the camera to two different observation orientations with an angle of 15° to 45° with the normal of the target plane, and two sets of high-resolution images are obtained respectively; (b) preprocessing operations are performed on the collected images, including paper edge detection based on adaptive threshold, Gaussian pyramid noise reduction and lens distortion correction, to extract the contour feature point set of the suspected bullet hole area in the image; (c) the point cloud data is converted into the local image of the target paper area; The candidate bullet hole coordinates are projected into the image coordinate system. A cost function is established for feature matching between the 3D point cloud and the 2D image, and the Euclidean distance deviation between the projected feature points and the actual image contour feature points is calculated. (d) The 3D coordinates of the bullet holes are iteratively adjusted using the Levenberg-Marquardt nonlinear optimization algorithm until the root mean square error of all projection deviations is less than 0.15 pixels. The optimized coordinates are then overwritten with the original curvature detection results. (e) If the projection deviation is still greater than the threshold after three iterations, the point is determined to be a false bullet hole and an alarm is triggered. The spatial location and environmental parameters of the outlier are also recorded for manual review. This setup builds a multimodal collaborative verification system based on the existing verification mechanism, effectively addressing the technical challenge of misidentifying bullet holes in complex environments.

[0089] This application has all-weather anti-interference capabilities: it innovatively combines meteorological sensors to dynamically calculate the intensity filter threshold (the formula involves rainfall intensity ρ and adjustable parameters α, β, γ), and starts a multi-frame weighted fusion mechanism during heavy rainfall (ρ>10mm / h), effectively eliminating the point cloud scattering noise caused by rain and fog; at the same time, it uses the inertial measurement unit to construct a three-dimensional compensation vector (based on the Rodriguez motion model of angular velocity ω and acceleration a), and cooperates with the wind speed-triggered scanning frequency adaptive improvement (increased to 30Hz when >6m / s) and Kalman filtering to effectively solve the problem of target paper shaking in strong wind environments, thereby improving the final recognition accuracy. At the same time, this application adopts a layered surface processing strategy, first realizing basic reconstruction through RANSAC plane fitting (inner point screening threshold <1mm), and then targeting the wrinkle area (curvature change rate >0.05mm -2 ) implements regional segmentation and hyperbolic surface patch fitting, supplemented by NURBS surface splicing to ensure continuity, and finally through point cloud encryption (local density increased to 3 times the baseline value) and moving least squares reconstruction, the reconstruction error of the complex deformation target surface is controlled within 0.3mm. In addition, this application uses curvature mutation detection (>0.25mm -1 ) and a 3D convolutional neural network collaborate to verify bullet hole characteristics (confidence threshold > 0.9) using a 16×16×16 voxel grid and a three-layer convolutional kernel structure, significantly reducing false alarm rates. This is supplemented by residual-driven verification (visible light projection matching is activated when > 0.3mm) to ensure that the three-dimensional coordinate positioning accuracy of bullet holes meets the ring value determination requirements. Finally, this application achieves stable output of target paper posture parameters through the establishment of a local coordinate system (with the plane normal as the Z axis), Hough transform boundary detection, and intersection average positioning, providing a spatial reference for automatic target reporting.

[0090] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

[0091] The above is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A target paper recognition method based on point cloud data, characterized in that: The following steps are involved: Step 1: Collect the original point cloud data of the shooting range environment through LiDAR; Step 2: Anti-rain and fog interference processing, specifically: Based on the real-time rainfall intensity ρ obtained by meteorological sensors, the dynamic intensity filtering threshold is calculated: ； Among them, α, β, γ are adjustable parameters, α∈[0.8,1.2], β∈[-0.15,-0.05], γ∈[15,25]; Eliminate reflections with intensity lower than T intensity Point cloud; Step 3: Compensation for strong wind sway, specifically: S301: Obtain the angular velocity ω and acceleration a of the target paper support through an inertial measurement unit; S302: Construct a motion compensation model to calculate the 3D compensation vector: ； Where dt refers to the time differential variable; Δt refers to the time difference between adjacent frames scanned by the lidar; The linear velocity vector derived from the angular velocity ω by the Rodriguez formula; S303: Applying reverse compensation to the point cloud; Step 4: Reconstruct the basic surface and perform random sampling consistent plane fitting; Step 5: Enhanced reconstruction of wrinkle areas, specifically: Detecting the curvature change rate on the base surface greater than 0.05mm -2 Mark the area as the area to be enhanced and pass it to the sub-surface processing flow; Step 6: Bullet hole identification: Calculate the local curvature of the point cloud and select the curvature value greater than 0.25mm -1 The mutation point of the bullet hole is fitted using the least square method; Step 7: Output the target paper attitude parameters and bullet hole ring value.

2. The target paper recognition method based on point cloud data according to claim 1, characterized in that: In step 2, when the rainfall intensity ρ>10 mm / h, multi-frame point cloud fusion is performed, specifically: Get 5 consecutive frames of point cloud data; Calculate the fusion weight based on the intensity variance of each frame point cloud; Generate fused point cloud by weighted averaging.

3. The target paper recognition method based on point cloud data according to claim 2, characterized in that: In step three, when the wind speed is greater than 6 m / s, the laser radar scanning frequency is increased to 30 Hz; and a Kalman filter is used to smooth the inertial measurement unit data.

4. The target paper recognition method based on point cloud data according to claim 3, characterized in that: In step 4, the specific process of performing random sampling consistency plane fitting is: (1) Randomly select three points to generate a candidate plane; (2) Calculate the distance from all points to the plane; (3) Screening of inliers with a distance less than 1 mm; (4) Verify that the angle difference between the normal vectors of the interior points is less than 5°; (5) Iteratively select the plane with the most internal points as the fitting result.

5. The target paper recognition method based on point cloud data according to claim 4, characterized in that: In step 5, the sub-surface processing flow includes: Region segmentation: The area to be enhanced is segmented based on the criterion that the angle difference between the normal vectors is less than 15°; Piecewise fitting: bilinear surface equation is used to fit each sub-region independently; Surface splicing: Connect adjacent subsurfaces through non-uniform rational B-splines and ensure the continuity of first-order derivatives.

6. The target paper recognition method based on point cloud data according to claim 5, characterized in that: The bullet hole identification in step 6 includes: The curvature mutation points were input into a pre-trained 3D convolutional neural network for verification: the network input was a 16×16×16 voxel grid; the feature extractor consisted of three layers of 3×3×3 convolutional kernels; When the neural network output probability value is greater than 0.9, the curvature detection result is covered.

7. The target paper recognition method based on point cloud data according to claim 6, characterized in that: The step 4 further includes: Establish a local coordinate system with the normal of the fitting plane as the Z axis; Detect the four boundary lines of the target paper through Hough transform; The arithmetic mean of the coordinates of the boundary line intersection points is taken to set the origin of the coordinate system.

8. The target paper recognition method based on point cloud data according to claim 7, characterized in that: Perform point cloud encryption on the wrinkle area in step 5, specifically: When the curvature change rate is greater than 0.05mm -2 The area generates a spherical domain with a radius of 5mm; Insert new points in the sphere to make the density reach 3 times the baseline value; The surface is reconstructed using the moving least squares method with Gaussian kernel function.

9. The target paper recognition method based on point cloud data according to claim 8, characterized in that: Also includes a validation mechanism: When the bullet hole fitting residual is greater than 0.3 mm, the visible light camera is started to collect images; Coordinate matching is achieved by minimizing the projection error from point cloud to image.

10. The target paper recognition method based on point cloud data according to claim 9, characterized in that: The verification mechanism further includes the following collaborative verification process: (a) When the bullet hole fitting residual error is greater than 0.3 mm, a visible light camera deployed coaxially with the lidar is synchronously triggered to capture a local image of the target area. The motorized gimbal is then controlled to adjust the camera to two different observation positions with angles ranging from 15° to 45° relative to the normal of the target plane, acquiring two sets of high-resolution images. (b) Preprocessing the captured image, including paper edge detection based on adaptive threshold, Gaussian pyramid noise reduction, and lens distortion correction, to extract the contour feature point set of the suspected bullet hole area in the image; (c) Projecting the candidate bullet hole coordinates in the point cloud data into the image coordinate system, and calculating the Euclidean distance deviation between the projected feature points and the actual image contour feature points by establishing a feature matching cost function between the 3D point cloud and the 2D image; (d) The Levenberg-Marquardt nonlinear optimization algorithm is used to iteratively adjust the three-dimensional coordinates of the bullet holes until the root mean square error of all projection deviation values ​​is less than 0.15 pixels, and the optimized coordinates are used to overwrite the original curvature detection results; (e) If the projection deviation is still greater than the threshold after three iterations, the point is determined to be a false bullet hole and an alarm is triggered. At the same time, the spatial location and environmental parameters of the abnormal point are recorded for manual review.

Images