A large-span passage construction safety early warning method based on key point linkage matching

Abstract

The application discloses a kind of large-span passage construction safety early warning method based on key point linkage matching, it is related to electric power engineering construction safety monitoring technical field, first, the three-dimensional point cloud data and image of large-span passage are collected, the lead point cloud is segmented by RANSAC algorithm, SIFT algorithm is used to extract image feature points, 2D-3D point pair set is obtained by projection association and RANSAC false matching elimination;Second, the initial pose is solved by UPnP algorithm, and the optimal pose is obtained by combining GICP algorithm to complete secondary fine calibration;Improved U-Net model is constructed to segment construction vehicle target, after extracting vehicle contour, profile key point is detected and converted into three-dimensional coordinates, optimal key point is selected by preliminary distance screening and variance reverse verification;The shortest Euclidean distance of the optimal key point and the lead point cloud is calculated using three-dimensional traversal algorithm, finally, the distance is compared with the preset alarm, early warning threshold, to realize the accurate judgment and response of safety, early warning, alarm state.

Description

Technical Field

[0001] This invention relates to the field of power engineering construction safety monitoring technology, and in particular to a method for early warning of construction safety in large-span tunnels based on key point linkage matching. Background Technology

[0002] Long-span crossings are critical nodes in high-voltage transmission networks. Construction sites present a dynamic, three-dimensional working environment, where unauthorized intrusion by construction vehicles can easily lead to safety accidents such as conductor scraping and electric shock. Traditional management, primarily manual, suffers from inaccurate positioning of personnel and machinery, incomplete physical protection, and reliance on experience for electric shock identification, which can no longer meet the safety management requirements of modern power construction.

[0003] With the rapid development of LiDAR and visual positioning technologies, scene registration and reconstruction of real-world scene information using visualization devices have become possible. The core of multi-dimensional fusion ranging stems from the fundamental theories and engineering applications of computer vision; essentially, it involves inferring three-dimensional spatial coordinates from two-dimensional image information. The key lies in establishing a stable mapping relationship between spatial points and pixels. Currently, technologies such as deep learning and point cloud processing have yielded numerous practical applications in computer vision and virtual reality, providing theoretical and technical support for the research of safety early warning technologies for long-span passages.

[0004] Currently, the power industry is constantly developing technologies related to safety early warning for transmission lines. However, there are still shortcomings in dedicated early warning technologies for dynamic three-dimensional operation scenarios in long-span channels. Existing technologies are mostly applicable to static scenarios or conventional transmission lines, and generally suffer from insufficient accuracy and adaptability, making it difficult to achieve accurate and reliable safety early warning. There is an urgent need for a targeted, accurate, and adaptable intelligent early warning method.

[0005] Existing technologies only locate construction vehicles by the vertices of a rectangular frame, without combining the contour features of the construction vehicles to extract the optimal key points related to the distance to the guide rail. This makes it impossible to accurately reflect the actual threat posed by the vehicles to the guide rail, and it is easy to cause false alarms or missed alarms due to positioning deviations. At the same time, the mapping relationship between pixels and point clouds is only established through imaging models and stereo matching algorithms, without setting a secondary pose alignment verification mechanism. The initial pose calculation process is easily affected by scene interference, resulting in deviations in the matching between point clouds and pixels, which in turn affects the ranging accuracy and cannot meet the high-precision monitoring requirements of long-span channels. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a construction safety early warning method for long-span tunnels based on key point linkage matching, comprising: Collect single-time 3D point cloud data and images of the long-span passage, automatically segment the 3D point cloud data into a guide wire to obtain a guide wire point cloud set, extract image feature points, and form an image feature point set. Project the wire point cloud set onto the image coordinate system, associate and match it with the image feature point set, eliminate the mis-matched points, and obtain the 2D-3D point pair set; Solve the initial pose of the visualization device based on the 2D-3D point pair set, and perform secondary fine calibration on the initial pose to obtain the optimal pose; Construct an object segmentation model based on the improved U-Net, perform construction vehicle object segmentation on the image, extract the vehicle contour, convert the key points of the vehicle contour into three-dimensional coordinates, and select the optimal key points through preliminary distance screening and variance reverse verification; Traverse each point in the wire point cloud set, calculate the Euclidean distance from it to the optimal key point, and take the minimum value as the shortest distance d between the construction vehicle and the wire; Set the alarm threshold D1 and the warning threshold D2; compare the shortest distance d with the alarm threshold D1 and the warning threshold D2: if d < D1, trigger an alarm; if D1 ≤ d < D2, trigger a warning; if d ≥ D2, it is in a safe state.

[0007] Furthermore, fit the spatial line model, calculate the Euclidean distance from the points in the three-dimensional point cloud data to the line, determine the points with a distance less than the set threshold as wire points, and after iteration, retain the point set corresponding to the spatial line model with the largest number of inliers as the wire point cloud set.

[0008] Furthermore, use the SIFT algorithm to extract image feature points, extract key points and 128-dimensional descriptors, and screen the extracted SIFT key points to retain the valid image feature points to form an image feature point set.

[0009] Furthermore, the association and matching of the wire point cloud set projected onto the image coordinate system with the image feature point set specifically includes: project the wire point cloud set to the image pixel coordinates in the image coordinate system based on the lidar extrinsic matrix and the camera intrinsic matrix, calculate the Euclidean distance from each projected point to the SIFT key point, set a matching threshold to screen the initial matching point pairs, then use the RANSAC algorithm to eliminate the mis-matched points, retain the inliers with a projection error less than or equal to 1 pixel, obtain a pure 2D-3D point pair set, and perform Gaussian filtering on the matched wire point cloud set.

[0010] Furthermore, use the UPnP algorithm, based on the camera pinhole imaging model, take the 2D-3D point pair set as the input, construct an optimization objective function, and solve the initial rotation matrix R and translation vector T that minimize the projection error.

[0011] Furthermore, perform secondary fine calibration on the initial pose, use the GICP algorithm, take the initial pose as the iteration initial value, consider the point-to-point, point-to-plane distance and covariance constraints, iterate the GICP optimization objective function, and output the optimally calibrated optimal pose.

[0012] Furthermore, an improved U-Net-based target segmentation model is constructed to segment construction vehicle targets in the image. Specifically, the SE-Net attention mechanism is introduced into the traditional U-Net model, and the cross-entropy loss function and Adam optimizer are used. The image is input into the trained target segmentation model, and the output is a binarized image of the vehicle outline. Then, the complete vehicle outline is extracted by the Canny edge detection algorithm.

[0013] Furthermore, the Harris corner detection algorithm is used to extract key points of the vehicle contour. The pixel coordinates of the key points are converted into three-dimensional coordinates through the camera pinhole imaging model and the optimal pose. The preliminary distance between each key point and the guide reference point is calculated, and multiple candidate key points with the closest distance are selected. The distance variance between each candidate key point and multiple reference points in the guide point cloud set is calculated, and the key point with the smallest distance variance and the shortest distance is selected as the optimal key point.

[0014] Furthermore, each point in the traverse point cloud set is traversed, and the Euclidean distance to the optimal key point is calculated. A three-dimensional traversal algorithm is used to calculate the Euclidean distance between each point in the traverse point cloud set and the optimal key point of the vehicle in turn. The minimum value among all distances is selected as the shortest distance d between the construction vehicle and the traverse.

[0015] Furthermore, the alarm threshold D1 is set based on safety regulations and standards, and the warning threshold D2 is set based on on-site operational needs and environmental complexity, with D2 > D1.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: 1. Based on an improved U-Net target segmentation model with SE-Net attention mechanism, combined with static images captured by visualization devices, vehicle contour key points are extracted. Through preliminary ranging and variance verification, the optimal key points of the vehicle closest to the guide are accurately located without the need for vehicle speed data. 2. The RANSAC algorithm is used to achieve automatic segmentation of traverse point clouds. Combined with UPnP initial pose calculation, GICP secondary pose fine calibration and SIFT feature-assisted matching, the matching accuracy and stability are improved. 3. Using point cloud data and visualized static images collected in a single transaction, a 3D traversal algorithm is employed to accurately calculate the shortest distance between the vehicle's optimal key points and the guide wire, ensuring comprehensive ranging accuracy. 4. For three-dimensional operation scenarios in long-span passages, a targeted early warning and judgment mechanism is constructed by combining national standard safety thresholds. This mechanism integrates point cloud and visualized image features to improve the reliability of early warnings, reduce false alarms and missed alarms, and meet the actual operational needs of front-line workers. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] Figure 1 The flowchart is a safety early warning method for the construction of long-span tunnels according to the present invention. Figure 2 To collect 3D point cloud data of a large-span passage in a single acquisition; Figure 3 To capture two-dimensional still images of the long-span passageway; Figure 4 The vehicle images collected at the construction site of the long-span passage were used as the model training dataset for this invention. Figure 5 The vehicle contour result extracted based on the Canny edge detection algorithm of this invention; Figure 6 The results of vehicle contour recognition and distance measurement in two-dimensional images of a 220kV line with channels 0013#-0014# as the pilot channel; Figure 7 This is the result of simulating vehicle position and measuring distance in point cloud data using rectangular boxes. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] This embodiment proposes a construction safety early warning method for large-span tunnels based on key point linkage matching and dynamic threshold calibration. Based on single-acquisition of 3D point cloud data and image data of the large-span tunnel, it integrates target segmentation, 2D-3D point pair matching, secondary pose verification, and 3D traversal ranging techniques to achieve accurate distance measurement and safety early warning between construction vehicles and the guide wire. Figure 1 As shown, the specific steps are as follows: Step 1: Collect single-shot 3D point cloud data and images of the long-span passage.

[0021] Using drones, three-dimensional point cloud data of a long-span passage can be collected in a single operation, such as... Figure 2 As shown, the point cloud density satisfies ≥100 points / m 2 This ensures that spatial location information of the conductor, surrounding environment, etc., can be clearly captured. The core principle of lidar point cloud acquisition is the time-of-flight (ToF) method, which obtains the three-dimensional spatial coordinates (X, Y, Z) of each point.

[0022] Using visualization equipment (resolution ≥ 1920×1080), capture two-dimensional static images of the long-span passage, such as... Figure 3 As shown, the image needs to clearly display the outline and texture features of the conductor. After acquisition, distortion correction is performed on the image based on the camera intrinsic matrix (obtained in advance through camera calibration) to eliminate errors caused by lens distortion. The corrected image is used for subsequent feature extraction and target segmentation. The expression for the camera intrinsic matrix K is: ; In the formula, , These are the focal lengths (in pixels) of the visualization device (camera) in the x and y directions, respectively. , These are the pixel coordinates of the camera's principal point.

[0023] Step 2: Perform automatic traverse segmentation on the 3D point cloud data to obtain a traverse point cloud set.

[0024] Considering the structural characteristics of the traverse lines, which are spatially linear or approximately cylindrical, the Random Sampling Consensus (RANSAC) algorithm is used to automatically segment the traverse lines from a single acquisition of 3D point cloud data. This algorithm has stronger resistance to background interference and is suitable for complex scenarios involving long-span passages, as detailed below: A spatial straight line model is fitted, and the RANSAC algorithm is used to fit a spatial straight line to segment the traverse point cloud. The Euclidean distance from each point in the 3D point cloud data to the straight line is calculated. Specifically, two sets of non-collinear points are randomly selected from the global point cloud, and a candidate spatial straight line is fitted. Its equation is: , Where (X,Y,Z) are the three-dimensional coordinates of any point in space. A reference point on the straight line coordinates It is the direction vector of the line; Calculate the Euclidean distance from all points to the candidate line. The distance calculation formula is: , In the formula, Let i be the i-th point in the 3D point cloud data. As a reference point on the straight line, The direction vector is a straight line. A distance threshold d = 0.03m is set. Points less than the threshold are denoted as interior points (traverse points), and points greater than the threshold are denoted as exterior points (background, interference points). After multiple iterations (≥1000 iterations), the spatial straight line model with the most interior points is retained, and the corresponding point set is the traverse point cloud W. The segmented traverse point cloud undergoes secondary denoising (removing isolated exterior points), and each traverse point is numbered and labeled, forming a traverse point cloud set W = {W1(X1,Y1,Z1),W2(X2,Y2,Z2),...,W...} m(X m ,Y m Z m (m≥1000), complete the segmentation and labeling of the traverse point cloud.

[0025] Step 3: Extract feature points from the static image to form an image feature point set.

[0026] The Scale Invariant Feature Transform (SIFT) algorithm is used to extract feature points from the visualized static image, making full use of the image's texture features to provide auxiliary support for subsequent 2D-3D point pair matching. The specific process is as follows: The image is converted to grayscale to reduce redundant information and improve feature extraction efficiency. At the same time, Gaussian smoothing is performed to remove image noise and avoid noise interference with feature point extraction. The Gaussian smoothing formula is the same as that used in the subsequent point cloud preprocessing, so it will not be repeated here.

[0027] The SIFT algorithm is used to extract key points and corresponding 128-dimensional descriptors from images. SIFT key points are scale-invariant and rotation-invariant, which can effectively adapt to feature matching requirements under different lighting and slightly occluded scenes. The 128-dimensional descriptors are used to accurately describe the local texture information of each feature point, ensuring the uniqueness of the feature points.

[0028] The extracted SIFT keypoints are filtered, removing feature points with low response values ​​and blurry edges, and retaining effective feature points F with high response values ​​and clear textures. i (i≥50), forming an image feature point set F={F1(u1,v1),F2(u2,v2),...,F... i (u i ,v i )}, where (u i ,v i () represents the pixel coordinates of the feature points, preparing for subsequent 2D-3D point pair matching.

[0029] Step 4: Project the conductor point cloud onto the image coordinate system, perform association matching with image feature points, remove mismatched points, and obtain a 2D-3D point pair set.

[0030] Based on the point cloud set W of the segmented and labeled guide lines and the extracted image feature point set F, we perform accurate 2D-3D point pair matching between the guide line point cloud and image pixels. SIFT feature assistance is introduced to improve the matching stability in complex scenes. The specific process is as follows: Based on the lidar extrinsic matrix [R|T], with rotation matrix R and translation vector T, each point in the traverse point cloud set W is converted into coordinates in the camera coordinate system. Then, it is projected onto the image pixel coordinate system through the camera intrinsic matrix K to obtain the preliminary projected pixel coordinates of the traverse point cloud. It should be noted that the extrinsic matrix adopts the chessboard calibration method. The two-dimensional corner coordinates of the visualized camera image are obtained through corner detection, and the corresponding three-dimensional corner coordinates are obtained through point cloud plane fitting and corner matching. The rotation matrix R and translation vector T are solved through the PnP algorithm and least squares optimization to form the extrinsic matrix [R|T].

[0031] The initial projected pixel coordinates of the conductor point cloud are associated with the image feature point set F. The Euclidean distance between each projected point and the SIFT key point is calculated. A matching threshold is set, preferably 0.5, and point pairs with high matching degree are selected to form an initial 2D-3D matching point pair set.

[0032] The RANSAC algorithm is used to filter the initial 2D-3D matching point pair set, eliminating mismatched points and retaining interior point matching pairs to ensure matching accuracy. Specifically, a subset of matching point pairs is iteratively selected, the projection error is calculated, and an error threshold is set, preferably 1 pixel. Interior points with projection errors less than the threshold are retained. After multiple iterations, the set of matching point pairs with the largest number of interior points is the final clean 2D-3D point pair set. .

[0033] Gaussian filtering is applied to the matched traverse point (X,Y,Z) cloud to remove high-frequency noise, preserve traverse structural features, and ensure the accuracy of subsequent pose calculation and ranging. The Gaussian filtering formula is: In the formula, standard deviation =0.005m, This is a Gaussian filter operator that smooths point cloud data using Gaussian filtering to eliminate noise interference.

[0034] Step 5: Solve the initial pose of the visualization device based on the 2D-3D point pair set.

[0035] The UPnP algorithm, based on a clean 2D-3D point pair set, solves for the initial extrinsic parameters of visualization devices such as cameras, including the rotation matrix R and translation vector T. It is more robust to noise and mismatched points, making it suitable for complex scenarios involving large-span channels. It does not require data assistance based on traverse wire oscillations. The specific process is as follows: Using a camera pinhole imaging model, the 3D coordinates of the duct point cloud are correlated with the image pixel coordinates. A least-squares optimization problem is constructed using the UPnP algorithm to directly solve for the globally optimal initial pose, avoiding the problem of traditional PnP algorithms easily getting trapped in local optima. The formula for the camera pinhole imaging model is: In the formula, Let (X,Y,Z) be the depth value of the target point in the camera coordinate system, (X,Y,Z) be the three-dimensional coordinates of the target point in the world coordinate system, (u,v) be the pixel coordinates of the target point on the visualized image, K be the camera intrinsic matrix, and [R|T] be the camera extrinsic matrix.

[0036] Using the 2D-3D point pair set obtained in step 5 as input, a least-squares optimization objective function is constructed to solve for the initial rotation matrix R and translation vector T that minimize the projection error. The least-squares optimization objective function is:

[0037] In the formula, Let K be the depth value of the j-th traverse point in the camera coordinate system, and K be the camera intrinsic parameter matrix. j ,v j ) represents the image pixel coordinates of the j-th point; n is the number of 2D-3D matching point pairs, preferably n≥6. init It is the initial rotation matrix; T init It is the initial translation vector; (X) wj Y wj Z wj ) represents the three-dimensional coordinates of the j-th point in the world coordinate system, and w represents the world coordinate system.

[0038] After the solution is completed, the initial rotation matrix R and translation vector T are output as the initial values ​​for the subsequent secondary pose verification iteration.

[0039] Step 6: Perform a second fine calibration on the initial pose to obtain the optimal pose.

[0040] To further eliminate initial pose calculation errors and improve pose accuracy, the GICP algorithm is used to perform secondary pose fine calibration on the initial pose (R,T), while considering point-to-point and point-to-surface distances and covariance constraints. The specific process is as follows: The objective function formula for GICP optimization is: In the formula: Let j be the coordinates of the j-th point in the original traverse point cloud. Let J be the coordinates of the j-th point in the point cloud after projection from the initial pose. , R and T are the local neighborhood covariance matrices of the original traverse point cloud and the projected point cloud, respectively, and the rotation matrix and translation vector to be optimized.

[0041] Using the R and T values ​​output by the UPnP algorithm as initial values ​​for iteration, the optimal pose (R,T) that minimizes the GICP objective function is iteratively solved. In each iteration, the rotation matrix R and translation vector T are updated, and the average distance error is calculated. Preferably, the iteration stops when the average distance error is <0.01m or the number of iterations is >50, and the finely calibrated optimal pose (R,T) is output to complete the secondary pose verification, providing an accurate pose basis for subsequent vehicle target segmentation and key point extraction.

[0042] Step 7: Construct a target segmentation model based on the improved U-Net to segment construction vehicle targets in the image and extract vehicle contours.

[0043] An improved U-Net-based target segmentation model is constructed. By combining the acquired images and the obtained optimal pose (R,T), the construction vehicle is accurately segmented, background interference is removed, and the complete outline of the vehicle is obtained. The specific process is as follows: Model building and training: Based on the traditional U-Net model, this model adds an attention mechanism module (SE-Net) to improve the segmentation accuracy of construction vehicle outlines, reduce background interference such as ground, vegetation, and other non-operational equipment, and adapt to static image scenes captured by visualization equipment.

[0044] The model training dataset uses vehicle images collected at the construction site of the long-span tunnel (captured by visualization equipment, such as...). Figure 4 The dataset includes static samples of different types of construction vehicles (cranes, excavators, transport vehicles, etc.), different lighting conditions, and different working angles, with a sample size of ≥1000 images. Data augmentation techniques (symmetry, rotation, brightness adjustment) are used to expand the dataset.

[0045] The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. During training, the cross-entropy loss function was used as the loss function, the Adam optimizer was employed, the learning rate was set to 0.001, and the number of iterations was 100 rounds, until the model's validation set accuracy was ≥95%, at which point training stopped.

[0046] The expression for the cross-entropy loss function L is: ; In the formula, N is the sample size. The true label for the i-th sample is 1 for the vehicle region and 0 for the background region. Let be the predicted probability of the model for the i-th sample.

[0047] The corrected static image is input into the trained target segmentation model, which outputs a binarized image of the vehicle contour. In the binarized image, the vehicle region is marked as 1, and the background region is marked as 0. Then, the edge points of the vehicle contour are extracted using the Canny edge detection algorithm to obtain the complete vehicle contour. Preferably, the threshold parameter of the Canny edge detection algorithm is 0.6. Figure 5 The image shows the vehicle outline extracted using the Canny edge detection algorithm. If the outline is smaller than the threshold parameter, it is directly filtered out and not displayed.

[0048] Step 8: Convert the key points of the vehicle outline into 3D coordinates, and select the optimal key points through preliminary distance screening and variance backtesting.

[0049] The Harris corner detection algorithm is used to extract multiple key points (≥20) of the vehicle contour. Each key point corresponds to image pixel coordinates (u,v). Then, using a camera pinhole imaging model and optimal pose (R,T), the pixel coordinates are converted into corresponding point cloud 3D coordinates (X,Y,Z), forming a set of vehicle key points: P={P1(X1,Y1,Z1),P2(X2,Y2,Z2),...,P n (X n ,Y n Z n )}, preferably n≥20.

[0050] Using the simplified Euclidean distance formula, the formula is as follows: ;in, Let be the initial distance between the i-th vehicle key point and the traverse reference point (the center coordinates of the traverse point cloud). The three-dimensional coordinates of the guide reference point, Given the three-dimensional coordinates of the i-th vehicle key point, calculate the preliminary distance between each key point in the vehicle key point set P and the guide point cloud region, and select the 3-5 key points closest to the guide as candidate optimal key points.

[0051] Calculate the distance between each candidate optimal keypoint and m reference points (m≥5) in the traverse point cloud set W. Verify the stability of the keypoints by calculating the variance. The formula for calculating the distance variance is: , The variance of the distances between the optimal keypoint candidate and m points on the traverse line is given, where m is the number of reference points on the traverse line. Let be the distance between the candidate optimal keypoint and the j-th traverse reference point. is the average distance between the candidate optimal key point and m wire reference points. Select the key point with the smallest variance and the shortest distance as the optimal key point P(X, Y, Z). This key point is the key position where the vehicle is closest to the wire.

[0052] Step 9: Traverse each point in the wire point cloud set, calculate the Euclidean distance from each point to the optimal key point, and take the minimum value as the shortest distance d between the construction vehicle and the wire.

[0053] Traverse each wire point in the wire point cloud set W to ensure that no wire point is missed and the ranging comprehensiveness is guaranteed. Using the Euclidean distance formula, calculate the distance between each wire point in the wire point cloud set W and the optimal key point P of the vehicle. The formula is: ; is the distance (m) between the optimal key point of the vehicle and the j-th wire point, are the three-dimensional coordinates of the optimal key point of the vehicle, are the three-dimensional coordinates of the j-th wire point. Select the minimum value d from all the distances as the shortest distance between the construction vehicle and the wire, that is: .

[0054] Step 10: Set the alarm threshold and the early warning threshold, and compare the shortest distance d with the alarm threshold and the early warning threshold to implement alarm judgment.

[0055] (1) Dual threshold setting: Set the alarm threshold D1 and the early warning threshold D2. The alarm threshold is the standard distance in the "Electric Power Safety Work Regulations of State Grid Corporation of China", and the early warning threshold is a custom distance threshold, which is flexibly set based on factors such as on-site operation requirements and environmental complexity (such as wind force, vegetation occlusion), etc. The early warning threshold should be greater than the alarm threshold corresponding to the voltage level.

[0056] (2) Early warning and alarm judgment logic: Compare the calculated shortest distance d with the alarm threshold D1 and the early warning threshold D2 respectively to distinguish three states: safe, early warning, and alarm. The specific judgment rules are as follows: 1. If d < D1: It is determined as a major safety hazard, immediately trigger an alarm, synchronously send it to the mobile phones of on-site management personnel and operation and maintenance responsible persons, and at the same time automatically record key information such as the alarm time, the shortest distance, the vehicle position, and the wire voltage level for subsequent traceability, and trigger the on-site operation suspension instruction until the hidden danger is eliminated; 2. If D1 ≤ d < D2: It is determined as a potential safety hazard, trigger an early warning, and a pop-up window prompt will appear on the background system, only notify the on-site management personnel, do not trigger the operation suspension, and remind the management personnel to promptly guide the construction vehicle to adjust its position to avoid approaching the alarm threshold; 3. If d ≥ D2: It is determined as a safe state, and no early warning and alarm are triggered.

[0057] Taking channel 0013#-0014# of a 220kV line as a pilot channel, point cloud data was collected using drones, such as... Figure 6 As shown, two-dimensional images of the channel are acquired from the visualization device on tower #013. Through data preprocessing, 2D-3D point pair matching, and pose calculation, a matching relationship is constructed between the three-dimensional point cloud and the two-dimensional image. Image target recognition and segmentation are used to segment the crane type in the visualization image, such as... Figure 7 As shown, the result of simulating the vehicle's position in the form of a rectangular box in the point cloud data and measuring the distance is 47.53m. Based on the power safety work regulations, at a voltage level of 220kV, the minimum safe distance between construction vehicles and overhead transmission lines and other live conductors is 6.0m. Therefore, the alarm threshold D1 is set to 6.0m, and the warning threshold D2 is set to 16m (the warning threshold is adjustable). The current shortest distance d between the crane and the conductor is... crane >D2, therefore it is determined to be a safe state and no warning or alarm will be triggered.

[0058] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of construction safety of long-span tunnels based on key point linkage matching, characterized in that, It includes: Collect the single - time three - dimensional point cloud data and images of the large - span channel, automatically segment the wire from the three - dimensional point cloud data to obtain a wire point cloud set, extract the image feature points, and form an image feature point set; Project the wire point cloud set into the image coordinate system, perform association matching with the image feature point set, and eliminate the mis - matched points to obtain a 2D - 3D point pair set; Solve the initial pose of the visualization device based on the 2D - 3D point pair set, and perform secondary fine calibration on the initial pose to obtain the optimal pose; Construct an object segmentation model based on the improved U - Net, perform object segmentation of construction vehicles on the image, extract the vehicle contour, convert the key points of the vehicle contour into three - dimensional coordinates, and select the optimal key points through preliminary distance screening and variance reverse verification; Traverse each point in the wire point cloud set, calculate the Euclidean distance from each point to the optimal key point, and take the minimum value as the shortest distance d between the construction vehicle and the wire; Set an alarm threshold D1 and a warning threshold D2; compare the shortest distance d with the alarm threshold D1 and the warning threshold D2: if d < D1, trigger an alarm; if D1 ≤ d < D2, trigger a warning; if d ≥ D2, it is in a safe state.

2. The method for early warning of construction safety of long-span passages based on key point linkage matching according to claim 1, characterized in that, Fit a spatial straight - line model, calculate the Euclidean distance from the points in the three - dimensional point cloud data to the straight line, determine the points with a distance less than the set threshold as wire points, and after iteration, retain the point set corresponding to the spatial straight - line model with the largest number of inlier points as the wire point cloud set.

3. The method for early warning of construction safety of large-span passages based on key point linkage matching according to claim 1, characterized in that, Use the SIFT algorithm to extract image feature points, extract SIFT key points and 128 - dimensional descriptors, and screen the extracted feature points to retain the valid image feature points to form an image feature point set.

4. The construction safety early warning method for large-span passages based on key point linkage matching according to claim 3, characterized in that, The association matching of projecting the wire point cloud set into the image coordinate system with the image feature point set specifically includes: project the wire point cloud set to the image pixel coordinates in the image coordinate system based on the lidar extrinsic matrix and the camera intrinsic matrix, calculate the Euclidean distance between each projected point and the SIFT key points, set a matching threshold to screen the initial matching point pairs, then use the RANSAC algorithm to eliminate the mis - matched points, retain the inlier points with a projection error less than or equal to 1 pixel, obtain a pure 2D - 3D point pair set, and perform Gaussian filtering on the matched wire point cloud set.

5. The method for early warning of construction safety of large-span passages based on key point linkage matching according to claim 1, characterized in that, Use the UPnP algorithm, based on the camera pinhole imaging model, take the 2D - 3D point pair set as the input, construct a least - squares optimization objective function, and solve the initial rotation matrix R and translation vector T that minimize the projection error.

6. The construction safety early warning method for large-span passages based on key point linkage matching according to claim 5, characterized in that, Perform secondary fine calibration on the initial pose. Use the GICP algorithm, take the initial pose as the iteration initial value, consider the point - to - point, point - to - plane distances and covariance constraints, and iterate the GICP optimization objective function to output the fine - calibrated optimal pose.

7. The method for early warning of construction safety of large-span passages based on key point linkage matching according to claim 1, characterized in that, Construct an object segmentation model based on the improved U - Net, perform object segmentation of construction vehicles on the image. Specifically: introduce the SE - Net attention mechanism into the traditional U - Net model, use the cross - entropy loss function and the Adam optimizer, input the image into the trained object segmentation model, output the binary image of the vehicle contour, and then extract the complete vehicle contour through the Canny edge detection algorithm.

8. The method for early warning of construction safety of long-span passages based on key point linkage matching according to claim 6, characterized in that, The Harris corner detection algorithm is used to extract key points of the vehicle contour. The pixel coordinates of the key points are converted into three-dimensional coordinates by the camera pinhole imaging model and the optimal pose. The preliminary distance between each key point and the guide reference point is calculated, and multiple candidate key points with the closest distance are selected. The distance variance between each candidate key point and multiple reference points in the guide point cloud set is calculated, and the key point with the smallest distance variance and the shortest distance is selected as the optimal key point.

9. The method for early warning of construction safety of long-span passages based on key point linkage matching according to claim 8, characterized in that, Traverse each point in the traverse point cloud set, calculate the Euclidean distance to the optimal key point, and use a three-dimensional traversal algorithm to calculate the Euclidean distance between each point in the traverse point cloud set and the optimal key point of the vehicle in turn. Select the minimum value from all distances as the shortest distance d between the construction vehicle and the traverse.

10. The method for early warning of construction safety of long-span passages based on key point linkage matching according to claim 1, characterized in that, The alarm threshold D1 is set based on safety regulations and standards, and the warning threshold D2 is set based on on-site operation requirements and environmental complexity, with D2 > D1.

Images