Flow guide facility monitoring method and system based on multi-source perception and dynamic feature recognition

Abstract

The invention relates to the technical field of safety monitoring, in particular to a diversion facility monitoring method and system based on multi-source sensing and dynamic feature recognition, and the method comprises the steps: obtaining construction scene synchronization data through laser radar and visible light camera cross-modal sensing; removing outliers based on point cloud statistics and screening and identifying diversion facility targets according to long-strip-shaped geometric features; realizing alignment of adjacent frames by adopting ISS key point detection and an iterative weighted registration algorithm; extracting multi-dimensional physical characteristics of point cloud density change rate, surface curvature difference and normal vector field divergence, and generating a comprehensive change index by combining dynamic weighted fusion of facility material rigidity; and based on historical monitoring data, adaptive adjustment of the weight and the threshold is realized through kernel density estimation and gradient descent optimization, and multi-level early warning is triggered. According to the invention, all-weather automatic monitoring and intelligent early warning of the state of the diversion facility are realized, and the risk identification accuracy and response speed in a complex construction environment are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology, and in particular to a method and system for monitoring diversion facilities using multi-source sensing and dynamic feature recognition. Background Technology

[0002] In recent years, with the continuous improvement of the highway network, the focus of expressway construction has gradually shifted from new construction to the reconstruction and expansion of existing roads. In reconstruction and expansion projects, construction teams must carry out operations while ensuring the normal passage of existing lanes. The site environment is complex, traffic organization is challenging, and safety risks are prominent. As key facilities for maintaining traffic order and separating construction areas from traffic areas, the stability of traffic diversion facilities directly affects operational and driving safety. Once traffic diversion facilities shift, collapse, or are damaged, it can easily cause traffic chaos or even safety accidents, posing significant hazards to both construction workers and drivers.

[0003] Under the current transportation development situation, real-time and automated monitoring technologies for the status of traffic diversion facilities in reconstruction and expansion projects are still insufficient. Currently, they rely heavily on manual inspections and on-site monitoring, which suffers from low efficiency, slow response, and limited coverage. To address these pain points, it is necessary to fully utilize information technology and intelligent sensing technologies, and through multidisciplinary integration, conduct research on real-time monitoring and risk early warning technologies for the morphological integrity, spatial stability, and structural status of traffic diversion facilities. This will enable a shift from "human-based" to "technology-based" approaches, comprehensively improving the safety management level of highway reconstruction and expansion projects. Summary of the Invention

[0004] Therefore, the present invention provides a method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition, in order to solve the aforementioned problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides a method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition, comprising: Step S1: The diversion facilities at the construction site are simultaneously scanned and imaged using a cross-modal sensing system composed of lidar and a visible light camera to obtain multi-source synchronous data. Step S2: Preprocess the multi-source synchronous dataset and identify the monitoring target dataset based on the target features; Step S3: Register the monitoring target datasets of adjacent monitoring frames to obtain cross-frame spatial registration results; Step S4: Calculate the local point cloud density change rate, surface average curvature difference, and local normal vector field divergence based on the cross-frame spatial registration results. Generate a comprehensive change index value through weighted fusion and generate an initial monitoring strategy based on a preset risk threshold. Step S5: Based on historical monitoring data and feedback information, the initial monitoring strategy is adaptively optimized to obtain an optimized monitoring strategy. When the changing characteristics exceed the optimized monitoring threshold, an early warning is triggered and relevant personnel are notified.

[0006] Furthermore, the process of step S2 includes: Calculate the distance statistics between each data point and its spatial neighbors based on the multi-source synchronous data, and remove outliers that deviate from the statistical mean by more than a preset outlier threshold to obtain filtered point cloud data. Cluster analysis is performed on the filtered point cloud data to calculate the minimum bounding box of each point cloud cluster. Candidate objects with a minimum bounding box aspect ratio greater than a preset geometric threshold are retained to obtain a geometric shape candidate set. For the candidate set of geometric shapes, the distribution dispersion of the point cloud in the vertical direction is analyzed by the voxelization method, and abnormal targets with vertical thickness variance greater than the preset consistency tolerance range are eliminated to obtain the monitoring target dataset.

[0007] Furthermore, the process of retaining candidate objects whose aspect ratio of the smallest circumscribed cuboid is greater than a preset geometric threshold to obtain a candidate set of geometric shapes includes: The maximum extension direction of the point cloud cluster projected onto the horizontal plane is calculated as the principal direction, and the direction orthogonal to this principal direction is calculated as the secondary direction. The ratio of the length in the main direction to the length in the secondary direction is used as the basis for determining the aspect ratio. Only point cloud clusters that meet the preset geometric constraints in terms of aspect ratio, absolute length, and absolute height are retained as real flow guiding facilities.

[0008] Furthermore, the process of step S3 includes: Perform ISS keypoint detection on the target dataset of the current frame, calculate the local shape feature descriptor of each keypoint, and obtain the feature set of the current frame; ISS keypoint detection and feature descriptor calculation are also performed on the target dataset of the previous frame to obtain the reference frame feature set; Feature matching is performed based on the current frame feature set and the reference frame feature set. Reliable matching point pairs are selected by the nearest neighbor distance ratio method, and coarse registration is completed by rigid transformation matrix estimation. Based on coarse registration, an iterative optimization algorithm is used for fine registration. Point-to-point distance constraints and point-to-area distance constraints are introduced into the iterative objective function. Fine registration is achieved by minimizing the comprehensive distance error between the source point cloud and the target point cloud, resulting in a time-aligned dataset.

[0009] Furthermore, the process of performing fine registration using an iterative optimization algorithm based on coarse registration includes: In each iteration, the weight coefficient of each matching point pair is dynamically calculated. This weight coefficient is inversely proportional to the spatial distance between the point pairs and directly proportional to the cosine of the angle between the normal vectors. Matching point pairs with large distances or significant differences in normal vectors are assigned lower weights. The optimal rigid body transformation matrix is ​​solved by the weighted least squares optimization method, so that the registration process is robust to local occlusion and noise interference.

[0010] Furthermore, the process of step S4 includes: Based on the time-series aligned dataset, spherical neighborhood search structures are established for the reference point cloud and the target point cloud respectively. The rate of change of the number of neighboring points of each sampling point in the spherical neighborhood is calculated to obtain the local density change characteristics. For each point in the temporally aligned dataset, calculate the covariance matrix of its neighboring points, obtain the principal curvature components through eigenvalue decomposition and calculate the average curvature, and obtain the absolute value of the curvature difference between the corresponding points of the reference point cloud and the target point cloud to obtain the surface curvature difference feature. For each point in the time-aligned dataset, calculate the covariance matrix of its neighboring point set, extract the eigenvector corresponding to the minimum eigenvalue as the normal vector, and statistically analyze the consistency of the normal vectors of the reference point cloud and the target point cloud in their corresponding neighborhoods to obtain the divergence feature of the normal vector field. The local density variation features, surface curvature difference features, and normal vector field divergence features are normalized to eliminate the differences in the dimensions of each feature, resulting in a normalized feature vector set. Based on the type attribute of the diversion facility, a weight coefficient is assigned to each normalized feature, and the comprehensive change index value is calculated by weighted summation. Combined with the preset risk threshold, an initial monitoring strategy is formed.

[0011] Furthermore, the process of assigning weight coefficients to each normalized feature based on the diversion facility type attribute includes: A correlation model between the physical properties of the diversion facility and its risk-sensitive characteristics is pre-established. Based on the material stiffness, geometric dimensions, and installation method of the current monitoring target, the basic weight values ​​of each normalized characteristic are calculated. Among them, the basic weight value of the curvature difference characteristic of rigid diversion facilities is higher than that of density change characteristics and normal vector divergence characteristics, while the basic weight value of the normal vector divergence characteristic of flexible diversion facilities is higher than that of curvature difference characteristics and density change characteristics. The basic weight values ​​are dynamically adjusted based on the real-time correction factors of vibration intensity and traffic flow at the construction site to obtain adaptive weight coefficients.

[0012] Furthermore, the process of step S5 includes: Collect the comprehensive change index value sequence and the corresponding false alarm label sequence generated within the historical monitoring period to establish a historical monitoring database; Statistical analysis is performed on the historical monitoring database to identify the characteristic distribution patterns under high-frequency false alarm conditions, calculate the dispersion contribution rate of each geometric feature in the false alarm samples, and obtain a false alarm feature analysis report. Based on the false alarm feature analysis report, the gradient descent method is used to iteratively optimize the feature fusion weight coefficients, so that the comprehensive change index value under false alarm conditions converges to the low-risk range, and a dynamic weight adjustment strategy is obtained. Probability density estimation is performed on the comprehensive change index values ​​in the historical monitoring database, a nonlinear mapping function for risk level is constructed, and the early warning threshold and alarm threshold are dynamically adjusted according to the safety level requirements of the current construction stage to form an adaptive threshold configuration scheme. The dynamic weight adjustment strategy and adaptive threshold configuration scheme are applied to the current monitoring data processing flow to obtain an optimized monitoring strategy. When the real-time calculated comprehensive change index value exceeds the adaptive threshold configuration scheme, a graded early warning is triggered and relevant personnel are notified through the communication module.

[0013] Furthermore, the process of estimating the probability density of comprehensive change index values ​​in the historical monitoring database, constructing a nonlinear mapping function for risk levels, and dynamically adjusting the early warning threshold and alarm threshold according to the safety level requirements of the current construction stage to form an adaptive threshold configuration scheme includes: A nonparametric probability distribution model is constructed using the kernel density estimation method, and the risk quantile is calculated as the initial threshold based on the distribution function. The initial threshold is bidirectionally corrected by the false alarm rate and missed detection rate fed back from the construction site. When the false alarm rate exceeds the preset tolerance, the warning threshold is increased, and when the missed detection rate exceeds the preset tolerance, the alarm threshold is decreased, forming an adaptive threshold configuration scheme with hysteresis characteristics.

[0014] On the other hand, the present invention also provides a multi-source sensing and dynamic feature recognition-based monitoring system for diversion facilities, comprising: The multi-source sensing module is used to synchronously scan and image the diversion facilities at the construction site through a cross-modal sensing system composed of lidar and visible light camera, and acquire multi-source synchronous data. The target recognition module is connected to the multi-source sensing module to preprocess the multi-source synchronous dataset and identify the monitoring target dataset based on the target features. The spatiotemporal registration module is linked to the target recognition module and is used to register the monitoring target datasets of adjacent monitoring frames to obtain cross-frame spatial registration results. The strategy generation module is connected to the spatiotemporal registration module and is used to calculate the local point cloud density change rate, surface average curvature difference and local normal vector field divergence based on the cross-frame spatial registration results. It generates a comprehensive change index value through weighted fusion and generates an initial monitoring strategy based on a preset risk threshold. The strategy optimization module, connected to the strategy generation module, is used to adaptively optimize the initial monitoring strategy based on historical monitoring data and feedback information to obtain an optimized monitoring strategy. When the changing characteristics exceed the optimized monitoring threshold, an early warning is triggered and relevant personnel are notified.

[0015] Compared with existing technologies, the advantages of this invention are as follows: This invention significantly reduces reliance on manual on-site monitoring and periodic inspections through fully automated processing of cross-modal automatic perception, intelligent identification and screening, temporal registration, multi-feature fusion, and adaptive optimization using lidar and visible light cameras. This reduces the safety risks and misjudgments associated with manual observation in complex traffic environments, and drastically shortens the response time from facility displacement, overturning, or damage to risk warning, thus meeting the all-weather real-time monitoring needs of fast-paced road reconstruction and expansion projects. It constructs a comprehensive analysis framework covering multiple dimensions of physical characteristics such as point cloud density changes, surface curvature differences, and normal vector field divergence, enabling accurate identification of various risk modes such as diversion facility defects, deformation, and overturning. It comprehensively captures subtle state changes under construction disturbances and environmental interference, thereby systematically improving the completeness of risk identification and the accuracy of early warning in the monitoring system. It is specifically designed for the high complexity of road construction scenarios. The monitoring method is designed to meet the high dynamic and high-risk requirements, making full use of the inherent physical properties of the diversion facilities, such as their elongated geometric structure and uniform vertical thickness. It effectively addresses harsh environmental factors such as construction dust, vehicle obstruction, and changes in lighting, significantly enhancing the environmental adaptability and reliability of the monitoring system in complex construction sites. Through a modular algorithm architecture and configurable feature weights and dynamic threshold adjustment mechanisms, it allows for customized parameter settings based on construction stage, traffic organization methods, and diversion facility types, flexibly adapting to diverse application scenarios such as road expansion, bridge maintenance, and tunnel construction. This effectively meets the refined safety management needs under specific engineering conditions. Through automated monitoring and intelligent early warning, it promotes the transformation and upgrading of construction safety management from traditional "human-based" to "technology-based" approaches, enabling proactive identification and early intervention of risks and hazards. This improves the overall safety control level and construction efficiency of highway reconstruction and expansion projects, ensuring the safety of life and property for both construction and traffic parties. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition provided by the present invention; Figure 2 A flowchart illustrating step S2 in a multi-source sensing and dynamic feature recognition method for monitoring diversion facilities provided by the present invention; Figure 3 This is a flowchart illustrating step S3 in a multi-source sensing and dynamic feature recognition method for monitoring diversion facilities provided by the present invention. Figure 4 This is a schematic diagram of the structure of a diversion facility monitoring method based on multi-source sensing and dynamic feature recognition provided by the present invention. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0020] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] Please see Figure 1 As shown, the present invention provides a method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition, comprising: Step S1: The diversion facilities at the construction site are simultaneously scanned and imaged using a cross-modal sensing system composed of lidar and a visible light camera to obtain multi-source synchronous data. Specifically, the cross-modal sensing system employs a cross-modal coupled sensing mechanism, referencing a precise time synchronization scheme based on the IEEE 1588PTPv2 protocol. In this multi-source sensing system, the lidar and industrial camera, based on a shared field-of-view optical design (optical axis angle 30°±2°), establish a μs-level time synchronization network via an industrial switch supporting the PTP protocol. The camera is configured with an 850-1100nm bandpass filter, forming a spectral response overlap region with the lidar's 905nm band. The data acquisition trigger signals for both are uniformly distributed by the PTP master clock, achieving a spatiotemporal alignment error ≤1.5μs. This scheme reduces hardware costs by approximately 10% compared to traditional FPGA triggering, while simultaneously meeting the stringent requirements of multi-source data synchronization for flow control facility monitoring.

[0022] Step S2: Preprocess the multi-source synchronous dataset and identify the monitoring target dataset based on the target features; Specifically, such as Figure 2 As shown, the process of step S2 includes: Step S21: Calculate the distance statistics between each data point and its spatial neighbors based on the multi-source synchronous data, and remove outliers that deviate from the statistical mean by more than a preset outlier threshold to obtain filtered point cloud data. Specifically, for 3D point cloud data in a multi-source synchronous dataset, a kd-tree spatial index structure is first established to accelerate neighborhood search. For each data point, its k nearest neighbors in space are retrieved (the value of k is adaptively determined according to the point cloud density, usually between 20 and 50), and the average Euclidean distance from the point to its k nearest neighbors is calculated. The average distance distribution of all points is statistically analyzed, and its statistical mean μ and standard deviation σ are calculated. An outlier determination threshold is set to μ + α·σ (α is a sensitivity coefficient used to adjust the removal intensity). Data points with an average distance exceeding this threshold are marked as outliers and removed, resulting in filtered point cloud data.

[0023] Specifically, α is typically set between 2.0 and 3.0. In dusty road construction scenarios, α is set to 2.5-3.0 to avoid accidentally deleting real target points; in relatively clean, enclosed construction areas, α is set to 2.0-2.5 to enhance noise suppression. This threshold is adaptively adjusted based on environmental factors such as visibility at the construction site and equipment vibration intensity.

[0024] Step S22: Perform cluster analysis on the filtered point cloud data, calculate the minimum bounding box of each point cloud cluster, and retain candidate objects whose aspect ratio of the minimum bounding box is greater than a preset geometric threshold to obtain a geometric shape candidate set. Specifically, the process of retaining candidate objects whose aspect ratio of the smallest circumscribed cuboid is greater than a preset geometric threshold to obtain a candidate set of geometric shapes includes: The maximum extension direction of the point cloud cluster projected onto the horizontal plane is calculated as the principal direction, and the direction orthogonal to this principal direction is calculated as the secondary direction. Specifically, the filtered point cloud data undergoes region-growing clustering analysis based on Euclidean distance. The system searches for spatially neighboring points starting from an initial seed point, aggregating points with a spatial distance less than a preset neighborhood threshold into a single connected region, forming point cloud clusters representing candidate objects. This neighborhood threshold is dynamically set based on the LiDAR scanning density, typically 10-15 cm, ensuring that points on the same flow guidance facility are completely aggregated while effectively separating different objects. For each point cloud cluster, principal component analysis is used to calculate its three-dimensional covariance matrix, and three orthogonal eigenvectors are obtained through eigenvalue decomposition. The eigenvector corresponding to the smallest eigenvalue is determined as the height direction, while the other two eigenvectors constitute the principal direction (corresponding to the largest eigenvalue) and the secondary direction (corresponding to the middle eigenvalue) in the horizontal plane. The point cloud cluster is projected onto the principal and secondary directions respectively, and the difference between the maximum and minimum projected coordinates is calculated to obtain the length and width of the circumscribed cuboid; the height is obtained by projecting onto the height direction.

[0025] The ratio of the length in the main direction to the length in the secondary direction is used as the basis for determining the aspect ratio. Only point cloud clusters that meet the preset geometric constraints in terms of aspect ratio, absolute length, and absolute height are retained as real flow guiding facilities.

[0026] Specifically, the aspect ratio is determined based on the ratio of the length in the primary direction to the width in the secondary direction. The system's preset geometric constraints include: an aspect ratio greater than 5:1, an absolute length greater than 1 meter, and an absolute height within the range of 0.5-2 meters. A point cloud cluster is only retained as a candidate set of geometric shapes if it simultaneously meets all three constraints.

[0027] Step S23: For the candidate set of geometric shapes, analyze the dispersion of the point cloud distribution in the vertical direction using the voxelization method, and remove abnormal targets whose vertical thickness variance is greater than the preset consistency tolerance range to obtain the monitoring target dataset.

[0028] Specifically, the thickness threshold is determined by using the variance of the vertical point density distribution of a statistical standard flow guidance facility sample. For plastic cones, the variance threshold is set at 8-12 cm; for metal guardrails, it is set at 5-10 cm. This threshold is adaptively adjusted based on the rigidity of the target material: the stronger the rigidity, the stricter (smaller) the threshold. This effectively identifies and eliminates interfering objects such as guardrails and curbs that, although elongated, have uneven thickness. These objects exhibit significant gradient changes in their vertical point density distribution, with variances far exceeding those of actual flow guidance facilities. In environments with strong vibrations, to avoid misjudging point cloud discretization caused by shaking as uneven thickness, this threshold can be temporarily relaxed by 10%-15%.

[0029] Specifically, for each point cloud cluster in the candidate set of geometric shapes, voxelization is performed along the vertical direction (Z-axis), dividing the space into equally spaced three-dimensional voxel grids. The number of points contained in each voxel grid is counted to obtain the point density distribution vector along the Z-axis. The variance of this distribution vector is calculated; a larger variance indicates a more dispersed distribution of the point cloud in the vertical direction and poorer structural consistency. A thickness variance threshold is set, and candidate objects with variances exceeding this threshold are identified as anomalous targets and removed.

[0030] Step S3: Register the monitoring target datasets of adjacent monitoring frames to obtain cross-frame spatial registration results; Specifically, such as Figure 3 As shown, the process of step S3 includes: Step S31: Perform ISS keypoint detection on the target dataset of the current frame, calculate the local shape feature descriptor of each keypoint, and obtain the feature set of the current frame; Specifically, a kd-tree spatial index structure for the point cloud is constructed to support fast neighborhood queries. For each data point, a search radius r is set (15-25 cm based on the point cloud density of the diversion facility), and all neighboring points within this radius are retrieved to form a local neighborhood. The covariance matrix of this neighborhood is calculated, specifically by calculating the 3D coordinate differences from all points in the neighborhood to the center point, constructing a 3×3 matrix, and averaging the results to obtain the covariance matrix. Eigenvalue decomposition is performed on the covariance matrix to obtain three eigenvalues ​​λ1, λ2, and λ3 (arranged in descending order). The core criterion of the ISS algorithm is that the ratio of λ2 to λ1 must be greater than the significance threshold, and the ratio of λ3 to λ2 must also be greater than the significance threshold. These two thresholds are typically set to 0.6 and 0.5, respectively, to ensure that the selected points have significant structural changes in all three principal directions. Simultaneously, only one point with the largest eigenvalue is retained within each radius r neighborhood as a key point to avoid excessive concentration of key points.

[0031] For each keypoint, a local shape feature descriptor is calculated within its extended neighborhood of twice the radius 2r. The covariance matrix of this extended neighborhood is calculated and eigenvalues ​​are decomposed. The three eigenvalues ​​are normalized to form a three-dimensional feature vector. At the same time, the local point density (number of points per unit volume) is calculated as the fourth dimension, and the angle between the normal vector (the eigenvector corresponding to the smallest eigenvalue) and the vertical direction is calculated as the fifth dimension, finally forming a five-dimensional feature descriptor vector.

[0032] Step S32: Perform ISS keypoint detection and feature descriptor calculation on the previous frame monitoring target dataset to obtain the reference frame feature set; Specifically, the same ISS detection process is performed on the target dataset of the previous frame to obtain a reference frame feature set. For each five-dimensional descriptor vector of a key point in the current frame, the two closest and second-closest descriptor vectors in the reference frame feature set are searched for using Euclidean distance. The ratio of the closest distance to the second-closest distance is calculated. If this ratio is less than a preset reliability threshold (usually set to 0.75), the matching pair is considered to have high uniqueness and is retained; otherwise, it is considered a fuzzy match and is discarded.

[0033] Step S33: Perform feature matching based on the current frame feature set and the reference frame feature set, use the nearest neighbor distance ratio method to screen reliable matching point pairs, and complete coarse registration by estimating the rigid transformation matrix; Specifically, for the retained matching point pairs, the Singular Value Decomposition (SVD) algorithm is used to estimate the optimal rigid body transformation. The centroid matrices of all matching point pairs are constructed, centered, and then a covariance matrix is ​​constructed. SVD decomposition is performed on the covariance matrix to obtain initial estimates of the rotation matrix R and the translation vector t. This transformation is applied to the current frame's point cloud to complete coarse registration, roughly aligning the two frames' point clouds. The root mean square error after coarse registration is typically controlled within the range of 5-10 cm.

[0034] Step S34: Based on coarse registration, fine registration is performed using an iterative optimization algorithm. Point-to-point distance constraints and point-to-surface distance constraints are introduced into the iterative objective function. Fine registration is achieved by minimizing the comprehensive distance error between the source point cloud and the target point cloud, resulting in a time-aligned dataset.

[0035] Specifically, the process of fine registration based on coarse registration using an iterative optimization algorithm includes: In each iteration, the weight coefficient of each matching point pair is dynamically calculated. This weight coefficient is inversely proportional to the spatial distance between the point pairs and directly proportional to the cosine of the angle between the normal vectors. Matching point pairs with large distances or significant differences in normal vectors are assigned lower weights. The optimal rigid body transformation matrix is ​​solved by the weighted least squares optimization method, so that the registration process is robust to local occlusion and noise interference.

[0036] Specifically, based on coarse registration, an iterative weighted nearest-point algorithm is used for fine registration. Each iteration includes three core operations: First, corresponding point search: For each point in the current frame, search for the nearest neighbor in the reference frame to establish a correspondence. At the same time, use kd-tree to accelerate the search and limit the maximum search distance (set to 5 cm) to avoid incorrect matching.

[0037] Second, dynamic weight calculation: For each pair of matched points, calculate its spatial distance d and spatial weight w_s. w_s equals the preset maximum distance threshold divided by the sum of d and the threshold, ensuring that the greater the distance, the smaller the weight. Calculate the angle θ between the normal vectors of the two corresponding points, and calculate the normal vector weight w_n, which is equal to the absolute value of the cosine of the angle, ensuring that the smaller the angle, the greater the weight. Multiply w_s and w_n to obtain the comprehensive weight w_i, and normalize w_i so that the sum of the squares of all weights is 1.

[0038] Matching point pairs whose distance exceeds the maximum threshold or whose normal vector angle exceeds 60 degrees are directly assigned zero weight and discarded.

[0039] Third, weighted least squares optimization: Construct a composite objective function containing point-to-point distance terms and point-to-surface distance terms. The point-to-point distance term is the squared Euclidean distance between the corresponding point coordinates, and the point-to-surface distance term is the squared perpendicular distance from the source point to the local fitted plane of the target point (fitted by the target point and its neighboring points). Multiply each distance term by its corresponding dynamic weight w_i and sum them. Use the LM algorithm iteratively to find the rigid body transformation parameters that minimize the objective function.

[0040] Specifically, higher weight refers to matching pairs with w_i > 0.7, which are considered high-confidence matches; medium weight refers to matching pairs with 0.3 ≤ w_i ≤ 0.7, which participate in optimization but have a moderate impact; lower weight refers to matching pairs with w_i < 0.3, which are considered unreliable matches and have a weak effect on optimization.

[0041] Specifically, the weighting coefficient w_i is composed of the product of two independent factors, achieving adaptive adjustment: Spatial distance factor: w_s = d_max / (d_i + d_max), where d_i is the spatial distance between the current matched point pairs, and d_max is the preset maximum distance threshold (set to 5 cm). This factor ensures that the smaller the distance between matched point pairs, the greater their registration contribution, while point pairs with a distance close to the threshold contribute almost nothing.

[0042] Normal vector consistency factor: w_n = |cos(θ_i)|, where θ_i is the angle between the local normal vectors of two corresponding points. The smaller the angle, the closer the cosine value is to 1, and the greater the weight; the larger the angle, the smaller the weight. When the angle exceeds 60 degrees, the weight approaches 0, and matching pairs with significant differences in normal vectors are automatically eliminated. In this embodiment, when θ_i = 0° (completely parallel), w_n = 1 (maximum weight); when θ_i = 30°, w_n ≈ 0.87; when θ_i = 60°, w_n = 0.5; when θ_i = 90° (orthogonal), w_n = 0.

[0043] The overall weights w_i = w_s × w_n are calculated and normalized so that the sum of squares of all weights is 1. Matching pairs with w_i less than 0.1 are assigned zero weights and discarded, and do not participate in subsequent optimization calculations.

[0044] Specifically, the point-to-point distance constraint measures the spatial positional deviation between corresponding points in the source point cloud and the target point cloud. For the i-th matching point pair, its point-to-point distance error is defined as the squared Euclidean distance between the coordinates T(p_i) of the source point p_i after the current transformation matrix T and the target point q_i. This constraint directly penalizes spatial positional deviations, ensuring that the corresponding points coincide as much as possible in three-dimensional space after registration, which is the fundamental guarantee for registration accuracy.

[0045] Point-to-plane distance constraints are a key technique for improving registration robustness. For the target point q_i in the i-th matching point pair, k neighboring points (k = 10-20) are searched in its reference frame, and these k points are used to fit a local plane. The fitting method uses the least squares method, calculating the covariance matrix of the point set, taking the eigenvector corresponding to its smallest eigenvalue as the plane normal vector, and using the centroid of the point set as the reference point on the plane.

[0046] For the transformed point T(p_i) of the source point p_i, the vertical distance from it to the fitted plane is calculated as the point-to-plane distance error. The advantage of this constraint is that when the point cloud in the corresponding region of the reference frame is sparse due to occlusion or incomplete scanning, the point-to-plane constraint can still utilize local planar structure information to guide the registration to converge in a reasonable direction, avoiding registration failure due to a small number of incorrect matches.

[0047] Specifically, a triple iteration termination condition is set: first, the maximum number of iterations (set to 50); second, the root mean square error change rate (the error decrease between two adjacent iterations is less than 0.1 mm); and third, the effective matching point pair ratio (stopping when it is less than 15% of the total number of points). The iteration terminates when any one of the conditions is met, and the final rotation matrix and translation vector are output as the cross-frame spatial registration result, at which point the registration accuracy is better than 5 cm.

[0048] Step S4: Calculate the local point cloud density change rate, surface average curvature difference, and local normal vector field divergence based on the cross-frame spatial registration results. Generate a comprehensive change index value through weighted fusion and generate an initial monitoring strategy based on a preset risk threshold. Specifically, step S4 includes the following process: Step S41: Based on the time-series aligned dataset, establish spherical neighborhood search structures for the reference point cloud and the target point cloud respectively, calculate the rate of change of the number of neighboring points of each sampling point in the spherical neighborhood, and obtain the local density change features. Specifically, kd-tree spherical neighborhood search structures are constructed for both the reference point cloud (previous frame) and the target point cloud (next frame). For each sampling point, a spherical neighborhood radius r is set (10-15 cm based on the point cloud density of the diversion facility), and all neighboring points within this radius are retrieved and their counts are recorded. For any point p in the target point cloud, the number of neighboring points N_target contained in its spherical neighborhood is counted; for the corresponding point q in the reference point cloud, the number of neighboring points N_ref within the same radius spherical neighborhood is counted. The local density change rate is defined as (N_target - N_ref) / (N_ref + ε), where ε is a minimum value (e.g., 0.01) to prevent division by zero. This ratio quantifies the temporal change in the density of the point cloud distribution. When the diversion facility is damaged or partially collapsed, the point cloud density in the corresponding area decreases significantly, and this feature value shows a clear positive response.

[0049] Step S42: Calculate the covariance matrix of each point in the temporal alignment dataset for its neighboring points, obtain the principal curvature components through feature decomposition and calculate the average curvature, and obtain the absolute value of the curvature difference between the corresponding points of the reference point cloud and the target point cloud to obtain the surface curvature difference feature. Specifically, for each valid point in the time-aligned dataset, its k nearest neighbors (k is 20-30) are retrieved to form a local neighborhood. A 3×3 covariance matrix is ​​calculated for this neighborhood, where each element is the average of the outer products of the differences between the coordinates of each point and the centroid coordinates. Eigenvalue decomposition is performed on the covariance matrix to obtain three eigenvalues ​​λ1, λ2, and λ3 (arranged in descending order). The maximum principal curvature κ1 is approximately λ2 / (λ1+λ2+λ3), and the minimum principal curvature κ2 is approximately λ3 / (λ1+λ2+λ3). The average curvature is calculated as (κ1 + κ2) / 2, characterizing the degree of surface bending. After calculating the average curvature for corresponding points in the reference and target point clouds, the absolute difference between the two is taken as the surface curvature difference feature. When the flow guidance facility undergoes bending deformation or surface damage, the local curvature changes significantly, and this feature effectively captures the deformation risk.

[0050] Step S43: For each point in the time-aligned dataset, calculate the covariance matrix of its neighboring point set, extract the eigenvector corresponding to the minimum eigenvalue as the normal vector, and statistically analyze the consistency of the normal vectors of the reference point cloud and the target point cloud in their corresponding neighborhoods to obtain the divergence feature of the normal vector field. Specifically, for each point in the temporally aligned dataset, a local covariance matrix is ​​constructed and eigenvalue decomposition is performed using the same set of neighboring points (k = 20-30) as used in the curvature calculation. The eigenvector corresponding to the smallest eigenvalue is the normal vector n of the local plane containing that point. To evaluate the consistency of the normal vector field, the cosine similarity of the angle between the target point cloud and the reference point cloud's corresponding neighborhood normal vectors is calculated: consistency = |n_target·n_ref|. For a single point, the divergence feature of the normal vector field is defined as 1 - consistency; the closer the value is to 1, the greater the difference in the normal vector direction. The average divergence feature of all corresponding points is further calculated to obtain the divergence of the normal vector field of the entire frame point cloud. When the diversion facility tilts or overturns as a whole, the angle between the surface normal vector and the vertical direction systematically deviates, and this feature value increases significantly, effectively identifying abnormal areas where surface continuity is interrupted.

[0051] Step S44: Normalize the local density change features, surface curvature difference features, and normal vector field divergence features to eliminate the differences in the dimensions of each feature and obtain a normalized feature vector set. Specifically, the three types of features are normalized to eliminate dimensional differences. For the local density variation feature, z-score standardization is applied: subtracting the historical mean and dividing by the standard deviation ensures the feature values ​​follow a standard normal distribution. For the surface curvature difference feature and the normal vector field divergence feature, min-max normalization is applied: linearly mapping to the [0,1] interval to ensure comparability of different features during subsequent fusion. After normalization, the three types of feature values ​​are within the same dimensional range, resulting in a normalized feature vector set. Feature fusion weight coefficients for the three types of features are established, denoted as: w_density: fusion weight for the local density variation feature; w_curvature: fusion weight for the surface curvature difference feature; w_normal: fusion weight for the normal vector field divergence feature.

[0052] Step S45: Assign weight coefficients to each normalized feature based on the diversion facility type attribute, calculate the comprehensive change index value through weighted summation, and form an initial monitoring strategy by combining it with the preset risk threshold.

[0053] Specifically, the process of assigning weight coefficients to each normalized feature based on the diversion facility type attribute includes: A correlation model between the physical properties of the diversion facility and its risk-sensitive characteristics is pre-established. Based on the material stiffness, geometric dimensions, and installation method of the current monitoring target, the basic weight values ​​of each normalized characteristic are calculated. Among them, the basic weight value of the curvature difference characteristic of rigid diversion facilities is higher than that of density change characteristics and normal vector divergence characteristics, while the basic weight value of the normal vector divergence characteristic of flexible diversion facilities is higher than that of curvature difference characteristics and density change characteristics. Specifically, samples of diversion facilities with different materials (plastic, metal, concrete), different geometric dimensions (height, wall thickness), and different installation methods (fixed, mobile) are collected in advance. Displacement, overturning, and damage risks are simulated in a laboratory environment, and the sensitivity to changes in three types of features is recorded. An empirical mapping table is established: for rigid materials (such as concrete), the sensitivity weight for curvature difference is set to 0.5, the weight for density change is 0.3, and the weight for normal vector divergence is 0.2; for flexible materials (such as plastic), the weight for normal vector divergence is set to 0.5, the weight for curvature difference is 0.2, and the weight for density change is 0.3. During system operation, the basic weight vector [w_density_base, w_curvature_base, w_normal_base] is obtained by querying the correlation model through input or automatic identification of the material type and size specifications of the current monitored target.

[0054] The basic weight values ​​are dynamically adjusted based on the real-time correction factors of vibration intensity and traffic flow at the construction site to obtain adaptive weight coefficients.

[0055] Specifically, the vibration intensity (measured by an accelerometer) and traffic flow (analyzed by camera video) at the construction site are collected in real time and quantified into correction factors in the 0-1 range. The vibration intensity correction factor γ_vib is linearly mapped to the vibration amplitude; the larger the amplitude, the larger the factor. The traffic flow correction factor γ_traf is calculated based on the number of vehicles per unit time; the more vehicles, the larger the factor. The final feature fusion weight calculation formula is as follows: w_density = w_density_base × (1 + 0.3·γ_vib + 0.2·γ_traf); w_curvature = w_curvature_base × (1 + 0.3·γ_vib + 0.2·γ_traf); w_normal = w_normal_base × (1 + 0.3·γ_vib + 0.2·γ_traf).

[0056] The three weights are then normalized so that their sum equals 1. This dynamic adjustment enables the system to automatically increase the weight of deformation-sensitive features under conditions of strong vibration or high flow, thereby improving the targeting of risk identification.

[0057] The normalized three types of feature values ​​are weighted and summed with the feature fusion weights: Comprehensive change index value = w_density·density_feature + w_curvature·curvature_feature + w_normal·normal_feature. This index value is dimensionless and ranges from [0,1]. The larger the value, the more significant the change in the state of the diversion facility.

[0058] Based on statistical analysis of historical monitoring data, a preset static risk threshold is established. Typically, the early warning threshold is set at 0.6, and the alarm threshold at 0.8. When the comprehensive change index value exceeds the early warning threshold, an initial monitoring strategy is generated and the system is marked as "attention"; when it exceeds the alarm threshold, it is marked as "dangerous." The initial monitoring strategy includes three types of characteristic values, a comprehensive index value, a risk level label, and corresponding timestamp information.

[0059] Step S5: Based on historical monitoring data and feedback information, the initial monitoring strategy is adaptively optimized to obtain an optimized monitoring strategy. When the changing characteristics exceed the optimized monitoring threshold, an early warning is triggered and relevant personnel are notified.

[0060] Specifically, step S5 includes the following process: Collect the comprehensive change index value sequence and the corresponding false alarm label sequence generated within the historical monitoring period to establish a historical monitoring database; Specifically, the system continuously collects the comprehensive change index value sequence and the corresponding false alarm marker sequence generated within each monitoring cycle (usually set to 24 hours). The comprehensive change index value sequence includes the local density change characteristics, surface curvature difference characteristics, normal vector field divergence characteristics, and the fused comprehensive index value calculated from each frame of data. The false alarm marker sequence is marked after confirmation by manual inspection or verification by video surveillance, and is divided into three categories: "true warning", "false alarm", and "missed detection". The database adopts a time-series storage structure, and each record includes: timestamp, three types of feature values, comprehensive change index value, initial risk level, final confirmation result, and construction site environmental parameters (vibration intensity, traffic flow, lighting conditions). The database capacity is set to store the data of the most recent 30 days in a rolling manner. When the data volume exceeds the threshold, the earliest date data is automatically deleted to maintain stable storage space. This database provides a data foundation for subsequent statistical analysis, ensuring that the analysis results reflect recent construction characteristics and avoiding interference from outdated data.

[0061] Statistical analysis is performed on the historical monitoring database to identify the characteristic distribution patterns under high-frequency false alarm conditions, calculate the dispersion contribution rate of each geometric feature in the false alarm samples, and obtain a false alarm feature analysis report. Specifically, segmented statistical analysis is performed on the historical monitoring database to identify high-frequency false alarm conditions. Data is divided into several condition categories based on environmental parameters: high vibration conditions (vibration intensity > 0.6), high traffic volume conditions (traffic flow > 0.7), and low light conditions at night (light intensity < 0.3), etc. For false alarm samples under each condition category, the dispersion contribution rate of each geometric feature is calculated. The dispersion contribution rate is defined as the proportion of the variance of a certain feature in the false alarm samples to the total variance of all features. For example, under high vibration conditions, if the variance of the divergent normal vector field feature accounts for 60% of the total variance, its dispersion contribution rate is 0.6, indicating that this feature has poor stability under high vibration environments and is prone to false alarms. A false alarm feature analysis report is generated by synthesizing the analysis results of various conditions, clarifying the reliability ranking of each feature under different conditions. The report format is a feature-condition contribution rate matrix, providing a quantitative basis for subsequent weight optimization.

[0062] Based on the false alarm feature analysis report, the gradient descent method is used to iteratively optimize the feature fusion weight coefficients, so that the comprehensive change index value under false alarm conditions converges to the low-risk range, and a dynamic weight adjustment strategy is obtained. Specifically, based on the false positive feature analysis report, batch gradient descent is used to iteratively optimize the feature fusion weight coefficients. Let the current weight vector be w = [w_density, w_curvature, w_normal], and define the objective function as the sum of the comprehensive change index values ​​of the false positive samples, i.e., minimizing L(w) = Σ(I_i), where I_i is the comprehensive change index value of the i-th false positive sample. The learning rate α of gradient descent is initially set to 0.01, decreasing by 10% every 100 iterations to prevent oscillation. Gradient calculation uses a numerical differentiation method: a small perturbation Δw = 0.001 is applied to each weight component, and the rate of change of the objective function ∂L / ∂w ≈ [L(w+Δw) - L(w)] / Δw is calculated. During iteration, the weight vector is updated in the negative gradient direction: w_new = w_old - α·∇L(w). After each update, the weight vector is projected onto the feasible region (each weight > 0 and the sum is 1) to ensure the physical meaning is valid. The process stops when the objective function decreases by less than 0.001 or the number of iterations reaches 500, yielding the optimal weight vector that converges the comprehensive index value of the false alarm samples towards the low-risk range. This optimization process is executed automatically once a week, allowing the system to continuously learn recent false alarm patterns, and the dynamic weight adjustment strategy is adaptively updated as the construction phase evolves.

[0063] Probability density estimation is performed on the comprehensive change index values ​​in the historical monitoring database, a nonlinear mapping function for risk level is constructed, and the early warning threshold and alarm threshold are dynamically adjusted according to the safety level requirements of the current construction stage to form an adaptive threshold configuration scheme. Specifically, the process of estimating the probability density of comprehensive change index values ​​in the historical monitoring database, constructing a nonlinear mapping function for risk levels, and dynamically adjusting the early warning threshold and alarm threshold according to the safety level requirements of the current construction stage to form an adaptive threshold configuration scheme includes: A nonparametric probability distribution model is constructed using the kernel density estimation method, and the risk quantile is calculated as the initial threshold based on the distribution function. Specifically, kernel density estimation is performed on the comprehensive change index values ​​in the historical monitoring database to construct a nonparametric probability distribution model. A Gaussian kernel function is selected, and the bandwidth parameter h is automatically determined using the Silverman rule: h = 1.06·σ·n^(-1 / 5), where σ is the standard deviation of the comprehensive index values ​​and n is the sample size. Risk quantiles are calculated based on the probability distribution function as initial thresholds. The initial warning threshold is set at the 95th quantile (i.e., 95% of normal data are below this value), and the initial alarm threshold is set at the 98th quantile, ensuring that the thresholds are based on actual data distribution rather than subjective settings.

[0064] The initial threshold is bidirectionally corrected by the false alarm rate and missed detection rate fed back from the construction site. When the false alarm rate exceeds the preset tolerance, the warning threshold is increased, and when the missed detection rate exceeds the preset tolerance, the alarm threshold is decreased, forming an adaptive threshold configuration scheme with hysteresis characteristics.

[0065] Specifically, the two-way correction mechanism dynamically adjusts the thresholds based on the false alarm rate and missed detection rate reported from the construction site. The system calculates the false alarm rate (number of false alarms / total number of warnings) and missed detection rate (number of missed detections / number of actual risks) every 24 hours. A tolerance of 15% is set for the false alarm rate, and 5% for the missed detection rate. When the false alarm rate exceeds 15%, it indicates that the current warning threshold is too low. The warning threshold is increased by Δ = 0.05 × (false alarm rate - 15%), and the alarm threshold is simultaneously increased by 0.5Δ to suppress oversensitivity. When the missed detection rate exceeds 5%, it indicates that the current alarm threshold is too high. The alarm threshold is decreased by Δ = 0.05 × (missed detection rate - 5%), and the warning threshold is simultaneously decreased by 0.3Δ to improve risk detection capabilities. Threshold increases take effect immediately, while threshold decreases take effect after a 6-hour delay to prevent frequent threshold oscillations due to short-term data fluctuations and ensure strategy stability.

[0066] The dynamic weight adjustment strategy and adaptive threshold configuration scheme are applied to the current monitoring data processing flow to obtain an optimized monitoring strategy. When the real-time calculated comprehensive change index value exceeds the adaptive threshold configuration scheme, a graded early warning is triggered and relevant personnel are notified through the communication module.

[0067] Specifically, a dynamic weight adjustment strategy and an adaptive threshold configuration scheme are deployed into the real-time monitoring process. Upon arrival of the current frame data, the latest weight vector is used to calculate the comprehensive change index value, which is then compared with the current effective threshold. If the index value exceeds the warning threshold but is below the alarm threshold, a yellow warning is triggered, and a notification is sent to the on-site safety officer via the 4G communication module. If the value exceeds the alarm threshold, a red warning is triggered, and an emergency response notification is simultaneously sent to the project manager and the supervision unit.

[0068] Specifically, this invention significantly reduces reliance on manual on-site monitoring and periodic inspections through a fully automated process involving cross-modal automatic perception, intelligent identification and screening, temporal registration, multi-feature fusion, and adaptive optimization using lidar and visible light cameras. This reduces the safety risks and misjudgments associated with manual observation in complex traffic environments, and drastically shortens the response time from facility displacement, overturning, or damage to risk warnings, thus meeting the all-weather real-time monitoring needs of fast-paced road reconstruction and expansion projects. It constructs a comprehensive analytical framework covering multi-dimensional physical characteristics such as point cloud density changes, surface curvature differences, and normal vector field divergence, enabling accurate identification of various risk modes such as diversion facility defects, deformation, and overturning. It comprehensively captures subtle state changes under construction disturbances and environmental interference, thereby systematically improving the completeness of risk identification and the accuracy of warnings in the monitoring system. It is specifically designed for the high complexity, high dynamism, and high risk of road construction scenarios. The risk-based demand design monitoring method fully utilizes the inherent physical properties of diversion facilities, such as their elongated geometric structure and uniform vertical thickness, to effectively handle harsh environmental factors such as construction dust, vehicle obstruction, and changes in lighting. This significantly enhances the environmental adaptability and reliability of the monitoring system in complex construction sites. Through a modular algorithm architecture and configurable feature weights and dynamic threshold adjustment mechanisms, customized parameter settings are allowed based on construction stage, traffic organization method, and type of diversion facility. This flexibly adapts to diverse application scenarios such as road expansion, bridge maintenance, and tunnel construction, effectively meeting the refined safety management needs under specific engineering conditions. Through automated monitoring and intelligent early warning, the method promotes the transformation and upgrading of construction safety management from traditional "human defense" to "technology defense," enabling proactive identification and early intervention of risks and hazards. This improves the overall safety control level and construction efficiency of highway reconstruction and expansion projects, ensuring the safety of life and property for both construction and traffic parties.

[0069] Specifically, such as Figure 4 As shown, the present invention also provides a multi-source sensing and dynamic feature recognition-based monitoring system for diversion facilities, comprising: The multi-source sensing module is used to synchronously scan and image the diversion facilities at the construction site through a cross-modal sensing system composed of lidar and visible light camera, and acquire multi-source synchronous data. The target recognition module is connected to the multi-source sensing module to preprocess the multi-source synchronous dataset and identify the monitoring target dataset based on the target features. The spatiotemporal registration module is linked to the target recognition module and is used to register the monitoring target datasets of adjacent monitoring frames to obtain cross-frame spatial registration results. The strategy generation module is connected to the spatiotemporal registration module and is used to calculate the local point cloud density change rate, surface average curvature difference and local normal vector field divergence based on the cross-frame spatial registration results. It generates a comprehensive change index value through weighted fusion and generates an initial monitoring strategy based on a preset risk threshold. The strategy optimization module, connected to the strategy generation module, is used to adaptively optimize the initial monitoring strategy based on historical monitoring data and feedback information to obtain an optimized monitoring strategy. When the changing characteristics exceed the optimized monitoring threshold, an early warning is triggered and relevant personnel are notified.

[0070] Specifically, the multi-source sensing and dynamic feature recognition method for monitoring diversion facilities provided by the present invention can execute the multi-source sensing and dynamic feature recognition system for monitoring diversion facilities in the embodiments of the present invention, and can achieve the same technical effect, which will not be repeated here.

[0071] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition, characterized in that, include: Step S1: The diversion facilities at the construction site are simultaneously scanned and imaged using a cross-modal sensing system composed of lidar and a visible light camera to obtain multi-source synchronous data. Step S2: Preprocess the multi-source synchronous dataset and identify the monitoring target dataset based on the target features; Step S3: Register the monitoring target datasets of adjacent monitoring frames to obtain cross-frame spatial registration results; Step S4: Calculate the local point cloud density change rate, surface average curvature difference, and local normal vector field divergence based on the cross-frame spatial registration results. Generate a comprehensive change index value through weighted fusion and generate an initial monitoring strategy based on a preset risk threshold. Step S5: Based on historical monitoring data and feedback information, the initial monitoring strategy is adaptively optimized to obtain an optimized monitoring strategy. When the changing characteristics exceed the optimized monitoring threshold, an early warning is triggered and relevant personnel are notified.

2. The method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition according to claim 1, characterized in that, The process of step S2 includes: Calculate the distance statistics between each data point and its spatial neighbors based on the multi-source synchronous data, and remove outliers that deviate from the statistical mean by more than a preset outlier threshold to obtain filtered point cloud data. Cluster analysis is performed on the filtered point cloud data to calculate the minimum bounding box of each point cloud cluster. Candidate objects with a minimum bounding box aspect ratio greater than a preset geometric threshold are retained to obtain a geometric shape candidate set. For the candidate set of geometric shapes, the distribution dispersion of the point cloud in the vertical direction is analyzed by the voxelization method, and abnormal targets with vertical thickness variance greater than the preset consistency tolerance range are eliminated to obtain the monitoring target dataset.

3. The method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition according to claim 2, characterized in that, The process of retaining candidate objects whose aspect ratio of the smallest bounding box is greater than a preset geometric threshold to obtain a candidate set of geometric shapes includes: The maximum extension direction of the point cloud cluster projected onto the horizontal plane is calculated as the principal direction, and the direction orthogonal to this principal direction is calculated as the secondary direction. The ratio of the length in the main direction to the length in the secondary direction is used as the basis for determining the aspect ratio. Only point cloud clusters that meet the preset geometric constraints in terms of aspect ratio, absolute length, and absolute height are retained as real flow guiding facilities.

4. The method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition according to claim 3, characterized in that, The process of step S3 includes: Perform ISS keypoint detection on the target dataset of the current frame, calculate the local shape feature descriptor of each keypoint, and obtain the feature set of the current frame; ISS keypoint detection and feature descriptor calculation are also performed on the target dataset of the previous frame to obtain the reference frame feature set; Feature matching is performed based on the current frame feature set and the reference frame feature set. Reliable matching point pairs are selected by the nearest neighbor distance ratio method, and coarse registration is completed by rigid transformation matrix estimation. Based on coarse registration, an iterative optimization algorithm is used for fine registration. Point-to-point distance constraints and point-to-area distance constraints are introduced into the iterative objective function. Fine registration is achieved by minimizing the comprehensive distance error between the source point cloud and the target point cloud, resulting in a time-aligned dataset.

5. The method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition according to claim 4, characterized in that, The process of fine registration based on coarse registration using an iterative optimization algorithm includes: In each iteration, the weight coefficient of each matching point pair is dynamically calculated. This weight coefficient is inversely proportional to the spatial distance between the point pairs and directly proportional to the cosine of the angle between the normal vectors. Matching point pairs with large distances or significant differences in normal vectors are assigned lower weights. The optimal rigid body transformation matrix is ​​solved by the weighted least squares optimization method, so that the registration process is robust to local occlusion and noise interference.

6. The method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition according to claim 5, characterized in that, The process of step S4 includes: Based on the time-series aligned dataset, spherical neighborhood search structures are established for the reference point cloud and the target point cloud respectively. The rate of change of the number of neighboring points of each sampling point in the spherical neighborhood is calculated to obtain the local density change characteristics. For each point in the temporally aligned dataset, calculate the covariance matrix of its neighboring points, obtain the principal curvature components through eigenvalue decomposition and calculate the average curvature, and obtain the absolute value of the curvature difference between the corresponding points of the reference point cloud and the target point cloud to obtain the surface curvature difference feature. For each point in the time-aligned dataset, calculate the covariance matrix of its neighboring point set, extract the eigenvector corresponding to the minimum eigenvalue as the normal vector, and statistically analyze the consistency of the normal vectors of the reference point cloud and the target point cloud in their corresponding neighborhoods to obtain the divergence feature of the normal vector field. The local density variation features, surface curvature difference features, and normal vector field divergence features are normalized to eliminate the differences in the dimensions of each feature, resulting in a normalized feature vector set. Based on the type attribute of the diversion facility, a weight coefficient is assigned to each normalized feature, and the comprehensive change index value is calculated by weighted summation. Combined with the preset risk threshold, an initial monitoring strategy is formed.

7. The method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition according to claim 6, characterized in that, The process of assigning weight coefficients to each normalized feature based on the diversion facility type attribute includes: A correlation model between the physical properties of the diversion facility and its risk-sensitive characteristics is pre-established. Based on the material stiffness, geometric dimensions, and installation method of the current monitoring target, the basic weight values ​​of each normalized characteristic are calculated. Among them, the basic weight value of the curvature difference characteristic of rigid diversion facilities is higher than that of density change characteristics and normal vector divergence characteristics, while the basic weight value of the normal vector divergence characteristic of flexible diversion facilities is higher than that of curvature difference characteristics and density change characteristics. The basic weight values ​​are dynamically adjusted based on the real-time correction factors of vibration intensity and traffic flow at the construction site to obtain adaptive weight coefficients.

8. The method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition according to claim 7, characterized in that, The process of step S5 includes: Collect the comprehensive change index value sequence and the corresponding false alarm label sequence generated within the historical monitoring period to establish a historical monitoring database; Statistical analysis is performed on the historical monitoring database to identify the characteristic distribution patterns under high-frequency false alarm conditions, calculate the dispersion contribution rate of each geometric feature in the false alarm samples, and obtain a false alarm feature analysis report. Based on the false alarm feature analysis report, the gradient descent method is used to iteratively optimize the feature fusion weight coefficients, so that the comprehensive change index value under false alarm conditions converges to the low-risk range, and a dynamic weight adjustment strategy is obtained. Probability density estimation is performed on the comprehensive change index values ​​in the historical monitoring database, a nonlinear mapping function for risk level is constructed, and the early warning threshold and alarm threshold are dynamically adjusted according to the safety level requirements of the current construction stage to form an adaptive threshold configuration scheme. The dynamic weight adjustment strategy and adaptive threshold configuration scheme are applied to the current monitoring data processing flow to obtain an optimized monitoring strategy. When the real-time calculated comprehensive change index value exceeds the adaptive threshold configuration scheme, a graded early warning is triggered and relevant personnel are notified through the communication module.

9. The method for monitoring diversion facilities using multi-source sensing and dynamic feature recognition according to claim 8, characterized in that, The process of estimating the probability density of comprehensive change index values ​​in the historical monitoring database, constructing a nonlinear mapping function for risk levels, and dynamically adjusting the early warning threshold and alarm threshold according to the safety level requirements of the current construction stage to form an adaptive threshold configuration scheme includes: A nonparametric probability distribution model is constructed using the kernel density estimation method, and the risk quantile is calculated as the initial threshold based on the distribution function. The initial threshold is bidirectionally corrected by the false alarm rate and missed detection rate fed back from the construction site. When the false alarm rate exceeds the preset tolerance, the warning threshold is increased, and when the missed detection rate exceeds the preset tolerance, the alarm threshold is decreased, forming an adaptive threshold configuration scheme with hysteresis characteristics.

10. A multi-source sensing and dynamic feature recognition system for monitoring diversion facilities, as described in any one of claims 1-9, characterized in that, include: The multi-source sensing module is used to synchronously scan and image the diversion facilities at the construction site through a cross-modal sensing system composed of lidar and visible light camera, and acquire multi-source synchronous data. The target recognition module is connected to the multi-source sensing module to preprocess the multi-source synchronous dataset and identify the monitoring target dataset based on the target features. The spatiotemporal registration module is linked to the target recognition module and is used to register the monitoring target datasets of adjacent monitoring frames to obtain cross-frame spatial registration results. The strategy generation module is connected to the spatiotemporal registration module and is used to calculate the local point cloud density change rate, surface average curvature difference and local normal vector field divergence based on the cross-frame spatial registration results. It generates a comprehensive change index value through weighted fusion and generates an initial monitoring strategy based on a preset risk threshold. The strategy optimization module, connected to the strategy generation module, is used to adaptively optimize the initial monitoring strategy based on historical monitoring data and feedback information to obtain an optimized monitoring strategy. When the changing characteristics exceed the optimized monitoring threshold, an early warning is triggered and relevant personnel are notified.

Images