A method for automatically identifying road surface cracks based on three-dimensional point cloud data

Abstract

This invention belongs to the field of computer vision and road surface inspection, specifically relating to an automatic road surface crack identification method based on 3D point cloud data. The method preprocesses the original 3D road point cloud by denoising and coordinate transformation, and constructs a spatial partition index using adaptive voxel mesh partitioning. It calculates multi-dimensional features such as the normal vector, curvature, and reflection intensity of sampling points to characterize road surface undulations. Subsequently, the feature-laden point cloud is input into a deep learning model, and candidate crack points are identified through multi-scale feature aggregation. Then, density space clustering and skeleton extraction algorithms are used to perform crack topology reconstruction, and logical connections are made to the fractured segments based on geometric continuity constraints. Finally, quantitative evaluation of the damage is achieved, and a vector map with mapping accuracy is output. This invention solves the problems of low efficiency and poor recognition accuracy in processing massive point clouds, achieving accurate crack identification and complete reconstruction, and improving the automation and digitization level of road surface defect detection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and road surface detection, specifically relating to an automatic identification method for road surface cracks based on three-dimensional point cloud data. Background Technology

[0002] With the continuous evolution of digital transportation infrastructure and smart highway maintenance technologies, acquiring three-dimensional road surface information using vehicle-mounted LiDAR inspection systems has become an important means of detecting road defects. Three-dimensional point cloud data, with its high-precision geometric representation capabilities and rich spatial dimensional information, provides reliable data support for the refined assessment of road surface conditions. In the field of highway health monitoring, the automated identification and mapping of road surface defects such as cracks and potholes has extremely high application value for ensuring driving safety and optimizing maintenance decisions.

[0003] Deep learning-based point cloud data processing technology is a core approach for achieving automatic road surface defect identification. It aims to construct a mapping relationship between road surface geometry and defect attributes by performing feature analysis and semantic segmentation on large-scale point cloud sequences. This technology typically relies on vehicle-mounted mobile measurement systems to collect high-density point clouds of the road surface, using specific algorithms to extract subtle surface undulations and perform localization and quantitative analysis of road damage.

[0004] The massive amounts of 3D point cloud data collected by vehicle-mounted LiDAR limit data processing efficiency. Directly performing global traversal and complex feature calculations on the raw point cloud incurs huge computational overhead and memory consumption, making it difficult to meet the timeliness requirements of rapid inspection in daily highway maintenance. Meanwhile, real-world road environments are plagued by noise interference from tree shadows, water stains, manhole covers, and worn road markings. Traditional segmentation algorithms based on single elevation thresholds or curvature exhibit poor robustness in complex backgrounds, easily leading to false positives and false negatives. Furthermore, the morphological characterization of fine cracks is affected by point cloud sparsity and occlusion; existing methods struggle to maintain the continuity of the disease's topological structure, resulting in identified cracks appearing discrete and isolated, failing to accurately extract the complete crack orientation and network distribution.

[0005] Therefore, there is a need for an automatic road surface crack identification method based on 3D point cloud data. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic road surface crack identification method based on three-dimensional point cloud data, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: an automatic identification method for road surface cracks based on three-dimensional point cloud data, comprising the following specific steps: Step 1: Acquire the original 3D point cloud data of the road and perform spatial preprocessing. Simultaneously collect the 3D coordinate information, reflection intensity information and absolute geographical location information of the road surface by the LiDAR scanning system and the integrated inertial navigation system. Then, denoise, resample and uniformly transform the coordinate system of the acquired discrete point cloud sequence to generate a standardized road surface point cloud dataset. Step 2: Construct a spatial partitioning index structure. Based on the standardized road surface point cloud dataset, use an adaptive voxel mesh partitioning method to divide the massive point cloud into spatial processing units of a preset size, and establish a topological association index for each spatial processing unit to achieve fast retrieval and parallel computing scheduling of large-scale point cloud data. Step 3: Extract multidimensional geometric features. For the point cloud in each spatial processing unit, calculate the local normal vector, surface curvature, information entropy, and reflection intensity distribution features of each sampling point. Construct a high-dimensional feature vector for each sampling point through a feature fusion algorithm to characterize road surface smoothness and abnormal undulations. Step 4: Construct and train a deep learning semantic segmentation model. Input the spatial processing unit carrying multi-dimensional geometric features into the preset deep neural network. The deep neural network performs classification label prediction on each sampling point in the point cloud through the multi-scale local feature aggregation module and the global semantic encoding module, and identifies the objects belonging to the crack candidate point cloud. Step 5: Perform crack topology reconstruction and noise removal. Perform density space-based clustering analysis on the crack candidate point cloud, obtain the central axis of the crack through the skeleton extraction algorithm, and logically connect the broken crack segments based on the preset geometric continuity constraints to eliminate discontinuity problems caused by occlusion or noise. Step 6: Quantitative evaluation and mapping output of road defects. Based on the reconstructed crack topology, calculate the crack length, average width, maximum depth, direction angle and affected area, and generate a road defect vector map and technical condition assessment report with mapping accuracy by combining absolute geographical location information.

[0008] Preferably, the process of acquiring the original 3D point cloud data of the road in step 1 is achieved through a multi-beam lidar sensor integrated into the mobile inspection platform. The sensor emits laser pulses of a preset frequency and receives the echo signals reflected from the road surface. By measuring the time difference between the round trip of the laser pulse and combining it with the beam angle at the moment of sensor emission, the relative spatial coordinates of each reflection point relative to the sensor center are calculated. Simultaneously, the integrated inertial navigation system records the inspection platform's attitude angle, acceleration, and GPS positioning coordinates in real time. Through timestamp synchronization technology, the relative spatial coordinates are converted into a unified geospatial coordinate system.

[0009] Preferably, the spatial preprocessing in step 1 includes statistical filtering. This statistical filtering calculates the average distance between each sampling point in the point cloud dataset and a predetermined number of neighboring points within its neighborhood, thus statistically distributing the average distance across all points. Sampling points whose average distance falls outside a predetermined standard deviation are identified as outliers and removed. Furthermore, the process includes reflection intensity-based filtering. By analyzing the reflectivity differences between road surface materials and common interfering elements such as water stains, road markings, and manhole covers, a reflection intensity filtering threshold is set to suppress the impact of specific noise on subsequent crack identification.

[0010] Preferably, in step 2, the process of constructing the spatial partitioning index structure first involves establishing a three-dimensional bounding box based on the total length and width of the road segment to be processed. The adaptive voxel mesh partitioning method dynamically adjusts the side length of the voxel unit according to local changes in point cloud density. Larger voxel sizes are used in sparse point cloud regions, while smaller voxel sizes are used in dense point cloud regions. To improve query efficiency, an octree data structure is used to recursively partition the voxel unit, with each non-leaf node containing 8 child nodes, until a preset termination condition is met, such as the number of point clouds in each leaf node being less than a preset threshold or the partitioning depth reaching a preset level.

[0011] Preferably, the process of calculating the local normal vector of each sampling point in step 3 employs the least squares plane fitting method. A covariance matrix is ​​constructed for the target point and all its neighboring points within a preset search radius. This covariance matrix reflects the spatial distribution characteristics of the neighborhood point set. By performing eigenvalue decomposition on the covariance matrix, three eigenvalues ​​and their corresponding eigenvectors are obtained. The eigenvector corresponding to the smallest eigenvalue is the local normal vector of the target point. The surface curvature is characterized by the ratio of the smallest eigenvalue to the sum of the three eigenvalues. The larger this ratio, the more severe the surface undulations in the region, and the more closely it conforms to the geometric characteristics of cracks or pits.

[0012] Preferably, the process of constructing the high-dimensional feature vector in step 3 also takes into account the local variance of the reflection intensity and the rate of change of the elevation gradient. By constructing a multi-dimensional vector, geometric position information, normal vector components, curvature values, and normalized reflection intensity information are concatenated. To enhance the rotation invariance of the features, the relative coordinates in the local coordinate system are transformed to ensure that the same crack feature has a consistent vector representation under different scanning perspectives.

[0013] Preferably, the deep learning semantic segmentation model in step 4 employs a convolutional neural network based on direct point cloud processing. This network structure includes multiple feature downsampling layers and upsampling layers. During downsampling, a point sampling strategy is used to reduce the data size, and a local feature extractor captures the spatial geometric relationships between points. The local feature extractor simulates the feature extraction process of traditional image convolution by performing a weighted summation operation within the neighborhood of each sampling point. To preserve the detailed information of subtle cracks, skip connections are established between the corresponding downsampling and upsampling layers, directly transferring high-resolution features from the shallow layers to the deeper layers, achieving multi-scale feature fusion.

[0014] Preferably, the training process of the deep neural network in step 4 uses a sample library containing a large number of manually labeled cracks. The loss function consists of two parts: cross-entropy loss and geometric consistency loss. The cross-entropy loss is used to measure the deviation between the predicted class and the true label; the geometric consistency loss is used to constrain the spatial continuity of the prediction results and avoid isolated misclassification points. The network weights are continuously updated using the stochastic gradient descent algorithm so that the loss function value reaches a preset convergence range.

[0015] Preferably, in step 5, the density-space-based clustering analysis groups candidate point clouds belonging to the same crack entity into one class by setting a core neighborhood radius and a minimum number of members. For the classified point clusters, if the total number of points in the point cloud is less than a preset size threshold, they are considered isolated noise and discarded. The skeleton extraction algorithm calculates the local centroid of each point cluster and iteratively searches along the principal component direction of the point cluster to generate a series of ordered skeleton points to represent the morphological midline of the crack.

[0016] Preferably, in step 5, the process of logically connecting the fractured crack segments based on preset geometric continuity constraints first calculates the Euclidean distance between the end skeleton points of different crack segments. If this distance is less than a preset connection threshold, the angle between the two segments in the tangent direction at their ends is further calculated. When both the distance and the angle meet preset conditions, a spline curve interpolation algorithm is used to generate connection segments, combining the discrete segments into a complete crack line string. For intersecting cracks, node connection logic in graph theory is used to construct a network topology model of the pavement cracks.

[0017] Preferably, the process of calculating the crack width in step 6 is achieved by searching for crack edge points within the normal plane of the skeleton points. For each skeleton point, the abrupt change in road surface elevation on the left and right sides is calculated within its cross-section perpendicular to the crack direction, and the horizontal distance between these abrupt change points is the crack width at that location. The average width of the crack is obtained by weighted averaging the width values ​​at all skeleton point locations along the entire crack.

[0018] Preferably, in step 6, the quantitative evaluation of road defects is conducted according to the National Road Maintenance Technical Specifications. Based on the detected crack length and width, a defect severity index is automatically calculated. If the crack width exceeds a preset first threshold, or the total length of cracks within a preset area exceeds a preset second threshold, it is marked as a high-risk defect. All identified cracks are associated with their high-precision coordinates in a global coordinate system, which can be directly imported into a geographic information system for visualization.

[0019] Preferably, the method further includes a real-time processing mode. During the inspection vehicle's operation, the collected point cloud stream is processed in real-time through an onboard embedded computing unit. A sliding window mechanism is employed to quickly extract and pre-identify features only within the current window, and the identification results are uploaded to a cloud server via a wireless communication network for refined reconstruction and archiving. In this real-time processing mode, by reducing sampling density and simplifying the number of network layers, the data processing speed is ensured to match the vehicle's speed, enabling immediate detection of road defects.

[0020] Preferably, the method also involves a multi-source data fusion verification mechanism. During crack identification, two-dimensional road surface images captured by a high-definition industrial camera mounted on the inspection vehicle are simultaneously retrieved. The three-dimensional point cloud is projected onto the two-dimensional image plane using exterior orientation elements, achieving precise alignment between three-dimensional geometric information and two-dimensional texture information. Color and brightness features in the image are used to assist in correcting false detections in point cloud recognition, especially when dealing with easily confused targets such as manhole cover edges and worn markings; visual feature verification improves the accuracy of recognition.

[0021] Preferably, the standardized road surface point cloud dataset also includes automatic correction for the road's cross slope and longitudinal slope. A road surface reference plane is fitted by extracting point clouds within a preset range on both sides of the road centerline. The elevation values ​​of all sampled points are subtracted from the corresponding elevation of this reference plane to obtain normalized road surface relative undulation data. This processing eliminates the influence of the road's own slope on crack depth extraction, enabling accurate capture of subtle pothole and crack geometric features even on road sections with large slopes.

[0022] Preferably, the deep learning semantic segmentation model further includes an adaptive attention mechanism module. This module automatically focuses on geometric features that contribute significantly to crack identification, such as elevation change features, by calculating the correlation weights between feature channels. By weighting and enhancing key features, background clutter caused by road surface roughness is suppressed, maintaining stable detection accuracy on roads with different paving materials.

[0023] Compared with the prior art, the present invention has the following beneficial effects: 1. By constructing an adaptive voxel mesh index and spatial partitioning processing mechanism, this invention effectively solves the problem of low processing efficiency for massive 3D point cloud data. It decomposes tens of millions of original point clouds into small, parallel-processable units, reducing computational complexity and peak memory usage. This makes rapid processing and real-time monitoring of large-scale road inspection data possible, improving the operational efficiency of highway maintenance.

[0024] 2. By utilizing deep neural networks to extract multidimensional geometric and reflection features, this invention significantly enhances the recognition accuracy in complex road surface environments. By fusing normal vectors, curvature, intensity, and spatial semantic information, it can effectively distinguish cracks from non-defective interference objects such as tree shadows, water stains, and road markings, overcoming the high false detection rate of traditional single-threshold segmentation methods and possessing strong environmental adaptability and robustness.

[0025] 3. By employing a topology reconstruction algorithm based on density clustering and geometric continuity constraints, this invention solves the problem of fracture and incompleteness in the identification of minute cracks. Through logical connection and skeleton extraction of discrete candidate points, the complete direction and complex network structure of cracks can be accurately reconstructed, providing reliable geometric topology support for crack length measurement, direction analysis, and accurate assessment of disease severity.

[0026] 4. This invention achieves full automation of the road defect process, from automatic identification to high-precision mapping. By fusing the identification results with high-precision geographic location information, the generated defect vector map can be directly used for navigation, positioning, and engineering maintenance decisions, reducing manual intervention costs, improving the digital management level of transportation infrastructure, and yielding both economic and social benefits. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram illustrating the core principle framework for crack candidate point identification based on a deep learning semantic segmentation model according to the present invention. Figure 3 This is a flowchart illustrating the logical process of constructing a spatial partitioning index structure based on adaptive voxel mesh partitioning and octree decomposition according to the present invention. Figure 4 This is a flowchart illustrating the logical flow of the crack candidate point cloud clustering analysis, skeleton extraction, and geometric continuity topology reconstruction according to the present invention. Figure 5 This is a schematic diagram of the multi-source data fusion verification data flow of three-dimensional point cloud geometric information and two-dimensional image texture information according to the present invention. Detailed Implementation

[0028] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0029] In implementing the automatic road surface crack identification method based on three-dimensional point cloud data described in this invention, the specific implementation process of step 1 involves multi-source acquisition and high-precision preprocessing of raw data. Specifically, step 1 involves the coordinated operation of a multi-beam lidar sensor, a combined inertial navigation system, and a central clock synchronization controller integrated on a mobile inspection platform. The multi-beam lidar sensor emits high-frequency laser pulses onto the road surface at a preset scanning frequency, for example, 20 to 100 times per second. Each laser pulse is reflected upon contact with the road surface, and the sensor receives the echo signal and records the time difference between the laser's round trip and the instantaneous horizontal and vertical angles of the emitted pulse. By logically calculating half the product of the speed of light and the time difference, the polar coordinate information of each sampling point relative to the optical center of the sensor is obtained.

[0030] In step 1 above, the integrated inertial navigation system includes a GPS receiver, a three-axis gyroscope, and a three-axis accelerometer. The GPS receiver acquires geographic coordinates at a first frequency (e.g., 1 Hz to 5 Hz); the three-axis gyroscope and three-axis accelerometer record the attitude angle, angular velocity, and instantaneous acceleration of the inspection platform in real time at a second frequency (e.g., 100 Hz to 200 Hz). The multi-sensor data is fused using a Kalman filter algorithm to calculate the six-degree-of-freedom pose information of the inspection platform at each moment. To ensure the accuracy of the spatial coordinates, the central clock synchronization controller uses a second pulse signal to align the time axis of the lidar and the integrated inertial navigation system at the microsecond level. Based on the six-degree-of-freedom pose information, the relative spatial coordinates are transformed to a unified geographic spatial coordinate system, such as a geocentric coordinate system or a projected plane coordinate system, by constructing a textual logic for the rotation and translation matrix, forming a discrete original point cloud sequence.

[0031] In the spatial preprocessing of step 1, statistical filtering is first performed to eliminate environmental shot noise. Specifically, for each sampling point in the point cloud dataset, a spatial neighborhood search algorithm is used to find 30 to 50 neighboring points within a preset radius. The average Euclidean distance between the sampling point and all neighboring points is calculated. After traversing all sampling points along the entire road segment, the global average and standard deviation of the average distances to all points are calculated. A distance threshold is set, which is equal to the global average plus 2 to 3 times the standard deviation. For any sampling point, if its average Euclidean distance is greater than the distance threshold, the point is determined to be an isolated outlier and is removed from the point cloud dataset.

[0032] Subsequently, in step 1, filtering based on reflection intensity is performed. Since road surface materials such as asphalt or concrete have significantly different reflectivities than water stains, traffic markings, and metal manhole covers in the laser band, the system analyzes the histogram distribution of reflection intensity information and sets a reflection intensity filtering threshold. For example, points with reflection intensity values ​​in a preset low range (corresponding to water stains) or a preset high range (corresponding to markings or metal) are marked as interference candidate points. Combining local elevation variance, if the elevation consistency between the interference candidate point and surrounding road surface points is extremely high, it is temporarily masked from the crack extraction set to reduce the probability of false detection in subsequent identification processes.

[0033] Furthermore, step 1 also includes automatic correction of the road's cross and longitudinal slopes. This process extracts point cloud data within a preset width range on both sides of the road centerline and constructs a road reference plane using the least squares plane fitting method. The normal vector of the reference plane and its slope parameters in the geographic coordinate system are calculated. The absolute geographic elevation of all sampling points is subtracted from their corresponding elevation projected onto the reference plane to obtain normalized road surface relative undulation data. This processing eliminates the interference of the road's structured slope on the identification of minute crack depths, ensuring that subsequent feature extraction can focus on microscopic geometric variations.

[0034] After completing the above preprocessing, step 2 is executed: constructing a spatial partitioning index structure. First, based on the total length, width, and relative elevation range of the road segment to be processed, an axis-aligned 3D bounding box is established. This bounding box completely contains the standardized road surface point cloud dataset. Based on the standardized road surface point cloud dataset, an adaptive voxel meshing method is used for spatial partitioning. In the initial stage, the 3D bounding box is divided into initial voxel units with side lengths of 1 to 2 meters. Subsequently, for each initial voxel unit, the local density distribution and curvature distribution of the internal point cloud are calculated. If the point cloud density within the voxel is greater than a preset density threshold, or if the local curvature of a sampling point is greater than a preset curvature sensitivity threshold, then the recursive partitioning logic is initiated.

[0035] Specifically, the adaptive voxel meshing method employs an octree data structure. Each non-leaf node voxel is uniformly divided into 8 sub-voxel units. This partitioning process continues until one of the following termination conditions is met: the side length of the voxel unit reaches a preset minimum precision size (e.g., 0.01 meters); the number of point clouds within the voxel is less than a preset minimum value (e.g., 10 points); or the partitioning depth reaches a preset maximum level. In this way, larger voxels are used in areas with flat road surfaces and sparse point clouds, while extremely small voxels are used in areas with rich details and dense point clouds, such as crack edges and potholes. A unique global index number is assigned to each final generated spatial processing unit, and the topological associations of its neighboring voxels are recorded, constructing a topological index model that supports fast spatial neighborhood queries and large-scale parallel computing scheduling.

[0036] Next, step 3 is executed: extracting multidimensional geometric features. For each sampling point within each spatial processing unit, its local normal vector is calculated. Specifically, this is achieved by constructing a local covariance matrix centered on the target point. The covariance matrix is ​​a 3×3 symmetric matrix obtained by centering the coordinate components of the target point and all its neighboring points within a preset search radius, and then calculating the sum of the outer products of the coordinate deviation vectors. By performing eigenvalue decomposition on this covariance matrix, three eigenvalues ​​and their corresponding three eigenvectors are obtained in descending order. The eigenvector corresponding to the smallest eigenvalue is used as the local normal vector of the sampling point, representing the orientation of the local plane.

[0037] In step 3 above, the surface curvature is quantitatively characterized by the ratio of the minimum eigenvalue to the sum of the three eigenvalues. When this ratio is close to zero, it indicates that the neighboring points are highly coplanar and the road surface is smooth; when the ratio increases, it indicates that there are significant elevation jumps or normal discontinuities in the local area, consistent with the geometric characteristics of cracks or potholes. Furthermore, the system calculates the information entropy feature for each point. Information entropy is measured based on the probability distribution of the normal vector's direction within the neighborhood: if the normal vector has a unidirectional direction, the information entropy value is low; if the normal vector's direction is randomly distributed, the information entropy value increases, used to distinguish between regular mechanical joints and irregular natural cracks.

[0038] In step 3, during the construction of the high-dimensional feature vector, the local variance of the reflection intensity and the rate of change of the elevation gradient are also taken into consideration. Specifically, the three-dimensional spatial coordinate components of the sampling point, the three components of the normal vector, the surface curvature value, the information entropy value, the normalized reflection intensity value, and the standard deviation of the reflection intensity in the neighborhood are concatenated to form a multi-dimensional feature vector. To enhance the rotation invariance of the features, the coordinates of the neighborhood points are transformed to a local coordinate system with the local normal vector as the reference axis before calculating the features. This high-dimensional feature vector, as a joint expression of the geometric and physical properties of the sampling point, can enhance the model's ability to capture abnormal undulations in road surface smoothness.

[0039] Then, step 4 is performed: constructing and training a deep learning semantic segmentation model. This deep learning semantic segmentation model employs a convolutional neural network architecture based on direct point cloud processing, with its core consisting of a multi-scale local feature aggregation module and a global semantic encoding module. During the data input stage, spatial processing units carrying high-dimensional feature vectors are used as the raw input. The multi-scale local feature aggregation module obtains the local structure of the point cloud under different receptive fields by setting spherical search spaces of different radii. For each scale of receptive field, a weighted summation operation is used to simulate a convolution operation, with the weights automatically learned through a multilayer perceptron with shared parameters, capturing hierarchical features from microscopic crack edges to macroscopic road surface morphology.

[0040] In step 4, the deep neural network includes multiple downsampling layers. Each layer retains the most representative subset of points using a farthest-point sampling strategy to reduce computation and increase the abstraction of features. Correspondingly, the network also includes an equal number of upsampling layers, which use a distance-weighted interpolation algorithm to restore high-level semantic features to the original point cloud resolution. To prevent the loss of subtle crack features during deep convolution, a skip connection is established between the corresponding downsampling and upsampling layers, directly transferring lower-level geometric details across layers and concatenating them with higher-level semantic features along the feature dimension.

[0041] Furthermore, the deep learning semantic segmentation model also includes an adaptive attention mechanism module. This module learns the importance weight coefficients of different feature channels by calculating the covariance relationship between feature channels. By applying these weight coefficients to high-dimensional features, the model can automatically enhance features that contribute significantly to crack identification (such as abrupt changes in the Z-component of the normal vector) and suppress background clutter caused by road surface roughness and particle protrusions. For each sampling point, the network ultimately outputs a probability vector, representing the probability distribution of whether the point belongs to the background, crack, pothole, or marking. The category with the highest probability value is selected as the predicted label for that point, thus identifying the crack candidate point cloud.

[0042] In the training phase of step 4, a large-scale manually labeled road defect dataset is used. The loss function used in the training process consists of two parts: the first part is a weighted cross-entropy loss, which is used to minimize the statistical deviation between the predicted category and the manually labeled label, and balances the positive and negative sample distribution by increasing the weight of the crack category; the second part is a geometric consistency loss, which is constrained based on the spatial smoothness of the predicted label. If the predicted label of a point is inconsistent with the labels of most points in its neighborhood, a penalty term is added to force the model to learn a recognition result with spatial continuity and avoid sporadic misclassifications.

[0043] Step 5: Perform crack topology reconstruction and noise removal. Since the crack candidate point cloud output by deep learning may contain scattered noise points or local topological breaks, it is first subjected to density-space-based clustering analysis. Specifically, a core neighborhood radius and a minimum point count threshold are set. For any candidate point, if the number of candidate points within its core neighborhood radius is greater than the minimum point count threshold, it is marked as a core point. By grouping mutually reachable core points into the same cluster, multiple independent crack point clusters are formed. Small point clusters with a total point cloud count less than a preset size threshold (e.g., 50 points) are identified as interference noise (e.g., road debris residue) misjudged by the algorithm and removed.

[0044] In step 5, a skeleton extraction algorithm is performed on each filtered crack point cluster. First, the geometric centroid of the point cluster is determined, and the principal component directions of its covariance matrix are calculated. A sliding window is moved along the principal component directions with a preset step size. Within each window, the local centroid of the point cloud in a plane perpendicular to the principal direction is calculated. By connecting this series of local centroids, an initial skeleton point sequence describing the crack center orientation is generated. Subsequently, logical connections are made to fracture fragments caused by occlusion or scan sparsity based on preset geometric continuity constraints.

[0045] Specifically, the three-dimensional Euclidean distance between the end skeleton points of different crack segments is calculated. If this distance is less than a set connection step size threshold (e.g., 0.5 meters), the cosine of the angle between the tangent direction vectors at the ends of the two segments at the fracture is further calculated. When the Euclidean distance is less than the threshold and the cosine of the angle is greater than a preset direction consistency threshold (e.g., 0.85), the two segments are determined to belong to the same physical crack. A spline curve interpolation algorithm or a Bézier curve completion algorithm is used to generate smooth connection segments in the fracture gap, combining discrete crack segments into a crack line string with a complete geometric topology. For the network structure formed by the intersection of multiple cracks, the node connection logic in graph theory is used, defining the intersection points as nodes of the graph and the crack skeleton segments as edges, to construct a network topological relationship model of the pavement cracks.

[0046] Finally, step 6 is executed: quantitative evaluation and mapping output of the defects. Based on the reconstructed complete crack topology, various geometric defect indicators are automatically extracted. For crack width calculation, a search is performed in the normal plane of each skeleton point along a direction perpendicular to the crack direction. Along the search path, abrupt changes in pavement elevation distribution or step changes in reflection intensity are identified; the horizontal distance between these two locations is defined as the instantaneous crack width at that skeleton point. The average crack width is obtained by weighted averaging of the width values ​​at all skeleton points along the entire crack; simultaneously, the maximum value during the search process is recorded as the maximum width.

[0047] In step 6, the crack depth is obtained by calculating the vertical sinking distance of the skeleton point relative to the surrounding smooth road surface reference plane. The total crack length is obtained by summing the Euclidean distances between all adjacent skeleton points. The affected area is obtained by multiplying the crack length by the average width, or by calculating the area of ​​the bounding polygon of the crack candidate point cloud on the horizontal projection plane. Combining the absolute geographic location information obtained in step 1, the identified crack skeleton lines are converted into vector geographic data (such as Shapefile format) with mapping accuracy.

[0048] According to the national road maintenance technical specifications, this invention automatically calculates the road damage severity index in step 6. If the crack width exceeds a first preset threshold (e.g., 3 mm), or the total length of cracks within a preset 1 square meter sampling area exceeds a second preset threshold (e.g., 1 meter), the system marks the area as a high-risk damage zone and automatically labels its absolute coordinates, azimuth, and recommended repair level in the generated assessment report. Finally, the system generates a digital vector map containing all identified damages and a technical condition assessment report, which can be directly imported into a geographic information system for maintenance decision analysis.

[0049] In practical engineering applications, this invention also includes a real-time processing mode. In this mode, the onboard embedded computing unit on the inspection vehicle employs a sliding window mechanism. As the vehicle moves, the LiDAR continuously generates point cloud streams. The computing unit performs rapid spatial preprocessing and downsampling feature extraction only on the point cloud slices within the current time window. Suspected defect areas are initially screened through a lightweight pre-identification network, and the point cloud data of these key areas are uploaded to a cloud server via a wireless communication network. The cloud server performs high-precision deep semantic segmentation and topology reconstruction, achieving a balance between real-time inspection and delayed refined analysis.

[0050] Furthermore, this invention also relates to a multi-source data fusion verification mechanism. A high-definition industrial camera is synchronously installed on a mobile inspection platform. During crack identification, the camera's interior and exterior orientation elements are used to establish a projection mapping relationship between the three-dimensional point cloud space and the two-dimensional image plane. For crack candidate areas identified by the point cloud, the corresponding two-dimensional image is automatically retrieved, and the grayscale gradient, texture features, and edge contrast in the image are extracted. The advantages of two-dimensional images in high-resolution texture are utilized to assist in correcting potential false detections in point cloud identification. For example, when point cloud data becomes ambiguous when processing the edges of severely worn traffic markings, the color distribution features in the image are used to verify whether it is a manual marking, further improving the accuracy and reliability of the identification.

[0051] Example 2: In Example 2, the present invention provides an enhanced recognition scheme for extremely complex road surface environments (such as road sections in old urban areas with a large number of manhole covers, drainage ditch covers, and repair marks). The basic steps of this example are consistent with those of Example 1, but targeted enhancements have been made at the feature extraction and model decision levels.

[0052] Specifically, when performing step 3 to extract multidimensional geometric features, local anisotropy features are introduced for the sampling points. These features are derived from the three eigenvalues ​​of the covariance matrix (…). , , Further textual logical operations are performed.

[0053] The linearity index is expressed by the following formula: ;in The first eigenvalue, It is the second eigenvalue; The formula for the flatness index is: ;in, It is the third eigenvalue; The formula for the scattering index is: .

[0054] For crack points, the linearity index is usually much greater than the flatness and scattering indices; while for manhole covers or smooth road surfaces, the flatness index dominates. By incorporating these three second-order structural features into the high-dimensional feature vector, it is possible to more accurately distinguish between linear defects and planar man-made structures at the geometric level.

[0055] In step 4, to address complex background interference, the deep learning model introduces a multi-task learning framework. While outputting classification labels for sampling points, the network simultaneously outputs regression predictions of local surface normal vectors. During training, a consistency loss term for normal vector prediction is added to the loss function. This loss term calculates the angle difference between the network-predicted normal vector and the geometric normal vector obtained in step 3 based on least-squares fitting. For real crack areas, due to their dramatic geometric undulations, the predicted normal vector and the geometric normal vector should maintain a high degree of local correlation. This multi-task constraint mechanism forces the network to understand the physical characteristics of the road surface micromorphology more deeply, suppressing interference from non-destructive targets with regular geometric contours, such as manhole cover edges.

[0056] In step 5 of Example 2, the crack topology reconstruction further considers the growth dynamics of the cracks. For complex network cracks (crazing), a graph-based shortest path search algorithm is employed. Each skeleton point is considered a vertex of the graph, and the product of the Euclidean distance between points and the consistency weight of the local normal vector is used as the edge weight. When connecting cracks, the system seeks the path connection scheme that minimizes the sum of the global weights. This method can handle the situation of multiple interwoven micro-cracks, accurately reconstruct the stress-damage network of the pavement structure, and provide richer data support for damage assessment at the road structure level.

[0057] In step 6, the pavement damage assessment, this embodiment adds a quantitative estimate of crack volume changes. The total volume of the crack cavity is calculated by integrating the reconstructed three-dimensional solid model of the crack. Compared to simple length and width, this indicator more accurately reflects the loss of pavement material, especially in the early stages of pothole formation, where volume monitoring allows for more precise early warning.

[0058] Example 3: In Example 3, the present invention describes an implementation scheme for a large-scale urban road defect survey based on cloud collaboration. This scheme focuses on optimizing the spatial partitioning and data scheduling logic in step 2.

[0059] Specifically, in step 2, to handle road point cloud data spanning thousands of kilometers across the city, the system constructs a global spatial partitioning index based on Geographic Information System (GIS) tiling logic. First, the urban road network is divided into geographic grid units of a preset area (e.g., 500 meters by 500 meters). Each grid unit acts as an independent data processing container, internally undergoing adaptive voxel grid partitioning. To achieve topological continuity of cracks across grids, a 5% to 10% overlap buffer is reserved between grids. During topological connection in step 5, the system automatically identifies and merges homogeneous crack segments located within the buffer, ensuring that the generated road defect vector map has strict logical consistency across the entire city.

[0060] Regarding model deployment in step 4, this embodiment employs a hybrid architecture combining a lightweight convolutional network and a densely connected network. On the in-vehicle front-end device, preliminary feature extraction is performed on the original point cloud using depthwise separable convolutions, and the feature map is compressed and uploaded. The cloud server utilizes a residual network with high computing power to decode the compressed features and perform deep semantic analysis.

[0061] Furthermore, in the output stage of step 6, the system adds a spatiotemporal comparison analysis module based on historical trajectories. By retrieving the identification results of the road segment in the previous inspection cycle, the system automatically calculates the crack propagation rate. The specific calculation logic is as follows: by matching coordinates, two cracks at the same location are found, and the difference between their length and average width is compared.

[0062] In summary, this invention systematically solves the core technical problems faced by 3D point cloud data in road crack identification, such as massive data volume, severe environmental noise interference, and incomplete representation of disease structure, through a closed-loop design of the entire process from raw point cloud acquisition, adaptive spatial index construction, multi-dimensional geometric feature fusion extraction, deep learning semantic segmentation, dynamic topology reconstruction to high-precision disease evaluation.

[0063] In the spatial preprocessing stage, a combination of statistical filtering and reflection intensity filtering was used to purify the raw data stream to the greatest extent possible, laying the foundation for subsequent high-precision identification. The introduction of adaptive voxel grids and octree indexes gave the processing algorithm excellent scalability, enabling it to flexibly handle computational needs ranging from microscopic road segments of a few meters to macroscopic road networks of hundreds of kilometers, thus optimizing the allocation of computational resources.

[0064] At the feature extraction and recognition level, this invention does not rely on a single elevation threshold. Instead, it integrates high-dimensional geometric features such as normal vectors, curvature, and information entropy, as well as physical features such as reflection intensity, and combines this with a deep neural network with an adaptive attention mechanism to construct a deep perception capability for multi-dimensional features of road surface defects. This enables the system to exhibit extremely strong robustness and discriminative power when faced with interference such as tree shadows, water stains, road repair marks, and complex road markings.

[0065] In the topology reconstruction and quantitative evaluation stage, through density-space-based clustering analysis and a segment connection algorithm based on geometric continuity constraints, this invention possesses "logical association" capabilities similar to manual inspection, aggregating discrete identification points into physically meaningful crack entities and accurately reconstructing their complex topological orientation. Combined with vectorized evaluation results generated from high-precision geographic location information, it directly interfaces with modern smart city geographic information systems, enhancing the digitalization and intelligence level of transportation infrastructure maintenance.

[0066] The embodiments described in this specification are merely representative solutions of the present invention. In practical applications, those skilled in the art can make targeted adjustments and optimizations to the sensor parameters, the number of layers in the deep learning model, the minimum size of the voxel division, and the threshold for defect assessment based on the specific road surface type (such as airport runways, bridge paving, highways, etc.). For example, when processing airport runways, the detection threshold for crack width can be significantly lowered to capture fatigue cracks at the micrometer level; when processing the area around bridge expansion joints, the curvature sensitivity threshold can be adjusted to distinguish normal structural joints.

[0067] The various algorithmic logics involved in this invention, including least squares fitting, covariance matrix decomposition, octree partitioning, forward inference and backpropagation of neural networks, density clustering, and spline curve interpolation, can all be implemented on computing platforms equipped with high-performance graphics processors or tensor processors, and their processing throughput can be further improved through parallel programming architectures. Simultaneously, the disease database generated by this invention can be used as training data to continuously feed back into deep learning models, achieving iterative improvements in recognition performance.

[0068] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic identification of road surface cracks based on three-dimensional point cloud data, characterized in that, Includes the following steps: Step 1: Acquire the original 3D point cloud data of the road and perform spatial preprocessing. Simultaneously collect the 3D coordinate information, reflection intensity information and absolute geographical location information of the road surface through a multi-beam lidar sensor and a combined inertial navigation system. Perform denoising, resampling and coordinate system transformation on the acquired discrete point cloud sequence to generate a standardized road surface point cloud dataset. Step 2: Construct a spatial partition index structure. Based on the standardized road surface point cloud dataset, use an adaptive voxel mesh partitioning method to divide the massive point cloud into spatial processing units of a preset size, and establish a topological association index for each spatial processing unit. Step 3: Extract multidimensional geometric features. For the point cloud in each spatial processing unit, calculate the local normal vector, surface curvature, information entropy, and reflection intensity distribution features of the sampling points. Construct a high-dimensional feature vector for each sampling point through a feature fusion algorithm. Step 4: Construct and train a deep learning semantic segmentation model, input the spatial processing unit carrying the multi-dimensional geometric features into a preset deep neural network, perform classification label prediction on each sampling point in the point cloud, and identify objects belonging to the crack candidate point cloud; Step 5: Perform crack topology reconstruction and noise removal, perform density space-based clustering analysis on the crack candidate point cloud, obtain the central axis of the crack through the skeleton extraction algorithm, and logically connect the broken crack segments based on the preset geometric continuity constraints. Step 6: Quantitative evaluation and mapping output of road defects. Based on the reconstructed crack topology, calculate the geometric indices of the cracks and generate a road defect vector map and technical condition assessment report with mapping accuracy by combining the absolute geographical location information. 2.The method of claim 1, wherein, The process of obtaining the original three-dimensional point cloud data of the road in step 1 includes: emitting laser pulses at a preset frequency and receiving the echo signal reflected from the road surface through the multi-beam lidar sensor; Based on the time difference of the laser pulse round trip and the beam angle at the moment of sensor emission, the relative spatial coordinates of each reflection point relative to the center of the sensor are calculated. The attitude angle, acceleration, and GPS coordinates of the mobile inspection platform are recorded in real time by a combined inertial navigation system, and the time axis of the multi-beam lidar sensor is aligned with that of the combined inertial navigation system using timestamp synchronization technology. Based on the six-degree-of-freedom pose information of the inspection platform, a rotation and translation matrix is ​​constructed to convert the relative spatial coordinates into a unified geographic spatial coordinate system. 3.The method of claim 1, wherein, The spatial preprocessing process in step 1 includes: Perform statistical filtering to calculate the average Euclidean distance between each sampling point in the point cloud dataset and a preset number of neighboring points in the spatial neighborhood. After traversing all sampling points of the entire road segment, calculate the global mean and standard deviation of the average Euclidean distance of all points; A distance threshold is set, which is equal to the sum of the global average and the standard deviation of a preset multiple; If the average Euclidean distance of the sampling points is greater than the distance threshold, the sampling point is determined to be an outlier and is removed. Perform filtering based on reflection intensity. By analyzing the difference in reflectivity between the road surface material and the interfering object in the laser band, set a reflection intensity filtering threshold. Points with reflection intensity values ​​in a preset low range or a preset high range are marked as interfering candidate points. Based on the local variance of elevation, if the elevation consistency between the interference candidate point and the surrounding road surface points is higher than a preset consistency threshold, then the interference candidate point will be removed from the set to be processed.

4. The method of claim 1, wherein the method further comprises: The process of constructing the spatial partition index structure in step 2 includes: establishing an axis-aligned three-dimensional bounding box based on the total length, width, and relative elevation range of the road segment to be processed; The adaptive voxel mesh generation method dynamically adjusts the side length of the voxel unit according to the local changes in point cloud density, using a large voxel size in sparse point cloud regions and a small voxel size in dense point cloud regions. An octree data structure is used to recursively decompose the voxel units, dividing each non-leaf node into 8 sub-voxel units evenly until the preset termination condition is met. The termination conditions include the voxel unit's side length reaching a preset minimum precision size, the number of point clouds within the voxel being less than a preset minimum value, or the subdivision depth reaching a preset maximum level. Each generated spatial processing unit is assigned a unique global index number, and its topological associations with neighboring voxels are recorded.

5. The method of claim 1, wherein the method further comprises: The process of calculating the local normal vector of each sampling point in step 3 adopts the least squares plane fitting method, including: centering the coordinate components of the target point and all its neighboring points within the preset search radius, and calculating the sum of the self-outer products of the coordinate deviation vectors to construct a three-row, three-column symmetric covariance matrix. The symmetric covariance matrix is ​​decomposed into eigenvalues ​​to obtain three eigenvalues ​​and three corresponding eigenvectors arranged in descending order; the eigenvector corresponding to the smallest eigenvalue is used as the local normal vector of the sampling point. The surface curvature is characterized by the ratio of the minimum eigenvalue to the sum of the three eigenvalues; The information entropy is measured based on the directional distribution probability of the local normal vector in the neighborhood. If the direction of the local normal vector is randomly distributed, the information entropy value increases, which is used to distinguish between regular mechanical seams and irregular natural cracks. 6.The method of claim 1, wherein, The process of constructing the high-dimensional feature vector of each sampling point in step 3 includes: splicing together the three-dimensional spatial coordinate components of the sampling point, the three components of the local normal vector, the surface curvature value, the information entropy value, the normalized reflection intensity information, and the standard deviation of the reflection intensity in the neighborhood. By introducing local anisotropy features, linearity index, flatness index, and scattering index are calculated using the three eigenvalues ​​of the symmetric covariance matrix. The linearity index expression formula is ; wherein is a first eigenvalue, is a second eigenvalue; The formula for the flatness index is: ;in, It is the third eigenvalue; The formula for the scattering index is: ; Before calculating the features, the coordinates of the neighboring points are transformed into a local coordinate system with the local normal vector as the reference axis to enhance the rotation invariance of the features.

7. The method for automatic identification of road surface cracks based on three-dimensional point cloud data according to claim 1, characterized in that, In step 4, the deep learning semantic segmentation model adopts a convolutional neural network based on direct point cloud processing. The network structure includes multiple feature downsampling layers and upsampling layers. During the downsampling process, the farthest point sampling strategy is used to reduce the data size, and the spatial geometric relationship between points is captured by the multi-scale local feature aggregation module. The multi-scale local feature aggregation module extracts hierarchical features by performing weighted summation operations within a spherical search space with different radii at each sampling point. A skip connection is established between the corresponding downsampling layer and upsampling layer to directly transfer the high-resolution features of the shallow layer to the deep layer and concatenate them with the semantic features of the deep layer in terms of feature dimension. The deep neural network also includes an adaptive attention mechanism module, which calculates the importance weight coefficients by calculating the covariance relationship between feature channels and applies them to high-dimensional features to enhance geometric features that contribute highly to crack identification. 8.The method of claim 1, wherein, In step 4, the training process of the deep neural network uses a sample library containing labeled cracks, and the loss function consists of two parts: weighted cross-entropy loss and geometric consistency loss. The weighted cross-entropy loss is used to measure the deviation between the predicted category and the true label, and the positive and negative sample distributions are balanced by increasing the weight coefficient of the crack category. The geometric consistency loss is constrained based on the spatial smoothness of the predicted label. If the predicted label of a sampling point is inconsistent with the labels of most sampling points in its neighborhood, a penalty term for loss is added. The deep learning semantic segmentation model also adopts a multi-task learning framework, which outputs the classification labels of sampling points and the regression prediction values ​​of local surface normal vectors in parallel. A normal vector prediction consistency loss term is constructed by calculating the angle difference between the regression predicted value and the local normal vector obtained through geometric fitting. 9.The method of claim 1, wherein, The density-space-based clustering analysis in step 5 includes: setting the core neighborhood radius and the minimum number of members threshold, grouping mutually reachable core points into the same crack point cluster, and removing isolated noise point clusters whose total number of point clouds is less than a preset size threshold. The skeleton extraction algorithm includes calculating the geometric centroid and principal component direction of the covariance matrix of each crack point cluster, moving a sliding window along the principal component direction with a preset step size, calculating the local centroid of the point cloud in a plane perpendicular to the principal direction within each sliding window, and generating an initial skeleton point sequence by connecting a series of local centroid points. The logical connection process includes calculating the three-dimensional Euclidean distance between the end skeleton points of different crack segments and the cosine value of the angle between the end tangent direction vectors. When the three-dimensional Euclidean distance is less than the connection step size threshold and the cosine value of the angle is greater than the preset direction consistency threshold, a spline curve interpolation algorithm is used to generate the connection segment. 10.The method of claim 1, wherein, The process of calculating the geometric parameters of the crack in step 6 includes: identifying abrupt changes in the road surface elevation distribution in the normal plane of each skeleton point along a direction perpendicular to the crack direction, and calculating the horizontal distance between two abrupt changes as the instantaneous crack width at that location. The average width is obtained by weighting the instantaneous crack width at all skeleton points along the entire crack. The crack depth is obtained by calculating the vertical sinking distance of the skeleton point relative to the reference plane of the surrounding smooth road surface; The total length of the crack is obtained by summing the Euclidean distances between all adjacent skeleton points; The method also includes a multi-source data fusion verification mechanism, which synchronously retrieves two-dimensional road surface images captured by a high-definition industrial camera mounted on the inspection vehicle, establishes a projection mapping relationship between the three-dimensional point cloud space and the two-dimensional image plane using the camera's internal and external orientation elements, extracts grayscale gradients and texture features from the image, and assists in correcting false detections in point cloud recognition.

Images