Pedestrian intelligent detection method and system based on three-dimensional point cloud

Abstract

The application discloses a kind of sidewalk intelligent detection method and system based on three-dimensional point cloud, it is related to point cloud model technical field, including: obtaining the original laser point cloud data of sidewalk of mobile surveying and mapping system acquisition, generates the point cloud data set to be detected;Remove roadside vegetation and building facade point cloud, extract sidewalk pavement point cloud;Sliding window mechanism is used to block sidewalk pavement point cloud, utilize polynomial curved surface fitting algorithm to construct the local design elevation plane of each block, and calculate the elevation residual of point cloud relative to local design elevation plane;Multidimensional feature vector graph containing geometric features and texture features is constructed;Based on multidimensional feature vector graph, get independent disease area;The geometric quantitative index of independent disease area is calculated, and combined with the preset road maintenance specification threshold, determine disease grade and output detection report;The application improves the precision of sidewalk pavement disease detection under complex scene.

Description

Technical Field

[0001] This invention relates to the field of point cloud model technology, specifically to a method and system for intelligent detection of pedestrian walkways based on three-dimensional point clouds. Background Technology

[0002] As a crucial component of urban road traffic, the smoothness and integrity of sidewalks directly impact pedestrian safety and the city's appearance. Currently, sidewalk pavement defect detection is gradually shifting from traditional manual inspections to intelligent 3D point cloud detection. Utilizing laser scanning equipment, 3D spatial data of the pavement can be rapidly acquired, effectively addressing the problems of low efficiency, strong subjectivity, and high missed detection rates associated with manual inspections. However, most existing 3D point cloud detection technologies are designed for industrial surfaces, building floors, and vehicular road surfaces, lacking specific optimization for the complex working environment of sidewalks, resulting in numerous technical shortcomings in practical applications.

[0003] First, sidewalk scenarios are subject to complex environmental interference, with numerous vegetation, building facades, and debris on the roadside. Existing point cloud processing technologies rely solely on a single elevation threshold to filter road surface point clouds, failing to accurately distinguish between the road surface and surrounding interfering structures. This easily leads to the inclusion of stray point clouds from vegetation and building facades in the detection range, resulting in low purity of effective road surface point cloud extraction and introducing systematic errors into subsequent baseline fitting and defect detection. Furthermore, traditional baseline fitting often employs a global planar fitting method. Sidewalk surfaces exhibit slight natural undulations and paving seams, making global fitting unsuitable for adapting to local terrain variations. This can easily cause distortion in elevation residual calculations and significant deviations in flatness detection. Second, existing sidewalk defect detection methods are limited in scope, mostly utilizing only point cloud geometric elevation features for threshold segmentation, neglecting the road texture features corresponding to laser reflection intensity. The elevation changes of minor defects such as cracks and shallow damage in sidewalks are subtle and difficult to identify based on geometric features alone. Furthermore, traditional segmentation algorithms are poor at distinguishing between contiguous and fragmented defects, easily leading to missed or over-segmentation of defects. They cannot accurately separate independent defect areas, and the accuracy of defect detection is insufficient to meet road maintenance standards. Finally, the existing detection process has a simple and crude data processing link, failing to utilize the spatiotemporal correlation of multi-frame point clouds for data enhancement. Single-frame point cloud acquisition is susceptible to scanning jitter and environmental noise, resulting in point elevation distortion and data sparsity. In addition, traditional technologies only complete defect identification and grade determination, failing to achieve defect geographic location and automatic generation of digital work orders. Detection results cannot be directly linked to on-site maintenance operations, resulting in poor linkage between detection, location, and operation and maintenance processes, making it difficult to achieve intelligent and standardized closed-loop detection of sidewalk defects. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides a method and system for intelligent pedestrian detection based on 3D point clouds. This technical solution solves the problems mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A smart pedestrian detection method based on 3D point clouds includes: The original laser point cloud data of the sidewalk collected by the mobile mapping system is acquired, and the original laser point cloud data is denoised and spatially downsampled to generate a point cloud dataset to be detected. Based on the local normal vector distribution characteristics of the point cloud dataset to be detected, the point clouds of roadside vegetation and building facades are removed, and the point cloud of the sidewalk surface is extracted. The point cloud of the pedestrian road surface is divided into blocks using a sliding window mechanism. A local design elevation surface for each block is constructed using a polynomial surface fitting algorithm, and the elevation residual of the point cloud relative to the local design elevation surface is calculated. By fusing the elevation residual with the laser reflection intensity information of the point cloud dataset to be detected, a multidimensional feature vector map containing geometric and texture features is constructed. Based on the multidimensional feature vector map, a multi-scale morphological reconstruction algorithm is used to extract initial seed points for pavement distress, and an improved watershed algorithm is used to perform regional growth on the initial seed points to obtain independent distress regions. Calculate the geometric quantitative index of the independent disease area, and combine it with the preset road maintenance standard threshold to determine the disease level and output a test report.

[0006] Furthermore, based on the local normal vector distribution characteristics of the point cloud dataset to be detected, the point clouds of roadside vegetation and building facades are removed, and the point cloud of the sidewalk surface is extracted, including: For each point in the point cloud dataset to be detected, a local neighborhood is constructed, the covariance matrix of the points in the neighborhood is calculated and the eigenvalues ​​are solved to obtain the normal vector tilt angle of each point. Set a threshold for the tilt angle of the ground point normal vector, and mark points whose tilt angle is less than the threshold as candidate ground points; Perform connectivity analysis on the candidate ground points and calculate the flatness index of each connectivity component; If the flatness index is lower than the preset flatness threshold, the connected region is determined to be a noise or non-road surface area and is removed, while the remaining point cloud is retained as the pedestrian road surface point cloud.

[0007] Furthermore, a sliding window mechanism is used to divide the pedestrian walkway point cloud into blocks, and a polynomial surface fitting algorithm is used to construct a local design elevation surface for each block, including: Set a fixed length of sliding window along the direction of the sidewalk, and set the overlap rate between adjacent windows; Within each sliding window, a random sampling consensus algorithm is used to iteratively fit the planar model and remove outlier points within the window; The least squares method is used to fit the point cloud within the window after removing outliers to a biquadratic polynomial surface to obtain the local design elevation surface equation corresponding to the window. A weighted smooth transition is applied to the edge regions of adjacent sliding windows, and the local design elevation surfaces of all windows are spliced ​​together to generate a continuous global-local design elevation surface.

[0008] Furthermore, by fusing the elevation residual with the laser reflection intensity information of the point cloud dataset to be detected, a multi-dimensional feature vector map containing geometric and texture features is constructed, including: The point cloud dataset to be detected is projected onto a two-dimensional grid plane to generate an elevation residual grid map and a reflection intensity grid map. The elevation residual raster is normalized to highlight areas of negative elevation anomalies; Histogram equalization is performed on the reflection intensity raster to enhance the contrast of crack texture; The processed elevation residual raster image is used as the first channel, and the processed reflection intensity raster image is used as the second channel. The channels are cascaded to construct a dual-channel multidimensional feature vector map.

[0009] Furthermore, based on the aforementioned multi-dimensional feature vector map, a multi-scale morphological reconstruction algorithm is used to extract initial seed points for pavement distresses, including: Define circular structural elements of different sizes, and reconstruct the geometric feature channels in the multidimensional feature vector map by performing an opening operation to obtain the background estimation image; The difference between the geometric feature channel and the background estimation image is calculated to obtain a geometric saliency map; A top-hat transform is performed on the texture feature channels in the multidimensional feature vector map to obtain a texture saliency map. The geometric saliency map and the texture saliency map are weighted and fused to obtain the disease saliency map; The local maximum detection operator is applied to the saliency map of the road surface defects to extract the pixel extreme points as the initial seed points of the road surface defects.

[0010] Furthermore, an improved watershed algorithm is used to perform region growing on the initial seed point to obtain independent disease regions, including: Calculate the gradient magnitude image of the multidimensional feature vector map, and use the gradient magnitude image as the topological topography of the watershed algorithm; The initial seed point is marked as the lowest point of the catchment basin; The process of simulating water level rise is carried out by flooding the topology starting from the initial seed point. When the water from different catchment basins is about to converge, a water-blocking dam is constructed. Traverse the topology until all areas above the initial seed point are submerged, extract the pixel coordinates of the water-blocking dam, and form a closed boundary contour of the disease. Closed contours with an area smaller than the preset noise area threshold are removed to obtain the final independent disease area.

[0011] Furthermore, the geometric quantitative indicators of the independent diseased areas are calculated, and combined with preset road maintenance standard thresholds, the disease level is determined, including: Extract the three-dimensional coordinates of all point clouds within the independent disease area, and calculate the projected area, maximum length, and maximum width of the area on the horizontal plane; Calculate the maximum and average depth of the point cloud within the independent disease area relative to the local design elevation surface; If the projected area is less than the first area threshold and the maximum width is less than the crack width threshold, then the disease type is determined to be a transverse or longitudinal crack. If the projected area is greater than the second area threshold and the average depth is greater than the pit depth threshold, then the disease type is determined to be a pit or subsidence. Based on the identified disease type and corresponding geometric quantitative index value, the damage level is determined by matching the damage grading standard in the road maintenance specification.

[0012] Furthermore, the method also includes: Obtain the position information of the centroid of the independent diseased area in the global coordinate system; Read the positioning and attitude data synchronously collected by the mobile mapping system and map the independent disease area onto the pedestrian walkway geographic information map; A digital maintenance work order containing disease type, disease level, geometric quantitative indicators and geographical location information is generated and sent to the terminal display device.

[0013] Furthermore, after acquiring the original laser point cloud data of the sidewalk collected by the mobile mapping system, a point cloud enhancement step based on spatiotemporal consistency is also included: Obtain the current frame's point cloud dataset to be detected and its corresponding N previous frames' historical point cloud datasets; Based on the pose data of the mobile mapping system, the previous N frames of historical point cloud dataset are registered to the current frame coordinate system. For any target point in the current frame's point cloud dataset, search for a set of matching points corresponding to its spatial location in the previous N frames' historical point cloud dataset. If the number of points in the matching point set meets the preset confidence requirement, then the mean elevation residual between the target point and each point in the matching point set is calculated, and the elevation value of the target point is corrected using the mean elevation residual. If the number of points in the matching point set does not meet the preset confidence requirement, then the elevation value of the target point remains unchanged; The corrected point cloud dataset is used to perform subsequent steps.

[0014] A pedestrian walkway intelligent detection system based on 3D point clouds includes: The data acquisition module acquires the original laser point cloud data of the sidewalk collected by the mobile mapping system, and performs noise reduction and spatial downsampling processing on the original laser point cloud data to generate a point cloud dataset to be detected. The road surface extraction module, based on the local normal vector distribution characteristics of the point cloud dataset to be detected, removes the point clouds of roadside vegetation and building facades, and extracts the point cloud of the sidewalk road surface. The reference surface construction module uses a sliding window mechanism to divide the pedestrian pavement point cloud into blocks, uses a polynomial surface fitting algorithm to construct a local design elevation surface for each block, and calculates the elevation residual of the point cloud relative to the local design elevation surface. The feature fusion module fuses the elevation residual with the laser reflection intensity information of the point cloud dataset to be detected, and constructs a multi-dimensional feature vector map containing geometric features and texture features; The disease segmentation module, based on the multi-dimensional feature vector map, uses a multi-scale morphological reconstruction algorithm to extract initial seed points for road surface diseases, and uses an improved watershed algorithm to perform regional growth on the initial seed points to obtain independent disease regions. The result determination module calculates the geometric quantitative indicators of the independent disease area, and determines the disease level by combining them with the preset road maintenance standard threshold and outputs a test report.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves precise removal of non-road surface interference point clouds by combining point cloud normal vector distribution characteristics with connected domain flatness analysis. It can effectively filter invalid point clouds such as roadside vegetation and building facades, avoiding the limitations of traditional single elevation screening methods, and significantly improving the purity and accuracy of pedestrian pavement point cloud extraction. It eliminates detection errors caused by environmental interference from the data source level. It uses a sliding window block fitting combined with weighted smoothing splicing to construct a local design elevation surface, replacing the traditional global plane fitting mechanism. This can adapt to the local natural undulation characteristics of pedestrian pavement and conform to the actual pavement pattern. It effectively solves the problems of inaccurate elevation residual calculation and large flatness detection deviation caused by global fitting, and improves the rationality and accuracy of pavement reference surface construction. 2. This invention integrates geometric features of elevation residuals and texture features of laser reflection intensity to construct a dual-channel multidimensional feature vector map. Normalization and histogram equalization are used to enhance the differences in disease characteristics, overcoming the technical limitation of single geometric features in identifying shallow and subtle pavement defects. This significantly improves the feature recognition of weak and textured defects. Multi-scale morphological reconstruction combined with an improved watershed algorithm is used to segment defect areas. Precise defect seed points are extracted through weighted fusion of dual saliency maps. Furthermore, a topological inundation mechanism is used to achieve precise extraction of region growth and boundaries. This effectively distinguishes between adherent and fragmented defects, eliminates noise-induced pseudo-defect areas, and effectively identifies various types of sidewalk cracks, potholes, and other defects. The independent and precise segmentation of subsidence defects significantly improves segmentation accuracy. By combining multi-dimensional geometric quantitative indicators with road maintenance standard thresholds for defect classification and grading, different defect types such as cracks, potholes, and subsidence can be accurately distinguished, achieving standardized grading of defect severity. This avoids the subjective errors of traditional manual threshold judgment, ensuring that the detection results are fully compatible with municipal road maintenance standards. A multi-frame point cloud enhancement mechanism based on spatiotemporal consistency is introduced, using historical frame point information to correct the current frame's elevation distortion points. This effectively solves the data deviation problems caused by jitter, sparsity, and noise in single-frame point cloud acquisition, improving the accuracy and stability of the overall point cloud data and providing high-quality data support for subsequent defect detection. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the present invention. Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Example 1: Refer to Figure 1 As shown, a smart pedestrian detection method based on 3D point clouds includes: Step 101: Acquire the original laser point cloud data of the sidewalk collected by the mobile mapping system, and perform denoising and spatial downsampling processing on the original laser point cloud data to generate a point cloud dataset to be detected. Specifically, a mobile mapping system composed of a 3D LiDAR and a positioning and attitude determination module mounted on a mobile inspection device is used to scan at a constant speed along the sidewalk direction to collect the original laser point cloud data of the entire sidewalk surface, roadside vegetation, building facades and ancillary facilities. The original point cloud data contains a large number of redundant points, long-distance shot noise points and densely repeated points. Statistical filtering algorithms are used to remove isolated discrete noise points, and voxel downsampling is used to uniformly downsample the dense point cloud. While preserving the three-dimensional geometric features of the road surface, the point cloud density is unified to generate a point cloud dataset with uniform resolution and no redundant interference to be detected. The spatial downsampling mentioned here refers to the sampling optimization method of filtering the original dense point cloud through a fixed-size voxel grid, retaining only one representative point in each voxel. Step 102: Based on the local normal vector distribution characteristics of the point cloud dataset to be detected, remove the point clouds of roadside vegetation and building facades, and extract the point cloud of the sidewalk surface. Specifically, the sidewalk surface is a flat and continuous horizontal plane structure, while roadside vegetation and building facades are mostly inclined, messy, and discrete non-planar structures, and the normal vector distribution characteristics of the two are significantly different. By constructing local neighborhoods point by point to solve the normal vector parameters, and combining the tilt angle threshold and the connected component flatness screening mechanism, the horizontal road surface points and non-road surface interference points are distinguished, and invalid interference point clouds are accurately removed to obtain a pure sidewalk surface point cloud, eliminating the interference of the lateral environment on subsequent detection. The local normal vector distribution characteristics mentioned here refer to the tilt angle, dispersion, and consistency distribution law of the normal vectors of all points in the local neighborhood of the point cloud. Step 103: The pedestrian pavement point cloud is divided into blocks using a sliding window mechanism. A local design elevation surface for each block is constructed using a polynomial surface fitting algorithm, and the elevation residual of the point cloud relative to the local design elevation surface is calculated. Specifically, the traditional global plane fitting method cannot adapt to the local undulations of the pavement. By sliding blocks along the direction of the pedestrian pavement, a reference surface is fitted separately for each local pavement segment. Combined with the random sampling consistency algorithm, the pavement protrusions and debris and paving high points are removed. The natural undulation of the pavement is fitted by a biquadratic polynomial fitting method to generate a continuous and smooth global and local design elevation surface. The difference between the actual pavement point cloud elevation and the design elevation surface elevation is calculated to obtain elevation residual data that can accurately reflect the pavement's unevenness, damage, and settlement. The elevation residual here refers to the difference between the actual pavement point elevation and the local reference design elevation. A positive residual represents a pavement protrusion, and a negative residual represents a pavement depression and damage. Step 104: Integrate the elevation residual and laser reflection intensity information of the point cloud dataset to be detected to construct a multi-dimensional feature vector map containing geometric and texture features. Specifically, the elevation residual contains road surface geometric deformation information, and the laser reflection intensity contains road surface material, texture, and damage crack information. The two features are complementary. The three-dimensional point cloud data is rasterized and projected to generate elevation residual raster map and reflection intensity raster map respectively. Feature enhancement is completed through normalization and histogram equalization. The two feature maps are channel-cascaded and fused to construct a dual-channel multi-dimensional feature vector map that takes into account both geometric deformation and texture differences, thus making up for the deficiency that a single geometric feature cannot identify minute cracks and shallow damage. Step 105: Based on the multi-dimensional feature vector map, a multi-scale morphological reconstruction algorithm is used to extract initial seed points for pavement distresses. An improved watershed algorithm is then used to perform region growth on the initial seed points to obtain independent distress regions. Specifically, considering the characteristics of sidewalk distresses, such as varying sizes, shapes, and susceptibility to adhesion interference, background estimation and feature difference are performed using multi-scale structural elements. Geometric saliency maps and texture saliency maps are fused to obtain a high-precision distress saliency map. Local maximum detection is used to extract core distress seed points, solving the problems of messy seed points and high false detection rate in traditional methods. Based on the improved watershed algorithm, region submergence growth is performed using the seed points as a benchmark to accurately segment adhesion distresses and fragmented distresses, and eliminate false distress noise regions, resulting in various types of pavement distress regions with complete edges and independent regions. Step 106: Calculate the geometric quantitative indicators of independent disease areas, and combine them with the preset road maintenance specification thresholds to determine the disease level and output a test report; Specifically, perform three-dimensional geometric parameter calculation on the segmented independent disease areas to accurately extract quantitative indicators such as disease area, length and width dimensions, maximum depth, and average depth; Based on the preset disease judgment thresholds and grading standards of municipal sidewalk maintenance technical specifications, automatically distinguish disease types such as cracks, potholes, and subsidence, and complete the determination of mild, moderate, and severe levels; integrate all disease detection data to generate a standardized and traceable intelligent test report; The above technical solution achieves accurate purification of road surface point clouds by relying on point cloud normal vector features, constructs a benchmark elevation surface adapted to road surface undulations by local polynomial fitting, enhances disease features by integrating geometric and texture multidimensional features, and achieves accurate segmentation and classification of diseases by combining multi-scale morphology and improved watershed algorithm. It effectively solves the technical problems of traditional sidewalk point cloud detection being susceptible to environmental interference, missing subtle diseases, being unable to distinguish adhering diseases, and having low detection accuracy, and realizes automated, high-precision, and full-dimensional intelligent detection of sidewalk pavement diseases.

[0019] In some embodiments, based on the local normal vector distribution characteristics of the point cloud dataset to be detected, the point clouds of roadside vegetation and building facades are removed, and the point cloud of the sidewalk surface is extracted, including: Step 201: Construct a local neighborhood for each point in the point cloud dataset to be detected, calculate the covariance matrix of the points within the neighborhood and solve for the eigenvalues ​​to obtain the normal vector tilt angle of each point; specifically, take any target point in the point cloud to be detected as the center, set a spherical neighborhood with a fixed radius, and select all neighboring points within the neighborhood to form a local point set; solve for the three-dimensional covariance matrix of this local point set, perform eigenvalue decomposition on the covariance matrix to obtain three sets of eigenvalues ​​and corresponding eigenvectors; take the eigenvector corresponding to the smallest eigenvalue as the normal vector of that point, and calculate the normal vector tilt angle of each point by calculating the angle between the normal vector and the horizontal plane; the normal vector tilt angle mentioned here refers to the angle between the normal vector of the point cloud and the horizontal ground, the tilt angle of the horizontal road surface points is close to 0, and the tilt angle of the vertical and cluttered points is close to 90 degrees; Step 202: Set the ground point normal vector tilt angle threshold, and mark points with a normal vector tilt angle less than the ground point normal vector tilt angle threshold as candidate ground points; specifically, based on the planar characteristics of the sidewalk pavement, the ground point normal vector tilt angle threshold is preset to 10°; traverse all point cloud points, determine points with a normal vector tilt angle less than 10° as candidate ground points with pavement characteristics, and initially remove tilted and vertical points with a tilt angle greater than or equal to 10° to complete the initial screening of non-pavement points; Step 203: Perform connected component analysis on the candidate ground points and calculate the flatness index of each connected component. Specifically, perform spatial connectivity clustering on all selected candidate ground points, dividing the spatially continuous and densely packed point cloud into the same connected component, and the discrete and sparse points into separate components. For all points within each connected component, fit a local plane and calculate the elevation deviation of each point relative to the plane. Solve the flatness index of the connected component by the dispersion of the elevation deviation. The smaller the flatness index value, the flatter the area is, and the more it conforms to the characteristics of a pedestrian road surface. The flatness index mentioned here is used to characterize the flatness and uniformity of the connected component point cloud, and can effectively distinguish between flat road surfaces and areas with messy vegetation and uneven debris. Step 204: If the flatness index is lower than the preset flatness threshold, the connected region is determined to be a noise or non-road area and is removed, while the remaining point cloud is retained as the pedestrian road surface point cloud. Specifically, the preset flatness threshold is 0.05cm. When the flatness index of the connected region is greater than 0.05cm, the connected region is determined to be uneven and messy, does not conform to the road surface characteristics, and belongs to vegetation, debris, or ground noise areas, and is directly removed. The connected region point cloud that meets the flatness standard and is continuously flat is retained, and finally a high-purity pedestrian road surface point cloud is obtained.

[0020] By using the above technical solution, a two-layer screening mechanism combining global normal vector tilt angle screening and local connected domain flatness verification is adopted to replace the traditional single elevation threshold screening method. This can accurately filter out interfering point clouds such as roadside vegetation, building facades, and scattered debris, greatly improving the accuracy of road surface point cloud extraction and avoiding detection errors caused by environmental interference from the source.

[0021] In some embodiments, a sliding window mechanism is used to divide the pedestrian pavement point cloud into blocks, and a polynomial surface fitting algorithm is used to construct a local design elevation surface for each block, including: Step 301: Set a sliding window of fixed length along the direction of the sidewalk extension, and set the overlap rate between adjacent windows; specifically, in combination with the conventional sidewalk paving width and road surface undulation scale, set the size of the sliding window to 2m×2m to ensure that a single window can cover the complete local road surface unit; set the overlap rate of adjacent sliding windows to 30% to avoid the problem of characteristic discontinuity and surface discontinuity at the window splicing, and ensure the smoothness of the global reference plane; Step 302: Within each sliding window, the random sampling consensus algorithm is used to iteratively fit the planar model and remove outlier points within the window. Specifically, in the road surface point cloud of a single sliding window, a small number of points are randomly selected to fit the initial planar model, and the fit between all points within the window and the plane is iteratively calculated. After multiple iterations of optimization, outlier points that deviate from the plane by a distance exceeding a preset deviation threshold are counted. These outlier points are mostly road surface stones, debris, and protruding paving defects. They are all removed, leaving a clean and flat main road surface point cloud. Step 303: Use the least squares method to perform biquadratic polynomial surface fitting on the point cloud within the window after removing outliers to obtain the local design elevation surface equation corresponding to the window; specifically, use the biquadratic polynomial as the fitting benchmark model, the least squares method as the optimization algorithm, and the minimum sum of squared residuals of point elevations as the fitting objective to perform surface fitting on the clean road surface point cloud within the window; solve to obtain the surface equation containing lateral, longitudinal, and intersection terms, which accurately adapts to the small natural undulations of the road surface and replaces the traditional rigid plane fitting model; Step 304: Apply weighted smooth transition processing to the edge areas of adjacent sliding windows, and stitch together the local design elevation surfaces of all windows to generate a continuous global local design elevation surface. Specifically, for the 30% overlap area of ​​adjacent windows, use distance-weighted interpolation to smooth the surface elevation of the overlap area. The closer to the center of the window, the greater the weight, and the closer to the edge of the overlap, the smaller the weight. Eliminate the step-like error caused by block fitting through weighted transition, and seamlessly stitch together all block local surfaces to generate a global local design elevation surface that is continuous overall and locally conforms to the road surface shape.

[0022] The above technical solution, based on sliding window block fitting and weighted smooth splicing mechanism, takes into account both the global smoothness and local undulation characteristics of the road surface, effectively solving the problem that traditional global plane fitting cannot adapt to the small terrain changes of the sidewalk and the distortion of elevation residual calculation, and greatly improving the accuracy and rationality of the road surface reference surface construction.

[0023] In some embodiments, the laser reflection intensity information of the elevation residual and the point cloud dataset to be detected is fused to construct a multi-dimensional feature vector map containing geometric and texture features, including: Step 401: Project the point cloud dataset to be detected onto a two-dimensional raster plane to generate an elevation residual raster map and a reflection intensity raster map. Specifically, the resolution of the two-dimensional raster is uniformly set to 5mm×5mm, and the planar coordinates, elevation residual values, and laser reflection intensity values ​​of the three-dimensional road surface point cloud are mapped to the corresponding raster cells. For blank raster cells without point cloud coverage, neighborhood interpolation is used to fill them, and elevation residual raster maps and reflection intensity raster maps with consistent spatial resolution and corresponding dimensions are generated respectively, so as to realize the two-dimensional structured representation of the three-dimensional point cloud features. Step 402: Normalize the elevation residual raster map to highlight negative elevation anomaly areas. Specifically, extract the residual extreme values ​​of all grids in the elevation residual raster map, perform maximum and minimum value normalization based on the extreme value interval, and map all residual values ​​to the 0-1 interval. For the negative elevation residuals corresponding to pavement damage and potholes, amplify the feature response of the concave area through weight enhancement processing, strengthen the geometric anomaly characteristics of the damage, and highlight the difference between the pavement damage area and the normal pavement. Step 403: Perform histogram equalization on the reflection intensity grid to enhance the contrast of crack texture; specifically, the laser reflection intensity of cracks and minor damage areas on the sidewalk pavement is slightly different from that of the intact paved area, and the original image has low contrast; by using histogram equalization to redistribute the pixel grayscale distribution of the reflection intensity grid, stretch the effective grayscale range, amplify the intensity difference of crack texture and damage marks, and achieve feature enhancement of minor texture defects; Step 404: Using the processed elevation residual raster map as the first channel and the processed reflection intensity raster map as the second channel, channel cascade is performed to construct a dual-channel multidimensional feature vector map. Specifically, the normalized geometric feature channel and the equalized texture feature channel are spliced ​​and fused in dimension. Each raster position simultaneously contains elevation deformation information and reflection texture information to form a dual-channel multidimensional feature vector map, realizing the synchronous representation of geometric defects and texture defects.

[0024] The above technical solution achieves deep integration of road surface geometric deformation features and texture damage features, solving the technical problem that single features are difficult to identify shallow cracks, minor damage and other weak defects, greatly enriching the dimensions of defect features and improving the sensitivity of weak defect identification.

[0025] In some embodiments, based on a multi-dimensional feature vector map, a multi-scale morphological reconstruction algorithm is used to extract initial seed points for pavement distresses, including: Step 501: Define circular structural elements of different sizes and reconstruct the geometric feature channels in the multidimensional feature vector map by performing opening operations to obtain the background estimation image. Specifically, based on the differences in the scale of sidewalk damage, define three sets of circular structural elements of different sizes: 3×3, 5×5, and 7×7, to adapt to different scales of damage such as minor cracks, moderate damage, and large potholes. Morphological reconstruction of the geometric feature channels is performed through multi-scale opening operations to filter out foreground damage features and retain the background features of the smooth road surface area, generating a background estimation image that conforms to the normal road surface condition. Step 502: Calculate the difference between the geometric feature channel and the background estimation image to obtain a geometric saliency map; specifically, perform pixel-by-pixel difference operation between the original geometric feature channel image and the background estimation image, remove the stable background information of the road surface, and retain only the foreground area with elevation anomalies and geometric deformations to generate a geometric saliency map that highlights the road surface unevenness, damage, and settlement defects. Step 503: Perform top-hat transformation on the texture feature channels in the multi-dimensional feature vector map to obtain a texture saliency map; specifically, top-hat transformation can effectively extract small areas of light and dark difference in the image. Perform multi-scale top-hat transformation on the texture feature channels to filter out the uniform texture background of the road surface, enhance texture abnormal areas such as cracks, wear, and peeling, and generate a texture saliency map that focuses on the damage to the road surface texture. Step 504: Perform weighted fusion of the geometric saliency map and the texture saliency map to obtain the disease saliency map; specifically, based on the characteristics of sidewalk disease detection, set the weight of the geometric saliency map to 0.6 and the weight of the texture saliency map to 0.4, and perform weighted superposition and fusion of the two saliency maps; take into account both geometric deformation defects and texture damage defects, suppress noise interference from single-channel features, and generate a disease saliency map with more comprehensive features and lower noise; Step 505: Apply the local maximum detection operator to the saliency map of the road surface defects and extract the pixel extreme points as the initial seed points of the road surface defects. Specifically, a 3×3 local maximum detection window is used to traverse all pixels in the saliency map of the road surface defects and mark the pixel with the largest gray value in the local neighborhood as the extreme point. These extreme points correspond to the core feature positions of the defect area and are free from background noise interference. They are uniformly used as the initial seed points of the road surface defects to provide a benchmark for subsequent accurate segmentation.

[0026] The above technical solution achieves accurate separation of background and foreground through multi-scale morphological reconstruction, integrates two-dimensional salient features to generate a high-quality disease saliency map, and extracts initial seed points to accurately focus on the core area of ​​the disease, effectively solving the problems of messy seed points, susceptibility to noise interference, and inaccurate disease location in traditional methods.

[0027] In some embodiments, an improved watershed algorithm is used to perform region growing on the initial seed point to obtain independent disease regions, including: Step 601: Calculate the gradient magnitude image of the multidimensional feature vector map and use the gradient magnitude image as the topology of the watershed algorithm. Specifically, the Sobel gradient operator is used to calculate the horizontal and vertical gradients of the multidimensional feature vector map respectively, and synthesizes them to obtain the overall gradient magnitude image. The larger the gradient magnitude, the more drastic the change in regional features, corresponding to the edge of the disease. The gradient magnitude image is used to construct the topology of the watershed algorithm, replacing the traditional single grayscale topology and improving the edge recognition accuracy. Step 602: Mark the initial seed point as the lowest point of the catchment basin; specifically, set the gradient magnitude of the pixel position corresponding to the extracted initial seed point of the disease to the minimum value and mark it as the lowest point of the catchment basin in the topology; ensure that each disease area corresponds to an independent catchment basin, and avoid confusion and adhesion of different disease areas from the root. Step 603: Simulate the water level rise process. Starting from the initial seed point, the topology is flooded. When the water from different catchment basins is about to converge, a water-blocking dam is constructed. Specifically, the water level is simulated to flood the topology step by step in order of gradient magnitude from small to large, and the flooding gradually spreads outward from the lowest point corresponding to each disease seed point. When the flooded areas of two adjacent catchment basins are about to connect, a pixel-level water-blocking dam is constructed at the boundary to separate the boundaries of different disease areas. Step 604: Traverse the topology until all areas above the initial seed point are submerged, extract the pixel coordinates of the water-blocking dams, and form a closed disease boundary outline; specifically, continue to iterate the submerging process until all non-disease background areas in the topology are completely submerged, and only the boundary lines formed by the water-blocking dams are retained; extract the continuous pixel coordinates of all water-blocking dams and stitch them together to form a complete and closed disease area boundary outline. Step 605: Remove closed contours with an area smaller than the preset noise area threshold to obtain the final independent disease area; specifically, the preset noise area threshold is 20 pixel units, filter the enclosing area of ​​all closed disease contours, and determine the small contours with an area smaller than the threshold as noise pseudo-diseases and remove them; retain disease contours with the correct size and real features to obtain an independent disease area with accurate segmentation and no noise interference.

[0028] By using the above technical solutions, the improved watershed algorithm uses precise disease seed points as the flooding benchmark and combines gradient topology to achieve accurate growth and boundary segmentation of disease areas. It can effectively separate complex diseases such as adhesion cracks and fine fragmentation, eliminate the problems of over-segmentation and under-segmentation of diseases, and significantly improve the segmentation accuracy of sidewalk diseases.

[0029] In some embodiments, the geometric quantitative index of an independent diseased area is calculated, and combined with a preset road maintenance specification threshold, the disease level is determined, including: Step 701: Extract the three-dimensional coordinates of all point clouds within the independent disease area, and calculate the projected area, maximum length, and maximum width of the area on the horizontal plane; specifically, traverse all valid point cloud points within the independent disease area and collect the X, Y, and Z three-dimensional coordinates of all points; project the three-dimensional points onto the horizontal reference plane, and calculate the projected area of ​​the disease area using the contour enclosing algorithm; solve for the maximum extension length and maximum lateral width of the disease area by traversing the coordinate extrema, and complete the extraction of the disease plane geometric parameters; Step 702: Calculate the maximum and average depth of the point cloud relative to the local design elevation surface within the independent disease area; specifically, retrieve the elevation data of the local design elevation surface corresponding to the disease area, calculate the negative elevation difference between the disease point and the reference elevation surface point by point, which is the depth of the disease depression; count the maximum value and arithmetic mean of all depth values ​​to obtain the maximum depth and average depth of the disease, which characterize the degree of depression and damage of the disease. Step 703: If the projected area is less than the first area threshold and the maximum width is less than the crack width threshold, the type of the defect is determined to be a transverse or longitudinal crack. Specifically, according to the municipal maintenance specifications, the first area threshold is preset to 0.5㎡ and the crack width threshold is 2cm. When the projected area of ​​the defect area is less than 0.5㎡, the maximum width is less than 2cm, and the area extends in a long strip shape, it is determined to be a road surface crack defect. The transverse crack and the longitudinal crack are further distinguished according to the direction of extension. Step 704: If the projected area is greater than the second area threshold and the average depth is greater than the pothole depth threshold, the type of damage is determined to be pothole or subsidence. Specifically, the second area threshold is preset to 0.5㎡ and the pothole depth threshold is 1.5cm. When the projected area of ​​the damaged area is greater than 0.5㎡, the average depression depth is greater than 1.5cm, and the area is in the form of block depression, it is determined to be pothole damage. Large-area, shallow and uniform depression areas are determined to be road subsidence damage. Step 705: Based on the determined type of damage and the corresponding geometric quantitative index value, match the damage level grading standard in the road maintenance specification to determine the damage level; specifically, retrieve the preset sidewalk damage grading standard, and match the corresponding area, depth, and width grading thresholds for different damage types such as cracks, potholes, and subsidence; compare the measured geometric quantitative index with the grading thresholds to automatically determine the damage to three levels: mild, moderate, and severe, and complete the standardized rating of the damage.

[0030] The above technical solutions enable accurate differentiation of disease types through multi-dimensional geometric quantitative indicators, and automatic disease grading is completed based on standardized maintenance specifications, avoiding subjective errors in manual judgment and ensuring that test results are compliant, consistent, and traceable.

[0031] In some embodiments, the method further includes: Step 801: Obtain the position information of the center of gravity of the independent disease area in the global coordinate system; specifically, extract the three-dimensional coordinates of all boundary points and internal points of the independent disease area, and solve the three-dimensional center of gravity coordinates of the disease area through the center of gravity calculation formula; use the center of gravity coordinates as the precise positioning coordinates of the disease area to characterize the core position of the disease on the road surface. Step 802: Read the positioning and attitude data synchronously collected by the mobile mapping system and map the independent disease area onto the pedestrian walkway geographic information map; specifically, the mobile mapping system synchronously acquires GPS positioning data and inertial attitude data during the point cloud acquisition process, and converts the disease coordinates in the global coordinate system of the point cloud into geographic coordinate system coordinates through the coordinate transformation matrix; accurately map the disease location onto the pre-loaded pedestrian walkway geographic information base map to achieve visualized geographic positioning of the disease; Step 803: Generate a digital maintenance work order containing information on the type, level, geometric quantitative indicators, and geographical location of the damage, and send it to the terminal display device; specifically, integrate the data from the entire process of damage detection, including key information such as damage type, damage level, area, depth, geographical location, and detection time, and automatically generate a digital work order according to the standard template for municipal maintenance work orders; push the work order to the on-site inspection terminal and the back-end management terminal in real time via wireless transmission for maintenance personnel to view and handle.

[0032] The above technical solutions have achieved a closed-loop process for disease detection, geolocation, work order generation, and terminal push, directly connecting intelligent detection results to maintenance and operation scenarios, solving the problem that traditional detection results cannot be applied, and greatly improving the efficiency of intelligent operation in sidewalk maintenance.

[0033] In some embodiments, after acquiring the original laser point cloud data of the sidewalk collected by the mobile mapping system, a point cloud enhancement step based on spatiotemporal consistency is further included: Step 901: Obtain the current frame point cloud dataset to be detected and its corresponding previous N frame historical point cloud datasets; specifically, set the number of historical frames N to 5, and while collecting the current frame pedestrian point cloud data, retrieve the previous 5 consecutively collected historical point cloud data of the same area from the device cache to construct a temporal point cloud dataset containing the current frame and historical frames. Step 902: Based on the pose data of the mobile mapping system, register the previous N frames of historical point cloud datasets to the coordinate system of the current frame; specifically, use the position and attitude time series data recorded in real time by the mobile mapping system to solve the coordinate transformation matrix between each historical frame and the current frame; use the rigid body transformation algorithm to uniformly register all historical frame point clouds to the global coordinate system of the current frame to achieve spatial alignment of multi-frame point clouds. Step 903: For any target point in the current frame's point cloud dataset, search for a set of matching points corresponding to the spatial location in the previous N frames' historical point cloud dataset; specifically, take any target point in the current frame as the center, set a spatial search radius of 3cm, search for matching points within the same location range in the registered historical frame point cloud, and summarize them to form a set of historical matching points for the target point. Step 904: If the number of points in the matching point set meets the preset reliability requirements, calculate the mean elevation residual between the target point and each point in the matching point set, and use the mean elevation residual to correct the elevation value of the target point; specifically, the preset threshold for the number of matching points is 3. When the number of valid points in the matching point set is ≥3, the time series data of that point is determined to be reliable; calculate the mean residual between the current target point elevation and the elevation of each historical matching point, and use the mean residual to compensate and correct the current point elevation to eliminate the elevation deviation caused by single-frame acquisition jitter; Step 905: If the number of points in the matching point set does not meet the preset confidence requirement, the elevation value of the target point remains unchanged; specifically, when the number of matching points is less than 3, it is determined that there is insufficient time series reference data for the point. In order to avoid over-correction and the introduction of new errors, the original elevation data of the current frame is directly retained. Step 906: Execute subsequent steps using the corrected point cloud dataset to be detected; specifically, after traversing all points in the current frame to complete the temporal correction, a high-precision and high-stability enhanced point cloud dataset to be detected is obtained, and subsequent reference surface fitting, feature fusion and disease detection processes are carried out based on the optimized point cloud data.

[0034] The above technical solution utilizes the spatiotemporal consistency characteristics of multi-frame point clouds to complete data enhancement and error correction, effectively solving the problems of jitter, sparsity distortion, and elevation deviation in single-frame laser point cloud acquisition. It improves the accuracy and stability of point clouds from the data source level, providing reliable data support for accurate disease detection.

[0035] Example 2: Refer to Figure 2 As shown, a pedestrian walkway intelligent detection system based on 3D point clouds includes: The data acquisition module is used to acquire the original laser point cloud data of the sidewalk collected by the mobile mapping system, and to perform noise reduction and spatial downsampling processing on the original laser point cloud data to generate a point cloud dataset to be detected. Specifically, the data acquisition module is connected to the mobile mapping lidar and positioning and attitude determination equipment to receive the original point cloud data of the entire sidewalk area in real time. It has built-in statistical filtering and voxel downsampling algorithms, which can automatically remove discrete noise points and simplify redundant points, and output a point cloud dataset with uniform resolution and extremely low noise, providing raw data support for subsequent processing. The road surface extraction module is used to extract the pedestrian road surface point cloud by removing the point clouds of roadside vegetation and building facades based on the local normal vector distribution characteristics of the point cloud dataset to be detected. Specifically, the road surface extraction module has built-in algorithms for normal vector solving, tilt angle threshold filtering, and connected component flatness analysis. It can automatically identify and filter non-road surface interference point clouds on the roadside, accurately purify the continuous and flat pedestrian road surface point cloud, and realize intelligent segmentation of road surface areas in complex scenarios. The reference surface construction module is used to divide the pedestrian pavement point cloud into blocks using a sliding window mechanism, construct the local design elevation surface for each block using a polynomial surface fitting algorithm, and calculate the elevation residual of the point cloud relative to the local design elevation surface. Specifically, the reference surface construction module can adaptively complete sliding window block division, outlier removal, biquadratic polynomial surface fitting, and window smoothing to generate a global and local design elevation surface that fits the real undulations of the pavement, accurately calculate the elevation residual data of each point, and quantify the pavement concavity and convexity deformation characteristics. The feature fusion module is used to fuse the laser reflection intensity information of the elevation residual and the point cloud dataset to be detected, and construct a multi-dimensional feature vector map containing geometric and texture features. Specifically, the feature fusion module has the functions of point cloud rasterization, feature normalization, histogram equalization, and channel cascade fusion, which can simultaneously enhance the geometric deformation features and texture damage features of the road surface, construct a dual-channel multi-dimensional feature map, and enrich the dimension of the disease feature expression. The disease segmentation module is used to extract initial seed points of pavement diseases based on multi-dimensional feature vector maps and a multi-scale morphological reconstruction algorithm. It then uses an improved watershed algorithm to perform region growth on the initial seed points to obtain independent disease regions. Specifically, the disease segmentation module integrates multi-scale morphological processing, dual feature saliency map fusion, extreme value seed point extraction, and an improved watershed region growth algorithm. It can accurately segment various types of sidewalk diseases, effectively distinguish between adhering diseases, eliminate false disease noise, and output independent and complete disease regions. The result determination module is used to calculate the geometric quantitative indicators of independent disease areas, and combine them with preset road maintenance standard thresholds to determine the disease level and output a test report. Specifically, the result determination module can automatically extract geometric quantitative indicators such as disease area, length, width, and depth, and has a built-in municipal sidewalk maintenance grading standard library. It can intelligently distinguish disease types, determine the damage level of diseases, automatically integrate all test data, and generate standardized and visualized intelligent test reports. Through the coordinated operation of the above system modules, the entire process of pedestrian walkway point cloud data acquisition, preprocessing, pavement purification, reference surface construction, feature fusion, defect segmentation, grade determination and report output can be completed automatically without manual intervention. This effectively solves the problems of low efficiency, large error and insufficient intelligence of traditional manual inspection, and realizes high-precision, automated and intelligent batch detection of pedestrian walkway defects.

[0036] 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 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 claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for intelligent detection of pedestrian walkways based on 3D point clouds, characterized in that, include: The original laser point cloud data of the sidewalk collected by the mobile mapping system is acquired, and the original laser point cloud data is denoised and spatially downsampled to generate a point cloud dataset to be detected. Based on the local normal vector distribution characteristics of the point cloud dataset to be detected, the point clouds of roadside vegetation and building facades are removed, and the point cloud of the sidewalk surface is extracted. The point cloud of the pedestrian road surface is divided into blocks using a sliding window mechanism. A local design elevation surface for each block is constructed using a polynomial surface fitting algorithm, and the elevation residual of the point cloud relative to the local design elevation surface is calculated. By fusing the elevation residual with the laser reflection intensity information of the point cloud dataset to be detected, a multidimensional feature vector map containing geometric and texture features is constructed. Based on the multidimensional feature vector map, a multi-scale morphological reconstruction algorithm is used to extract initial seed points for pavement distress, and an improved watershed algorithm is used to perform regional growth on the initial seed points to obtain independent distress regions. Calculate the geometric quantitative index of the independent disease area, and combine it with the preset road maintenance standard threshold to determine the disease level and output a test report.

2. The intelligent pedestrian detection method based on three-dimensional point clouds according to claim 1, characterized in that, Based on the local normal vector distribution characteristics of the point cloud dataset to be detected, the point clouds of roadside vegetation and building facades are removed, and the point cloud of the sidewalk surface is extracted, including: For each point in the point cloud dataset to be detected, a local neighborhood is constructed, the covariance matrix of the points in the neighborhood is calculated and the eigenvalues ​​are solved to obtain the normal vector tilt angle of each point. Set a threshold for the tilt angle of the ground point normal vector, and mark points whose tilt angle is less than the threshold as candidate ground points; Perform connectivity analysis on the candidate ground points and calculate the flatness index of each connectivity component; If the flatness index is lower than the preset flatness threshold, the connected region is determined to be a noise or non-road area and is removed, while the remaining point cloud is retained as the pedestrian road surface point cloud.

3. The intelligent pedestrian walkway detection method based on three-dimensional point clouds according to claim 1, characterized in that, The pedestrian walkway point cloud is divided into blocks using a sliding window mechanism, and a local design elevation surface for each block is constructed using a polynomial surface fitting algorithm, including: Set a fixed length of sliding window along the direction of the sidewalk, and set the overlap rate between adjacent windows; Within each sliding window, a random sampling consensus algorithm is used to iteratively fit the planar model and remove outlier points within the window; The least squares method is used to fit the point cloud within the window after removing outliers to a biquadratic polynomial surface to obtain the local design elevation surface equation corresponding to the window. A weighted smooth transition is applied to the edge regions of adjacent sliding windows, and the local design elevation surfaces of all windows are spliced ​​together to generate a continuous global-local design elevation surface.

4. The intelligent pedestrian detection method based on three-dimensional point clouds according to claim 1, characterized in that, By fusing the elevation residual with the laser reflection intensity information of the point cloud dataset to be detected, a multi-dimensional feature vector map containing geometric and texture features is constructed, including: The point cloud dataset to be detected is projected onto a two-dimensional grid plane to generate an elevation residual grid map and a reflection intensity grid map. The elevation residual raster is normalized to highlight areas of negative elevation anomalies; Histogram equalization is performed on the reflection intensity grid to enhance the contrast of crack texture; The processed elevation residual raster image is used as the first channel, and the processed reflection intensity raster image is used as the second channel. The channels are cascaded to construct a dual-channel multidimensional feature vector map.

5. The intelligent pedestrian detection method based on three-dimensional point clouds according to claim 1, characterized in that, Based on the aforementioned multidimensional feature vector map, a multi-scale morphological reconstruction algorithm is used to extract initial seed points for pavement distresses, including: Define circular structural elements of different sizes, and reconstruct the geometric feature channels in the multidimensional feature vector map by performing an opening operation to obtain the background estimation image; The difference between the geometric feature channel and the background estimation image is calculated to obtain a geometric saliency map; A top-hat transform is performed on the texture feature channels in the multidimensional feature vector map to obtain a texture saliency map. The geometric saliency map and the texture saliency map are weighted and fused to obtain the disease saliency map; The local maximum detection operator is applied to the saliency map of the road surface defects to extract the pixel extreme points as the initial seed points of the road surface defects.

6. The intelligent pedestrian walkway detection method based on three-dimensional point clouds according to claim 5, characterized in that, An improved watershed algorithm is used to perform region growing on the initial seed point to obtain independent disease regions, including: Calculate the gradient magnitude image of the multidimensional feature vector map, and use the gradient magnitude image as the topological topography of the watershed algorithm; The initial seed point is marked as the lowest point of the catchment basin; The process of simulating water level rise is carried out by flooding the topology starting from the initial seed point. When the water from different catchment basins is about to converge, a water-blocking dam is constructed. Traverse the topology until all areas above the initial seed point are submerged, extract the pixel coordinates of the water-blocking dam, and form a closed boundary contour of the disease. Closed contours with an area smaller than the preset noise area threshold are removed to obtain the final independent disease area.

7. The intelligent pedestrian detection method based on three-dimensional point clouds according to claim 1, characterized in that, Calculate the geometric quantitative indicators of the independent diseased areas, and determine the disease level by combining them with preset road maintenance standard thresholds, including: Extract the three-dimensional coordinates of all point clouds within the independent disease area, and calculate the projected area, maximum length, and maximum width of the area on the horizontal plane; Calculate the maximum and average depth of the point cloud within the independent disease area relative to the local design elevation surface; If the projected area is less than the first area threshold and the maximum width is less than the crack width threshold, then the disease type is determined to be a transverse or longitudinal crack. If the projected area is greater than the second area threshold and the average depth is greater than the pit depth threshold, then the disease type is determined to be a pit or subsidence. Based on the identified disease type and corresponding geometric quantitative index value, the damage level is determined by matching the damage grading standard in the road maintenance specification.

8. The intelligent pedestrian detection method based on three-dimensional point clouds according to claim 1, characterized in that, The method further includes: Obtain the position information of the centroid of the independent diseased area in the global coordinate system; Read the positioning and attitude data synchronously collected by the mobile mapping system and map the independent disease area onto the pedestrian walkway geographic information map; A digital maintenance work order containing disease type, disease level, geometric quantitative indicators and geographical location information is generated and sent to the terminal display device.

9. The intelligent pedestrian detection method based on three-dimensional point clouds according to claim 1, characterized in that, After acquiring the raw laser point cloud data of the sidewalk collected by the mobile mapping system, the process also includes a point cloud enhancement step based on spatiotemporal consistency: Obtain the current frame's point cloud dataset to be detected and its corresponding N previous frames' historical point cloud datasets; Based on the pose data of the mobile mapping system, the previous N frames of historical point cloud dataset are registered to the current frame coordinate system. For any target point in the current frame's point cloud dataset, search for a set of matching points corresponding to its spatial location in the previous N frames' historical point cloud dataset. If the number of points in the matching point set meets the preset confidence requirement, then the mean elevation residual between the target point and each point in the matching point set is calculated, and the elevation value of the target point is corrected using the mean elevation residual. If the number of points in the matching point set does not meet the preset confidence requirement, then the elevation value of the target point remains unchanged; The corrected point cloud dataset is used to perform subsequent steps.

10. A pedestrian walkway intelligent detection system based on three-dimensional point clouds, characterized in that, To implement the intelligent pedestrian detection method based on three-dimensional point clouds as described in claims 1-9, the method includes: The data acquisition module acquires the original laser point cloud data of the sidewalk collected by the mobile mapping system, and performs noise reduction and spatial downsampling processing on the original laser point cloud data to generate a point cloud dataset to be detected. The road surface extraction module, based on the local normal vector distribution characteristics of the point cloud dataset to be detected, removes the point clouds of roadside vegetation and building facades, and extracts the point cloud of the sidewalk road surface. The reference surface construction module uses a sliding window mechanism to divide the pedestrian pavement point cloud into blocks, uses a polynomial surface fitting algorithm to construct a local design elevation surface for each block, and calculates the elevation residual of the point cloud relative to the local design elevation surface. The feature fusion module fuses the elevation residual with the laser reflection intensity information of the point cloud dataset to be detected, and constructs a multi-dimensional feature vector map containing geometric features and texture features; The disease segmentation module, based on the multi-dimensional feature vector map, uses a multi-scale morphological reconstruction algorithm to extract initial seed points for road surface diseases, and uses an improved watershed algorithm to perform regional growth on the initial seed points to obtain independent disease regions. The result determination module calculates the geometric quantitative indicators of the independent disease area, and determines the disease level by combining them with the preset road maintenance standard threshold and outputs a test report.

Images