Petrochemical scene-based multi-line laser radar ground pit detection method and device

Abstract

The present application provides a kind of based on petrochemical scene's multi-line laser radar ground pit detection method and device, and explosion-proof inspection robot is periodically scanned to obtain the original three-dimensional point cloud data set of surrounding environment, carries out multi-stage adaptive filtering, obtains global point cloud;Based on attitude compensation extraction ground candidate area;Introduce curvature weighted sampling and multi-constraint screening mechanism, separate ground point cloud;According to the average height of ground and the height deviation threshold of pre-set, separate the candidate point cloud set containing pit point;Carry out clustering, obtain several candidate point cloud clusters;According to average curvature, height variation, point cloud density, calculate pit score;According to pit score, select pit cluster.The present application highly matches the multi-traffic inspection demand of explosion-proof inspection robot in petroleum and chemical industry, simultaneously solves the three major core pain points of poor environmental adaptability, high computing power requirement and weak generalization ability of traditional scheme, and improves the robustness of pit detection by constructing pit score model.

Description

Technical Field

[0001] This invention belongs to the field of robotics technology, and in particular relates to a method and device for detecting ground potholes using multi-line lidar in petrochemical scenarios. Background Technology

[0002] In high-risk areas such as petrochemical plants where flammable and explosive materials are present, explosion-proof inspection robots are gradually replacing manual labor for routine inspections. However, the ground conditions in these areas are extremely complex, often riddled with potholes and cracks caused by long-term heavy-load compaction, posing a significant challenge to the autonomous movement of explosion-proof inspection robots. (1) Lack of decision-making ability: Most existing robots can only detect obstacles and stop, lacking quantitative judgment ability and not knowing how to deal with obstacles in front of them, such as whether to "cross over" or "go around" potholes.

[0003] (2) Extremely high safety risks: First, there is a risk that the robot may accidentally enter a pothole, which may cause the robot chassis to bottom out, the motor to be damaged, or even the robot to overturn, causing secondary accidents in dangerous areas; second, due to explosion-proof requirements, the robot cannot carry or be equipped with high-power, high-heat, and heavy computing hardware, making it difficult to run complex algorithms and accurately calculate pothole data and then autonomously determine whether to cross the obstacle or detour.

[0004] (3) Extremely high fault tolerance cost: High-risk chemical site environment is dangerous and space is limited. If the robot gets stuck or breaks down due to misjudgment, it is difficult and costly for manual rescue to enter the site.

[0005] To address the challenges posed by complex ground conditions, existing technologies employ methods such as traditional point cloud analysis, multi-sensor fusion, and deep learning-based computer vision to detect ground potholes. However, these technologies still have some limitations when applied to the identification of ground potholes in petrochemical areas by explosion-proof inspection robots: I. Traditional point cloud solutions can only handle flat roads and other similar scenarios. They rely on simple height separation logic and are completely ineffective on uneven surfaces such as gravel roads and brick roads in petrochemical areas.

[0006] Second, the multi-sensor fusion solution, where multiple sensors (such as sound wave + camera + light sensor) are inside the enclosed body of the explosion-proof robot, still faces problems such as multi-source data synchronization delay and poor compatibility of sensor explosion-proof modification. In addition, the data processing link is long and the latency is high, which cannot meet the needs of explosion-proof inspection robots for real-time obstacle avoidance and early warning.

[0007] Third, deep learning-based computer vision solutions (such as the YOLO series and U-Net) are easily interfered with by the texture of gravel roads and gaps in brick roads, mistakenly identifying normal road textures as potholes; moreover, they are greatly affected by lighting (backlight in petrochemical areas, strong light inspection at night) and water accumulation on rainy days, shadows are easily misjudged as potholes, and water reflections can directly lead to detection failure; when the explosion-proof inspection robot is moving and inspecting, high-speed travel will cause the computer vision solution to blur the image, and vehicle bumps will cause image jitter, requiring additional vibration compensation algorithms, which not only increases the computational complexity, but also consumes the limited computing power of the explosion-proof robot.

[0008] In addition, deep learning models are highly dependent on large-scale, scenario-specific labeled datasets. Existing datasets are mostly concentrated on municipal roads and highways. When they are migrated to mixed road conditions such as asphalt roads, gravel roads, and brick roads in petrochemical areas, the accuracy will drop significantly. Furthermore, deep learning models have a large number of parameters and require the combination of optimization techniques such as Dice loss and Focal loss, which places extremely high demands on the computing power of the terminal. Multi-sensor fusion solutions need to handle data synchronization and registration, which further increases the computing power burden and makes it difficult to achieve real-time detection on the embedded computing power platform of explosion-proof robots. Summary of the Invention

[0009] This invention proposes a multi-line lidar ground pothole detection method and device based on petrochemical scenarios. By relying on the technical approach of multi-dimensional geometric feature recognition, it is highly matched with the multi-condition inspection needs of petrochemical explosion-proof inspection robots. At the same time, it solves the three core pain points of traditional solutions: poor environmental adaptability, high computing power requirements, and weak generalization ability. Furthermore, by constructing a pothole scoring model, the robustness of pothole detection is improved.

[0010] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A multi-line lidar method for detecting ground potholes in petrochemical environments includes: S1. The explosion-proof inspection robot periodically scans to acquire the original three-dimensional point cloud data set of the surrounding environment. The laser reflection intensity information is built into the point cloud data structure. A KD-Tree index structure is established for the point cloud data. S2. Perform multi-level adaptive filtering on the original 3D point cloud data to obtain the global point cloud; the multi-level adaptive filtering includes distance-adaptive statistical filtering and scan angle-compensated radius filtering. S3. Extract ground candidate regions from the global point cloud based on the attitude compensation of the explosion-proof inspection robot; S4. Curvature-weighted sampling and multi-constraint screening mechanisms are introduced to perform curvature-guided multi-constraint RANSAC ground fitting on the point cloud of ground candidate regions to separate the ground point cloud. S5. Calculate the average ground height based on the ground point cloud; in the point cloud within a certain range forward of the ground point cloud, separate a candidate point cloud set containing potholes based on the average ground height and a preset height deviation threshold. S6. Based on preset spatial distance thresholds and height thresholds, cluster the candidate point cloud set to obtain several candidate point cloud clusters; S7. Calculate the pothole score based on the average curvature, height change, and point cloud density of each candidate point cloud cluster; select the pothole cluster based on the pothole score, and output the pothole cluster as the pothole point cloud.

[0011] Furthermore, in step S1, when periodically scanning to acquire the original three-dimensional point cloud data set of the surrounding environment, the explosion-proof inspection robot also performs dynamic field-of-view cropping. Based on the robot's current driving speed, it dynamically adjusts the acquisition range and removes sparse point clouds outside the field of view and those that are too far away.

[0012] Furthermore, in step S2, the distance-adaptive statistical filtering includes: for any point in the original 3D point cloud data, searching for its k nearest neighbor set in the KD-Tree, calculating the neighborhood average distance and neighborhood standard deviation; and calculating the adaptive relaxation coefficient of the point based on the distance from the point to the lidar origin; and determining whether the point is outlier noise based on the neighborhood average distance, neighborhood standard deviation, and the adaptive relaxation coefficient of the point, and if so, removing it.

[0013] Furthermore, in step S2, the radius filtering based on scan angle compensation includes: After distance-adaptive statistical filtering, an adaptive search radius is defined for each point based on the scanning angle and angle compensation coefficient of the LiDAR. The number of neighboring points and the density of the neighboring point cloud within the adaptive search radius of each point are counted. Points with fewer than a preset minimum number of neighboring points or a lower than a preset minimum density are removed.

[0014] Furthermore, step S3 includes: introducing the attitude information of the explosion-proof inspection robot's IMU, using the pitch and roll angle information therein, and combining it with a preset attitude compensation coefficient to compensate the reference ground range, thereby obtaining a ground candidate height range as a ground candidate area.

[0015] Furthermore, step S4 includes: constructing a neighborhood covariance matrix for ground candidate points in the ground candidate region, then performing eigenvalue decomposition, and calculating the curvature of the ground candidate point based on the eigenvalues; The sampling probability is calculated based on curvature. The smaller the curvature, the higher the sampling probability, which concentrates the sampling points in a flat area. When performing RANSAC ground fitting, fitting samples are formed through the sampling points, a fitting plane is formed, and the fitted plane is checked for normal constraints and roughness constraints. The plane with the most interior points and satisfying the constraints is selected as the ground point cloud.

[0016] Furthermore, in step S5, the height difference between each point in the candidate point cloud set and the average height of the ground is less than the height deviation threshold, and the height deviation threshold is adjusted according to the obstacle-crossing capability of the explosion-proof inspection robot.

[0017] Furthermore, step S6 includes: for every two points in the candidate point cloud set, if the spatial distance between them is less than the spatial distance threshold and the height difference between them is less than the height threshold, then they are clustered into a cluster; thereby obtaining several candidate point cloud clusters.

[0018] Furthermore, step S7 includes: calculating the weighted sum of the average curvature, height change, and point density of each candidate point cloud cluster to obtain the pit score of the point cloud cluster.

[0019] In another aspect, the present invention also proposes a multi-line lidar ground pothole detection device based on petrochemical scenarios, comprising: Data acquisition module: used by the explosion-proof inspection robot to periodically scan and acquire the original 3D point cloud data set of the surrounding environment. The point cloud data structure has built-in laser reflection intensity information; and a KD-Tree index structure is established for the point cloud data. The preprocessing module performs multi-level adaptive filtering on the raw 3D point cloud data to obtain a global point cloud. This multi-level adaptive filtering includes distance-adaptive statistical filtering and radius filtering based on scan angle compensation. Ground candidate regions are extracted from the global point cloud based on the posture compensation of the explosion-proof inspection robot. A curvature-weighted sampling and multi-constraint screening mechanism is introduced to perform curvature-guided multi-constraint RANSAC ground fitting on the point cloud of the ground candidate regions, separating the ground point cloud. The average ground height is calculated based on the ground point cloud. Within a certain forward range of the ground point cloud, a candidate point cloud set containing potholes is separated based on the average ground height and a preset height deviation threshold. Pothole identification module: Based on preset spatial distance thresholds and height thresholds, the candidate point cloud set is clustered to obtain several candidate point cloud clusters; a pothole score is calculated based on the average curvature, height change, and point cloud density of each candidate point cloud cluster; a pothole cluster is selected based on the pothole score, and the pothole cluster is output as a pothole point cloud.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention designs a multi-dimensional geometric feature recognition technology path, which is highly compatible with the multi-condition inspection needs of explosion-proof inspection robots in petrochemical scenarios. At the same time, it solves the three core pain points of traditional solutions: poor environmental adaptability, high computing power requirements, and weak generalization ability. Furthermore, it constructs a pothole scoring model that considers spatial distance, height change, local curvature, and point cloud density to detect and identify potholes, avoiding the problem of traditional Euclidean clustering mistaking slopes or obstacles for potholes, thus improving the accuracy and robustness of pothole detection. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the structure of Embodiment 2 of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0023] The present invention will be further described below with reference to specific embodiments and accompanying drawings.

[0024] Example 1: This embodiment provides a method for detecting ground potholes using a multi-line lidar system in a petrochemical environment, such as... Figure 1 As shown, it includes: Step 1: Point cloud data acquisition and multi-source information initialization: During the inspection or autonomous movement of the explosion-proof inspection robot, it first uses a multi-line 3D LiDAR mounted on it to periodically scan the surrounding environment, thereby obtaining the original 3D point cloud data set at the current moment: ; in: Indicates time t The acquired point cloud set; Represents the first in the point cloud set i One point; Indicates the first i The three-dimensional spatial coordinates of a point in the robot coordinate system; Indicates the first i The laser reflection intensity corresponding to each point.

[0025] Because metal pipes, storage tanks, and reflective structures are common in petrochemical environments, their reflective properties differ significantly from those of ground materials. Therefore, this embodiment retains laser reflection intensity information in the point cloud data structure for subsequent structure differentiation. Simultaneously, to support subsequent real-time calculations, a KD-Tree index structure is established on the original 3D point cloud to improve neighborhood query efficiency.

[0026] When periodically scanning to acquire the raw 3D point cloud data set for the current moment, dynamic field-of-view clipping is also performed. Based on the current travel speed of the explosion-proof inspection robot, the real-time acquisition range ROI is dynamically adjusted to remove sparse point clouds outside the field of view (FOV) and those too far away, thereby reducing the computational load of subsequent processing. When the speed is high, the robot's inertia is relatively large, and the braking distance will increase accordingly, so it is necessary to consider point cloud data at a long distance; when the speed is low, only point cloud data at a slightly closer distance is considered for calculation, which can reduce the amount of computation.

[0027] The specific process includes: first, obtaining the current speed of the explosion-proof inspection robot. Establish a dynamic acquisition range model: ; ; in The robot's rated speed is configured according to actual needs. , This refers to the robot's performance at its rated speed. x and y Point cloud distance in the direction is configured according to the actual scenario; for x and y The directional attenuation coefficient is configured empirically based on the fixed value of the actual configuration. Calculated according to the above formula The real-time acquisition range (ROI) is used as the data acquisition area. Then, these two parameters are used as parameters for box filtering to obtain the filtered point cloud data, which serves as the original 3D point cloud data set for use in the preprocessing module.

[0028] Step 2: Multi-stage adaptive filtering with distance-angle coupling: Although the obtained raw 3D point cloud data set has undergone dynamic field-of-view clipping, problems such as sparse point clouds and enhanced noise still occur due to the long-range and large-angle scanning of the LiDAR. If left unprocessed, these issues will directly affect the stability of ground modeling. Therefore, this step performs multi-level filtering on the raw 3D point cloud data, including statistical filtering and radius filtering, and introduces a distance-scan angle coupling mechanism.

[0029] 2.1 Distance-adaptive statistical filtering.

[0030] For any point in the original 3D point cloud data First, search for it in KD-Tree. k Nearest neighbor set: ; in: Point of nearest neighbor set : Number of neighboring points.

[0031] against k Calculate the average distance within a neighborhood formed by the set of nearest neighbors: ; in: Represents the average neighborhood distance. Indicates Euclidean distance; Represents neighboring points; Calculate the neighborhood standard deviation, where Indicates the degree of dispersion of neighborhood distance distribution: .

[0032] After obtaining the average neighborhood distance and neighborhood standard deviation, it is necessary to calculate the distance adaptive coefficient using the following formula: ;in This represents the adaptive relaxation factor. Point Distance to the origin of the lidar This indicates the maximum effective processing distance after adjusting the real-time acquisition range (ROI) and entering the statistical filtering calculation: The distance can be determined based on the robot's current speed, braking distance, obstacle-crossing ability, and safe obstacle avoidance distance. In the petrochemical explosion-proof robot's road pothole detection scenario, a distance of 4m to 8m is preferred. In this formula, =2.5 is the classic default value, but it is easily misjudged as noise and deleted, resulting in incomplete contours, broken edges, loss of details, and thinner models. Therefore, the formula is followed by... Adaptive improvements are made because, after adjusting the real-time acquisition range ROI in step 1, it exceeds... The distant sparse point cloud has been removed and will not participate in subsequent statistical filtering calculations. Therefore, The value of is limited to the range of 0 to 1. The value ranges from 2.5 to 3.0, and is only used to slightly compensate for the sparsity of point clouds within the effective detection distance range, without constituting an unlimited threshold amplification of distant noise points. Through adaptive improvement, it can still maintain the dense area of ​​the point cloud and strictly eliminate isolated flying points and scattered noise.

[0033] When satisfied If the condition is met, the point is determined to be outlier noise and is removed.

[0034] 2.2 Radius filtering based on scan angle compensation.

[0035] Statistical filtering is used to handle the overall sparsity changes in point clouds as the detection distance increases; this step is used to handle the local sparsity changes within the same distance range caused by the laser beam deviating from the center of the effective field of view of the ROI. The two steps compensate for different things.

[0036] set up Indicates the first point in the point cloud set Each point corresponds to a unit direction vector of the laser beam. The unit direction vector representing the radar's preset center line of sight is used to uniquely determine the scan angle by defining the angle between the two. : ; in, The unit is rad.

[0037] The effective scanning angle is determined based on the effective field of view of the real-time acquisition area ROI and the field of view of the lidar. For angles exceeding the effective scanning angle Point clouds are discarded before entering radius filtering to avoid the cosine denominator approaching zero. An adaptive search radius is set based on the scan angle. ; in, Indicates the first point in the point cloud set The neighborhood search radius of each point is expressed in meters. This represents the basic search radius near the central line of sight, in meters. This represents the angle compensation weight, which is a dimensionless coefficient. and These represent the preset minimum search radius and the preset maximum search radius, respectively, in meters (m). The parameters satisfy: ; In the formula The approximate compensation ratio used to describe the change in point spacing caused by an increase in the scanning angle, under the conditions of an effective scanning angle range and an approximate local surface, is not a universal physical law that is not subject to any conditions.

[0038] in, Represents the amplitude limiting function: ; Point cloud frame at the current moment Statistical search radius The number of neighboring points within the range, excluding the point to be judged itself: ; in, Indicates the first point in the point cloud set The number of neighboring points of each point; Represents the first in the point cloud set j One point; To avoid simply increasing the search radius to satisfy the fixed number of neighborhood points condition for sparse noise, the th... Neighborhood point cloud density of each point : ; When satisfied or At that time, the point Identify and remove points that are isolated or locally sparse noise; otherwise, retain the point. , These are the minimum quantity value and the minimum density value, respectively, both of which are empirical constants based on the laser radar's beam settings. , , , , and The calibration is based on the lidar angular resolution, installation height, dynamic ROI range, and measured point cloud density on flat ground.

[0039] After radius filtering, a filtered point cloud is obtained for subsequent ground candidate region extraction.

[0040] Step 3: Ground candidate region extraction for attitude compensation: After filtering, to reduce the computational complexity of subsequent ground fitting, candidate ground regions need to be extracted from the filtered point cloud. Considering the inevitable pitch angle changes when the explosion-proof inspection robot moves in a petrochemical environment, IMU attitude information is introduced for compensation. IMU attitude information refers to the real-time three-dimensional attitude angles of the robot body calculated by the IMU, including roll angle, pitch angle, and yaw angle. The IMU outputs the roll angle, pitch angle, and yaw angle of the robot body, and based on this, the attitude compensation rotation matrix from the robot coordinate system to the gravity-aligned coordinate system is obtained. For points in a filtered point cloud After attitude compensation, it is represented as : ; The coordinate system is represented as: Roll and pitch angles are used to correct for the offset of the ground normal relative to the vertical direction; yaw angle mainly affects the horizontal direction and is used during global coordinate transformation, without changing the upper and lower limits of the ground candidate altitude independently.

[0041] Set the ground candidate height range in the gravity-aligned coordinate system. : ; in, and These represent the lower and upper limits of the candidate ground height relative to the robot's installation reference, respectively. These limits can be calibrated based on the LiDAR installation height, the robot's permissible body posture range, and measured ground undulations. This avoids neglecting the impact of roll attitude changes on ground height distribution while only using pitch angle compensation.

[0042] The calculated range of candidate ground heights is used as the candidate ground region.

[0043] Step 4: Curvature-guided multi-constraint RANSAC ground fitting: Based on the ground candidate points in the ground candidate region, this embodiment improves the traditional RANSAC method by introducing curvature-weighted sampling and multi-constraint screening mechanism to enhance the robustness of ground segmentation.

[0044] 4.1 Curvature Calculation: Construct the neighborhood covariance matrix for candidate ground points: ; in Represents the covariance matrix. Representing neighborhood points, It represents the neighborhood centroid.

[0045] Then eigenvalue decomposition is performed, where Eigenvalues: .

[0046] Curvature derived from eigenvalues : ; if Smaller points are found on flat ground, while larger points are found in sharp or noisy areas.

[0047] 4.2 Curvature-guided sampling: The purpose of this step is to guide the sampling probability by curvature, so that the sampling points are more concentrated in the flat area, thereby improving the convergence speed and stability of RANSAC fitting.

[0048] ; in, Indicates the sampling probability. It represents a very small constant.

[0049] 4.3 Planar Model and Constraint Selection: The RANSAC method is used, sampling 5 points in each round, and selecting the points with high sampling probability in step 4.2 to form the fitting samples, thereby reducing the influence of outliers.

[0050] The plane equation is used to fit the plane: ; Where the coefficient a,b,c The normal vectors constituting the plane are solved by least squares fitting of the point set; the parameter d represents the position offset parameter of the plane relative to the origin of the coordinate system.

[0051] Then, the fitted plane is subjected to constraint checks, including: First, apply normal constraints: ; The normal constraint is executed after attitude compensation is completed in step 3. The Z-axis of the gravity-aligned coordinate system represents the vertical direction, so the normal constraint is used to determine whether the normal of the candidate plane is close to the vertical direction; attitude compensation can prevent the robot from misjudging the real ground as non-ground when it tilts.

[0052] Then use roughness constraints: ; roughness R The smaller the size, the flatter the surface. R The larger the surface, the more likely it is to be curved or noisy.

[0053] Then, the number of inlier points is counted. In each round of RANSAC fitting, the number of points that satisfy the normal and roughness constraints is counted. After multiple rounds of RANSAC fitting, the fitted plane with the most inlier points that satisfy the constraints is selected as the final ground point cloud. .

[0054] Step 5: Obtain a candidate point cloud set containing potholes: Based on the ground point cloud separated in step 4 Calculate the average ground height ; Then, select a point cloud within a certain range in front of the explosion-proof inspection robot. For example, if we set the x-axis as the direction the robot is moving forward, then, for instance, we can use... The point cloud within a three-meter radius in front of the explosion-proof inspection robot was separated. ; average ground height For reference, set a height deviation threshold. This threshold can be adjusted based on the actual obstacle-crossing capability of the explosion-proof inspection robot. Based on the average ground height and height deviation threshold, a candidate point cloud set containing potholes is obtained. : ; The candidate point cloud set obtained by this formula includes both pits and depressions below the ground level and points above and near the ground level. Points significantly above the ground level (with a height difference greater than the average ground level threshold) are excluded. This is because the ground in petrochemical areas is not perfectly flat, and to extract as many pits and depressions as possible, some points near the ground level are included. However, these points will be removed in subsequent clustering and scoring steps.

[0055] Step 6: Topological Constraint Clustering: First, set the spatial distance threshold. and height threshold Then, for every two points in the candidate point cloud set... Determine if it satisfies , The spatial distance between two points The height difference between two points is used. Only spatially adjacent points with similar heights are grouped into the same cluster to avoid erroneous merging of points with vertical structures (such as slopes or obstacles), ultimately resulting in several candidate point cloud clusters. .

[0056] Step 7: Multi-feature pit scoring: For each candidate point cloud cluster Calculate pothole score : ; in To obtain the normalized value of the average curvature of each candidate point cloud cluster, first, the curvature of each point in the candidate point cloud cluster is obtained according to step 4.1. Then, the average value is calculated to obtain the average curvature of the candidate point cloud cluster. Finally, the average curvature is normalized by combining the maximum curvature and the minimum curvature in the candidate point cloud cluster.

[0057] This is a normalized value for the height change of each candidate point cloud cluster. The height change of the candidate point cloud cluster is obtained by the height difference between the highest and lowest points. The larger the value, the more concave it is.

[0058] The point cloud density anomaly normalization value is calculated for each candidate point cloud cluster. Based on the number of points contained in the candidate point cloud cluster and its projection range on the locally fitted ground, the point cloud density characteristics of the candidate point cloud cluster are determined. Then, based on the reference density of normal ground point clouds within the corresponding distance range and scan angle range, the point cloud density characteristics are normalized to obtain the point cloud density anomaly normalization value. Low density may indicate the edge of a pit.

[0059] The weighting coefficients for the normalized values ​​of the average curvature, the normalized values ​​of the height variation, and the normalized values ​​of the point cloud density anomalies are respectively, and can be adjusted according to actual needs.

[0060] Set a rating threshold ,according to The selection criteria are used to select pothole clusters, and all points in the pothole clusters can be output as pothole point clouds, which are the detected potholes. Since the actual height of the points above and near the ground included in step 5 is relatively small, the average curvature and height difference after clustering are relatively small, and the resulting pothole scores are also small, so they will not be selected as pothole clusters.

[0061] This method considers density, height, and curvature simultaneously, avoiding the problem of traditional Euclidean clustering mistaking slopes or obstacles for potholes, thus improving the accuracy of pothole detection.

[0062] After selecting the pothole clusters, a real-time inspection report can be generated, including the geographical coordinates of the pothole center, surface area, maximum depth, and hazard level assessment. The identified pothole boundaries are overlaid on a real-time 3D map in the background, in the form of bounding boxes or heat maps. Because the explosion-proof inspection robot itself has some off-road capability, it can directly pass through smaller potholes. If the pothole depth exceeds the preset obstacle avoidance threshold, the module immediately sends a deceleration command to the underlying actuator and calculates a local obstacle avoidance path, guiding the autonomous driving system to make a detour or crossing decision.

[0063] This embodiment is highly compatible with the multi-condition inspection requirements of explosion-proof inspection robots in petrochemical scenarios, and solves the three core pain points of traditional solutions: poor environmental adaptability, high computing power requirements, and weak generalization ability.

[0064] Example 2: This embodiment proposes a multi-line lidar ground pothole detection device based on a petrochemical scenario, such as... Figure 2 As shown, it includes: Data acquisition module: used by the explosion-proof inspection robot to periodically scan and acquire the original 3D point cloud data set of the surrounding environment. The point cloud data structure has built-in laser reflection intensity information; and a KD-Tree index structure is established for the point cloud data. The preprocessing module performs multi-level adaptive filtering on the raw 3D point cloud data to obtain a global point cloud. This multi-level adaptive filtering includes distance-adaptive statistical filtering and radius filtering based on scan angle compensation. Ground candidate regions are extracted from the global point cloud based on the posture compensation of the explosion-proof inspection robot. A curvature-weighted sampling and multi-constraint screening mechanism is introduced to perform curvature-guided multi-constraint RANSAC ground fitting on the point cloud of the ground candidate regions, separating the ground point cloud. The average ground height is calculated based on the ground point cloud. Within a certain forward range of the ground point cloud, a candidate point cloud set containing potholes is separated based on the average ground height and a preset height deviation threshold. Pothole identification module: Based on preset spatial distance thresholds and height thresholds, the candidate point cloud set is clustered to obtain several candidate point cloud clusters; a pothole score is calculated based on the average curvature, height change, and point cloud density of each candidate point cloud cluster; a pothole cluster is selected based on the pothole score, and the pothole cluster is output as a pothole point cloud.

[0065] It also includes an output module: After selecting a cluster of potholes, it can generate a real-time inspection report containing the geographical coordinates of the pothole center, surface area, maximum depth, and hazard level assessment. The identified pothole boundaries are overlaid on a real-time 3D map in the background, either as bounding boxes or heatmaps. Because the explosion-proof inspection robot itself has some off-road capability, it can directly pass through smaller potholes. If the pothole depth exceeds a preset obstacle avoidance threshold, this module immediately sends a deceleration command to the underlying actuator and calculates a local obstacle avoidance path, guiding the autonomous driving system to make a detour or obstacle-crossing decision.

[0066] In the data acquisition module, when periodically scanning to acquire the original three-dimensional point cloud data set of the surrounding environment, the explosion-proof inspection robot also performs dynamic field-of-view clipping. Based on the robot's current driving speed, it dynamically adjusts the acquisition range and removes sparse point clouds outside the field of view and those that are too far away.

[0067] In the preprocessing module, the distance-adaptive statistical filtering includes: for any point in the original 3D point cloud data, searching for its k nearest neighbor set in the KD-Tree, calculating the average neighborhood distance and neighborhood standard deviation; calculating the adaptive relaxation coefficient of the point based on the distance from the point to the lidar origin; and determining whether the point is outlier noise based on the average neighborhood distance, neighborhood standard deviation, and the adaptive relaxation coefficient of the point, and removing it if so.

[0068] In the preprocessing module, the radius filtering based on scan angle compensation includes: After distance-adaptive statistical filtering, an adaptive search radius is defined for each point based on the scanning angle and angle compensation coefficient of the LiDAR. The number of neighboring points and the density of the neighboring point cloud within the adaptive search radius of each point are counted. Points with fewer than a preset minimum number of neighboring points or a lower than a preset minimum density are removed.

[0069] The preprocessing module also includes: introducing the attitude information of the explosion-proof inspection robot's IMU, using the pitch and roll angle information therein, and combining it with the preset attitude compensation coefficient, to compensate the reference ground range and obtain the ground candidate height range as the ground candidate area.

[0070] The preprocessing module also includes: constructing a neighborhood covariance matrix for ground candidate points in the ground candidate region, then performing eigenvalue decomposition, and calculating the curvature of the ground candidate point based on the eigenvalues; The sampling probability is calculated based on curvature. The smaller the curvature, the higher the sampling probability, which concentrates the sampling points in a flat area. When performing RANSAC ground fitting, fitting samples are formed through the sampling points, a fitting plane is formed, and the fitted plane is checked for normal constraints and roughness constraints. The plane with the most interior points and satisfying the constraints is selected as the ground point cloud.

[0071] The preprocessing module also includes: the height difference between each point in the candidate point cloud set and the average height of the ground is less than the height deviation threshold, and the height deviation threshold is adjusted according to the obstacle-crossing capability of the explosion-proof inspection robot.

[0072] The pothole identification module includes: for every two points in the candidate point cloud set, if the spatial distance between them is less than the spatial distance threshold and the height difference between them is less than the height threshold, then they are clustered into a cluster; thereby obtaining several candidate point cloud clusters.

[0073] The pothole identification module also includes: calculating the weighted sum of the average curvature, height change and point density of each candidate point cloud cluster to obtain the pothole score of the point cloud cluster.

[0074] The multi-line lidar ground pothole detection device based on petrochemical scenarios proposed in this embodiment can achieve the same technical effect as the multi-line lidar ground pothole detection method based on petrochemical scenarios described in Embodiment 1.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-line lidar method for detecting ground potholes in petrochemical scenarios, characterized in that, include: S1. The explosion-proof inspection robot periodically scans to acquire the original three-dimensional point cloud data set of the surrounding environment. The laser reflection intensity information is built into the point cloud data structure. A KD-Tree index structure is established for the point cloud data. S2. Perform multi-level adaptive filtering on the original 3D point cloud data to obtain the global point cloud; the multi-level adaptive filtering includes distance-adaptive statistical filtering and scan angle-compensated radius filtering. S3. Extract ground candidate regions from the global point cloud based on the attitude compensation of the explosion-proof inspection robot; S4. Curvature-weighted sampling and multi-constraint screening mechanisms are introduced to perform curvature-guided multi-constraint RANSAC ground fitting on the point cloud of ground candidate regions to separate the ground point cloud. S5. Calculate the average ground height based on the ground point cloud; in the point cloud within a certain range forward of the ground point cloud, separate a candidate point cloud set containing potholes based on the average ground height and a preset height deviation threshold. S6. Based on preset spatial distance thresholds and height thresholds, cluster the candidate point cloud set to obtain several candidate point cloud clusters; S7. Calculate the pothole score based on the average curvature, height change, and point cloud density of each candidate point cloud cluster; select the pothole cluster based on the pothole score, and output the pothole cluster as the pothole point cloud.

2. The multi-line lidar ground pothole detection method based on petrochemical scenarios according to claim 1, characterized in that, In step S1, while periodically scanning to acquire the original three-dimensional point cloud data set of the surrounding environment, the explosion-proof inspection robot also performs dynamic field-of-view cropping. Based on the robot's current driving speed, it dynamically adjusts the acquisition range and removes sparse point clouds outside the field of view and those that are too far away.

3. The multi-line lidar ground pothole detection method based on petrochemical scenarios according to claim 1, characterized in that, In step S2, the distance-adaptive statistical filtering includes: for any point in the original 3D point cloud data, searching for its k nearest neighbor set in the KD-Tree, calculating the neighborhood average distance and neighborhood standard deviation; and calculating the adaptive relaxation coefficient of the point based on the distance from the point to the lidar origin; and determining whether the point is outlier noise based on the neighborhood average distance, neighborhood standard deviation, and the adaptive relaxation coefficient of the point, and removing it if it is.

4. The multi-line lidar ground pothole detection method based on petrochemical scenarios according to claim 1, characterized in that, In step S2, the radius filtering based on scan angle compensation includes: After distance-adaptive statistical filtering, an adaptive search radius is defined for each point based on the scanning angle and angle compensation coefficient of the LiDAR. The number of neighboring points and the density of the neighboring point cloud within the adaptive search radius of each point are counted. Points with fewer than a preset minimum number of neighboring points or a lower than a preset minimum density are removed.

5. The multi-line lidar ground pothole detection method based on petrochemical scenarios according to claim 1, characterized in that, Step S3 includes: introducing the attitude information of the explosion-proof inspection robot's IMU, using the pitch and roll angle information therein, and combining it with the preset attitude compensation coefficient to compensate the reference ground range, thereby obtaining the ground candidate height range as the ground candidate area.

6. The multi-line lidar ground pothole detection method based on petrochemical scenarios according to claim 1, characterized in that, Step S4 includes: constructing a neighborhood covariance matrix for ground candidate points in the ground candidate region, then performing eigenvalue decomposition, and calculating the curvature of the ground candidate point based on the eigenvalues; The sampling probability is calculated based on curvature. The smaller the curvature, the higher the sampling probability, which concentrates the sampling points in a flat area. When performing RANSAC ground fitting, fitting samples are formed through the sampling points, a fitting plane is formed, and the fitted plane is checked for normal constraints and roughness constraints. The plane with the most interior points and satisfying the constraints is selected as the ground point cloud.

7. The multi-line lidar ground pothole detection method based on petrochemical scenarios according to claim 1, characterized in that, In step S5, the height difference between each point in the candidate point cloud set and the average height of the ground is less than the height deviation threshold, which is adjusted according to the obstacle-crossing capability of the explosion-proof inspection robot.

8. The multi-line lidar ground pothole detection method based on petrochemical scenarios according to claim 1, characterized in that, Step S6 includes: for every two points in the candidate point cloud set, if the spatial distance between them is less than the spatial distance threshold and the height difference between them is less than the height threshold, they are clustered into a cluster; thus obtaining several candidate point cloud clusters.

9. The multi-line lidar ground pothole detection method based on petrochemical scenarios according to claim 1, characterized in that, Step S7 includes: calculating the weighted sum of the average curvature, height change, and point density of each candidate point cloud cluster to obtain the pit score of the point cloud cluster.

10. A multi-line lidar ground pothole detection device for petrochemical scenarios, characterized in that, include: Data acquisition module: used by the explosion-proof inspection robot to periodically scan and acquire the original 3D point cloud data set of the surrounding environment. The point cloud data structure has built-in laser reflection intensity information; and a KD-Tree index structure is established for the point cloud data. Preprocessing module: used to perform multi-level adaptive filtering on raw 3D point cloud data to obtain global point cloud; the multi-level adaptive filtering includes distance-adaptive statistical filtering and scan angle-compensated radius filtering; Ground candidate regions are extracted from the global point cloud based on the posture compensation of the explosion-proof inspection robot; a curvature-weighted sampling and multi-constraint screening mechanism is introduced to perform curvature-guided multi-constraint RANSAC ground fitting on the point cloud of the ground candidate regions to separate the ground point cloud; the ground average height is calculated based on the ground point cloud; in the point cloud within a certain range forward of the ground point cloud, a candidate point cloud set containing potholes is separated based on the ground average height and a preset height deviation threshold. Pothole identification module: Based on preset spatial distance thresholds and height thresholds, the candidate point cloud set is clustered to obtain several candidate point cloud clusters; a pothole score is calculated based on the average curvature, height change, and point cloud density of each candidate point cloud cluster; a pothole cluster is selected based on the pothole score, and the pothole cluster is output as a pothole point cloud.

Images