A method for processing obstacles in a mine roadway based on a solid-state three-dimensional laser radar

Abstract

The application discloses a kind of based on solid three-dimensional laser radar's underground roadway obstacle processing method, it is related to obstacle detection technical field, for monitoring underground roadway obstacle, including obtaining original point cloud data, constructs anti-distortion point cloud set;Establish roadway local coordinate system, anti-distortion point cloud set is converted to roadway local coordinate system in;Build reference envelope model, calculate the spatial deviation of each point cloud data point;Extract abnormal occupancy point, calculate the obstacle saliency defect amount and reference envelope invasion amount of abnormal occupancy point;Through the balance relation of effective coverage and residual time sequence variable, the integral length of candidate integral length is adaptively determined, and the optimal integral length is obtained;All abnormal occupancy points in the optimal integral length are clustered and segmented, and the obstacle point cloud is output, and the data information of the obstacle point cloud is solved, to generate underground roadway obstacle detection result.The application improves the accuracy and stability of underground roadway obstacle detection.

Description

Technical Field

[0001] This invention relates to the field of obstacle detection technology, and more particularly to the field of real-time obstacle monitoring technology in underground roadways, specifically to a method for handling obstacles in underground roadways based on solid-state three-dimensional lidar. Background Technology

[0002] Coal mine underground roadways are characterized by narrow spaces, poor lighting, and strong interference from dust and water mist. They may also contain obstacles such as fallen rocks, stockpiled materials, parked equipment, foreign objects on tracks, and personnel or vehicles obstructing the way. If these obstacles are not detected in time, they can easily affect transportation, inspection, and operational safety, and may even cause collisions, blockages, or production interruptions.

[0003] Existing methods for detecting obstacles in underground mines mainly include manual inspection, video surveillance, two-dimensional laser scanning, and three-dimensional perception methods based on mobile platforms. Manual inspection is inefficient, highly subjective, and difficult to achieve continuous monitoring; video surveillance is easily affected by lighting, dust, water mist, and obstructions, and lacks stability in complex underground environments; two-dimensional laser scanning can only obtain local cross-sectional information, making it difficult to fully reflect the three-dimensional shape and spatial occupancy of obstacles; conventional three-dimensional point cloud detection methods usually rely on the mapping and localization results of vehicles or robots themselves, and their detection accuracy and stability remain limited in long underground tunnels, curved tunnels, weak texture environments, and under conditions of localization drift.

[0004] Furthermore, underground roadways have significant space constraints for passage. Actual safety risks depend not only on the presence of abnormal point clouds, but also on whether obstacles encroach on the effective passage envelope of the roadway, whether they affect vehicle or equipment passage, and their temporal continuity. Existing methods primarily focus on point cloud target identification or distance alarms, lacking a comprehensive assessment of the roadway centerline, local coordinate system, passage envelope, and temporal continuity, which easily leads to false alarms, missed alarms, or inaccurate risk level assessments.

[0005] Therefore, there is an urgent need for a systematic detection method that can adaptively handle motion distortion, establish a local coordinate system for the roadway, optimize the integration time, and accurately identify intruding obstacles, so as to achieve stable identification of roadway obstacles and risk output. Summary of the Invention

[0006] The purpose of this invention is to provide a method for handling obstacles in underground roadways based on solid-state three-dimensional lidar, so as to solve the problems of low detection accuracy and poor detection stability of underground obstacles in the prior art.

[0007] To achieve the above objectives, the present invention provides a method for handling obstacles in underground roadways based on solid-state three-dimensional lidar, comprising: S1. Obtain the original point cloud data with timestamps within the target area, perform motion distortion removal processing on the original point cloud data, and construct an anti-distortion point cloud set; S2. Establish a local coordinate system for the roadway based on the roadway extension direction curve, and transform the anti-distortion point cloud set to the local coordinate system for the roadway to obtain the local three-dimensional coordinates of the point cloud data points in the anti-distortion point cloud set; S3. Construct a reference envelope model of the passage space in the local coordinate system of the passage, and calculate the spatial deviation of each point cloud data point; S4. Extract the abnormal occupied points based on the spatial deviation calculated in S3, and calculate the significant defect amount of the obstacle and the intrusion amount of the reference envelope at the abnormal occupied points; S5. Set the candidate integration duration. Based on the local three-dimensional coordinates of the point cloud data points obtained in S2 and the baseline envelope model constructed in S3, calculate the effective coverage and residual time series variables of the candidate integration duration. Adaptively determine the candidate integration duration that satisfies the balance relationship between the effective coverage and residual time series variables. The smallest candidate integration duration that satisfies the balance relationship is taken as the optimal integration duration. S6. Based on the significant defect quantity of the abnormal occupied points and the baseline envelope intrusion quantity calculated in S4, cluster and segment all abnormal occupied points within the optimal integration time, output the obstacle point cloud, solve the data information of the obstacle point cloud, and generate the obstacle detection results of the underground roadway.

[0008] In S1, the solid-state 3D lidar collects the original point cloud data with timestamps within the target area, extracts all the original point cloud data in the integration time domain before the reference time, and constructs the original point cloud set in the target integration time domain. The original point cloud set includes the original measurement coordinates and sampling time of each point cloud data point. For the original point cloud set in the target integral time domain, a continuous motion model is constructed to map each point cloud data point in the original point cloud set from the sampling time to the reference time in reverse distortion, thus obtaining the reverse distortion point cloud set. The anti-distortion mapping satisfies: ; In the formula, For point cloud data point indexing, Indicates the first Sampling time of each point cloud data point Indicates reference time. Indicates the first The spatial coordinates of a point cloud data point after anti-distortion mapping Indicates the first The original measured coordinates of each point cloud data point; Indicates the sensor at the first The pose matrix relative to the global coordinate system at each sampling time. This represents the pose matrix of the sensor relative to the global coordinate system at the reference time.

[0009] In S2, the curve representing the direction of tunnel extension is: ; In the formula, A curve indicating the direction of the tunnel's extension. For arc length parameters, Represents curve In arc length The three-dimensional coordinates of the curve point at that location. For the set of real numbers, Belongs to; The local 3D coordinates of each point cloud data point in the anti-distortion point cloud set are: ; ; ; ; ; In the formula, Indicates the first Local 3D coordinates of a point cloud data point express The vertical coordinate, express The horizontal coordinate, express The height coordinates, Indicates the first Each point cloud data point is in The corresponding arc length parameter, Represents the transpose of a vector or matrix. , They represent In arc length The horizontal unit direction vector and the vertical unit direction vector at the location; Represents the Euclidean norm. Indicates the search for arc length parameter Make the vector and The Euclidean distance between them is the smallest.

[0010] The baseline envelope model for the passage space in S3 is as follows: ; In the formula, The baseline envelope model representing the passage space of the alleyway. This represents the coordinates of a spatial point in the local coordinate system of the tunnel. express The vertical coordinate, express The horizontal coordinate, express The height coordinates, express The horizontal left boundary function, express The horizontal right boundary function, express The lower boundary function of the height, express The height of the upper boundary function; Extract the reference surface of the tunnel and the reference envelope boundary of the reference envelope model, and calculate the spatial deviation of each point cloud data point. The spatial deviation includes the deviation of the point cloud data point relative to the reference surface of the tunnel and the deviation of the point cloud data point relative to the reference envelope boundary.

[0011] The deviation of the point cloud data points relative to the roadway reference surface is: ; In the formula, Indicates the first The deviation of each point cloud data point from the roadway reference surface. Indicates the reference surface of the tunnel. express arrive European distance, This is a function for calculating Euclidean distance; The deviation of the point cloud data points from the reference envelope boundary is: ; In the formula, Indicates the first The deviation of each point cloud data point from the baseline envelope boundary. Indicates the reference envelope boundary. express arrive The signed distance, This is a function for calculating signed distance; When, it indicates Located inside the baseline envelope model, When, it indicates Located on the boundary of the reference envelope, When, it indicates It is located outside the baseline envelope model.

[0012] In S4, all anomalous occupied points are extracted from the anti-distortion point cloud set, an anomalous point cloud set is constructed, and the feature quantities of the anomalous occupied points are calculated. The feature quantities include the obstacle salience defect quantity and the baseline envelope intrusion quantity. The significant defect quantity of the obstacle is: ; In the formula, Indicates the index of the abnormal occupying point. Indicates the first Significant defect quantity of obstacles at each abnormally occupied point. Indicates the first Local three-dimensional coordinates of the abnormal occupied points; express Compared to The indicator function, when lie in Inside The value is 1, when lie in The outside, The value is 0; The function for finding the maximum value. Indicates the reference surface of the tunnel at Adaptive tolerance at the location, Indicates the first Adaptive weights for each abnormally occupied point; The baseline envelope intrusion is: ; In the formula, Indicates the first The baseline envelope intrusion of each anomalous occupancy point express arrive Euclidean distance.

[0013] The effective coverage of candidate integration duration is: ; In the formula, Indicates the candidate integral duration. express Effective coverage, Indicates the longitudinal integration interval of the tunnel. It is a natural exponential function. To represent the differential, Represents the integral. This represents the differential element when integrating along the longitudinal direction of the tunnel. Indicates will vertical coordinate As an integration variable, it is used for integration calculation along the longitudinal integration interval of the roadway; Indicates that in a given Two-dimensional cross-section of the reference envelope model, Indicates in The differential element when performing double integration with respect to the horizontal and vertical directions. Let be the probability density function. This indicates that the integration time is When, the vertical coordinate is The point cloud data points are in The conditional probability density function within; This represents the logarithmic function with base e.

[0014] The discrete form is: ; In the formula, Indicates the local unit index of the roadway. This indicates that the integration time is At that time, local units of the tunnel The number of valid observation points in the data; To cover saturation scale parameters, This indicates the local units of all roadways. Summation, Indicates a local unit of the roadway The weights; Indicates a local unit of the roadway Normalized coverage Indicates the weighted total coverage. This represents the sum of the weights of all local units in the roadway; The discrete form of the residual time series variable of the candidate integration duration is: ; In the formula, express The residual time-series variables, This represents a set of local units of a roadway. Indicates a local unit of the roadway The residual variable, Indicates a local unit of the roadway The sampling time variable, For covariance, Indicates a local unit in a given roadway Under the conditions, and The square of the covariance; For variance, , These represent local units in a given roadway. Under the conditions, , The variance; It is a preset positive number.

[0015] Construct a balance relationship between effective coverage and residual time series variables, select a positive integral duration that satisfies the balance relationship between effective coverage and residual time series variables, construct a discrete candidate set, and select the minimum value in the discrete candidate set as the optimal integral duration: The balance between effective coverage and residual time series variables is as follows: ; Optimal integration time for: ; In the formula, To find the minimum value of the function, The integral duration index represents the index of the discrete candidate set. Each point duration.

[0016] In S6, based on the significant defect quantity of obstacles at the abnormal occupancy points and the baseline envelope intrusion quantity, all abnormal occupancy points within the optimal integration time are obtained, and an obstacle point cloud set is constructed. : ; In the formula, The threshold for the significant defect quantity of the obstacle. The threshold for the intrusion amount into the passage envelope; Cluster the obstacle point cloud set to obtain at least one obstacle point cloud cluster. Calculate the 3D center position, minimum bounding box size, lateral occupancy width, height occupancy range, longitudinal occupancy length, contour boundary, and distance relative to the roadway traffic axis or planned traffic path for each obstacle point cloud cluster. Calculate the feature dimensions of the corresponding obstacle based on the 3D center position, minimum bounding box size, lateral occupancy width, height occupancy range, longitudinal occupancy length, contour boundary, and distance relative to the roadway traffic axis or planned traffic path of the obstacle point cloud cluster. Clustering and segmenting of obstacle point cloud sets includes spatial connectivity analysis, Euclidean clustering, density clustering, or voxel connectivity clustering of obstacle point cloud sets; The characteristic dimensions of an obstacle include the amount of intrusion into the passage envelope, the obstacle height, the distance between the obstacle and the passage axis, and the time duration of the obstacle.

[0017] Compared with the prior art, the present invention has the following advantages: This invention uses the centerline of the underground roadway, the local coordinate system, and the passage envelope model to determine whether an obstacle has encroached on the actual passage space. The detection results directly reflect the impact on the passage of vehicles and personnel, avoiding false alarms or missed alarms caused by relying solely on point cloud distance alarms. This invention establishes a continuous motion model to map the point cloud anti-distortion at each sampling time in the integral time domain to a unified reference time, which significantly reduces point cloud stretching, ghosting and spatial misalignment caused by platform motion, sensor attitude changes or scanning timing differences. It is especially suitable for downhole mobile platform inspection scenarios. This invention adaptively determines the optimal integration time by constructing a balance between effective coverage and residual temporal variables, avoiding insufficient point cloud coverage and incomplete contours due to excessively short integration, and avoiding the accumulation of motion model errors and blurred boundaries due to excessively long integration. It achieves a dynamic balance between detection accuracy and real-time performance under complex conditions such as dust, water mist, and partial occlusion. This invention outputs information such as the three-dimensional position, size, intrusion amount of the passage envelope, distance from the passage axis, and duration of the obstacle in the local coordinate system of the roadway, and classifies the risk level (high / medium / low) accordingly, providing a direct and quantifiable basis for autonomous driving of underground vehicles, obstacle avoidance of inspection robots, and alarm of the dispatching system; This invention effectively distinguishes between fixed structures, normal floor undulations, dust and noise, and real obstacles by modeling the reference surface of the tunnel and filtering the spatial connectivity and temporal continuity of abnormal occupancy points. It maintains a high detection success rate and a low false detection rate even under conditions such as dust / water mist, uneven floor, interference from fixed structures, partial obstruction, and abnormal reflection. This invention systematically improves the accuracy, stability, and engineering practicality of obstacle detection in underground roadways from four levels: "passage envelope constraint + motion anti-distortion + adaptive integration + comprehensive risk assessment". Attached Figure Description

[0018] Figure 1 This is an overall flowchart of a method for handling obstacles in underground roadways based on solid-state three-dimensional lidar provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] A method for handling obstacles in underground roadways based on solid-state 3D lidar includes: S1. Obtain the original point cloud data with timestamps within the target area, perform motion distortion removal processing on the original point cloud data, and construct an anti-distortion point cloud set; S2. Establish a local coordinate system for the roadway based on the roadway extension direction curve, and transform the anti-distortion point cloud set to the local coordinate system for the roadway to obtain the local three-dimensional coordinates of the point cloud data points in the anti-distortion point cloud set; S3. Construct a reference envelope model of the passage space in the local coordinate system of the passage, and calculate the spatial deviation of each point cloud data point; S4. Extract the abnormal occupied points based on the spatial deviation calculated in S3, and calculate the significant defect amount of the obstacle and the intrusion amount of the reference envelope at the abnormal occupied points; S5. Set the candidate integration duration. Based on the local three-dimensional coordinates of the point cloud data points obtained in S2 and the baseline envelope model constructed in S3, calculate the effective coverage and residual time series variables of the candidate integration duration. Adaptively determine the candidate integration duration that satisfies the balance relationship between the effective coverage and residual time series variables. The smallest candidate integration duration that satisfies the balance relationship is taken as the optimal integration duration. S6. Based on the significant defect quantity of the abnormal occupied points and the baseline envelope intrusion quantity calculated in S4, cluster and segment all abnormal occupied points within the optimal integration time, output the obstacle point cloud, solve the data information of the obstacle point cloud, and generate the obstacle detection results of the underground roadway.

[0021] In S1, the solid-state 3D lidar collects the original point cloud data with timestamps within the target area, extracts all the original point cloud data in the integration time domain before the reference time, and constructs the original point cloud set in the target integration time domain. The original point cloud set includes the original measurement coordinates and sampling time of each point cloud data point. For the original point cloud set in the target integral time domain, a continuous motion model is constructed to map each point cloud data point in the original point cloud set from the sampling time to the reference time in reverse distortion, thus obtaining the reverse distortion point cloud set. The anti-distortion mapping satisfies: ; In the formula, For point cloud data point indexing, Indicates the first Sampling time of each point cloud data point Indicates reference time. Indicates the first The spatial coordinates of a point cloud data point after anti-distortion mapping Indicates the first The original measured coordinates of each point cloud data point; Indicates the sensor at the first The pose matrix relative to the global coordinate system at each sampling time. This represents the pose matrix of the sensor relative to the global coordinate system at the reference time.

[0022] In S2, the curve representing the direction of tunnel extension is: ; In the formula, A curve indicating the direction of the tunnel's extension. For arc length parameters, Represents curve In arc length The three-dimensional coordinates of the curve point at that location. For the set of real numbers, Belongs to; The local 3D coordinates of each point cloud data point in the anti-distortion point cloud set are: ; ; ; ; ; In the formula, Indicates the first Local 3D coordinates of a point cloud data point express The vertical coordinate, express The horizontal coordinate, express The height coordinates, Indicates the first Each point cloud data point is in The corresponding arc length parameter, Represents the transpose of a vector or matrix. , They represent In arc length The horizontal unit direction vector and the vertical unit direction vector at the location; Represents the Euclidean norm. Indicates the search for arc length parameter Make the vector and The Euclidean distance between them is the smallest.

[0023] The baseline envelope model for the passage space in S3 is as follows: ; In the formula, The baseline envelope model representing the passage space of the alleyway. This represents the coordinates of a spatial point in the local coordinate system of the tunnel. express The vertical coordinate, express The horizontal coordinate, express The height coordinates, express The horizontal left boundary function, express The horizontal right boundary function, express The lower boundary function of the height, express The height of the upper boundary function; Extract the reference surface of the tunnel and the reference envelope boundary of the reference envelope model, and calculate the spatial deviation of each point cloud data point. The spatial deviation includes the deviation of the point cloud data point relative to the reference surface of the tunnel and the deviation of the point cloud data point relative to the reference envelope boundary.

[0024] The deviation of the point cloud data points relative to the roadway reference surface is: ; In the formula, Indicates the first The deviation of each point cloud data point from the roadway reference surface. Indicates the reference surface of the tunnel. express arrive European distance, This is a function for calculating Euclidean distance; The deviation of the point cloud data points from the reference envelope boundary is: ; In the formula, Indicates the first The deviation of each point cloud data point from the baseline envelope boundary. Indicates the reference envelope boundary. express arrive The signed distance, This is a function for calculating signed distance; When, it indicates Located inside the baseline envelope model, When, it indicates Located on the boundary of the reference envelope, When, it indicates It is located outside the baseline envelope model.

[0025] In S4, all anomalous occupied points are extracted from the anti-distortion point cloud set, an anomalous point cloud set is constructed, and the feature quantities of the anomalous occupied points are calculated. The feature quantities include the obstacle salience defect quantity and the baseline envelope intrusion quantity. The significant defect quantity of the obstacle is: ; In the formula, Indicates the index of the abnormal occupying point. Indicates the first Significant defect quantity of obstacles at each abnormally occupied point. Indicates the first Local three-dimensional coordinates of the abnormal occupied points; express Compared to The indicator function, when lie in Inside The value is 1, when lie in The outside, The value is 0; The function for finding the maximum value. Indicates the reference surface of the tunnel at Adaptive tolerance at the location, Indicates the first Adaptive weights for each abnormally occupied point; The baseline envelope intrusion is: ; In the formula, Indicates the first The baseline envelope intrusion of each anomalous occupancy point express arrive Euclidean distance.

[0026] The effective coverage of candidate integration duration is: ; In the formula, Indicates the candidate integral duration. express Effective coverage, Indicates the longitudinal integration interval of the tunnel. It is a natural exponential function. To represent the differential, Represents the integral. This represents the differential element when integrating along the longitudinal direction of the tunnel. Indicates will vertical coordinate As an integration variable, it is used for integration calculation along the longitudinal integration interval of the roadway; Indicates that in a given Two-dimensional cross-section of the reference envelope model, Indicates in The differential element when performing double integration with respect to the horizontal and vertical directions. Let be the probability density function. This indicates that the integration time is When, the vertical coordinate is The point cloud data points are in The conditional probability density function within; This represents the logarithmic function with base e.

[0027] The discrete form is: ; In the formula, Indicates the local unit index of the roadway. This indicates that the integration time is At that time, local units of the tunnel The number of valid observation points in the data; To cover saturation scale parameters, This indicates the local units of all roadways. Summation, Indicates a local unit of the roadway The weights; Indicates a local unit of the roadway Normalized coverage Indicates the weighted total coverage. This represents the sum of the weights of all local units in the roadway; The discrete form of the residual time series variable of the candidate integration duration is: ; In the formula, express The residual time-series variables, This represents a set of local units of a roadway. Indicates a local unit of the roadway The residual variable, Indicates a local unit of the roadway The sampling time variable, For covariance, Indicates a local unit in a given roadway Under the conditions, and The square of the covariance; For variance, , These represent local units in a given roadway. Under the conditions, , The variance; It is a preset positive number.

[0028] Construct a balance relationship between effective coverage and residual time series variables, select a positive integral duration that satisfies the balance relationship between effective coverage and residual time series variables, construct a discrete candidate set, and select the minimum value in the discrete candidate set as the optimal integral duration: The balance between effective coverage and residual time series variables is as follows: ; Optimal integration time for: ; In the formula, To find the minimum value of the function, The integral duration index represents the index of the discrete candidate set. Each point duration.

[0029] In S6, based on the significant defect quantity of obstacles at the abnormal occupancy points and the baseline envelope intrusion quantity, all abnormal occupancy points within the optimal integration time are obtained, and an obstacle point cloud set is constructed. : ; In the formula, The threshold for the significant defect quantity of the obstacle. The threshold for the intrusion amount into the passage envelope; Cluster the obstacle point cloud set to obtain at least one obstacle point cloud cluster. Calculate the 3D center position, minimum bounding box size, lateral occupancy width, height occupancy range, longitudinal occupancy length, contour boundary, and distance relative to the roadway traffic axis or planned traffic path for each obstacle point cloud cluster. Calculate the feature dimensions of the corresponding obstacle based on the 3D center position, minimum bounding box size, lateral occupancy width, height occupancy range, longitudinal occupancy length, contour boundary, and distance relative to the roadway traffic axis or planned traffic path of the obstacle point cloud cluster. Clustering and segmenting of obstacle point cloud sets includes spatial connectivity analysis, Euclidean clustering, density clustering, or voxel connectivity clustering of obstacle point cloud sets; The characteristic dimensions of an obstacle include the amount of intrusion into the passage envelope, the obstacle height, the distance between the obstacle and the passage axis, and the time duration of the obstacle.

[0030] The threshold values ​​for significant obstacle defects and traffic envelope intrusion are adaptively determined by the noise scale of the roadway reference surface, local point cloud density, roadway cross-sectional dimensions, or historical background statistics. Specifically, the threshold value for significant obstacle defects is determined by multiplying the local noise standard deviation of the roadway reference surface by a preset multiple (e.g., 3 times), or by determining it based on a preset quantile of the cumulative distribution function of defects in the historical background point cloud; the threshold value for traffic envelope intrusion is determined based on vehicle safety margin, point cloud positioning accuracy, or a preset fixed distance (e.g., 0.05 meters).

[0031] In S1, a solid-state 3D lidar installed at a monitoring location in the underground roadway or on an underground mobile platform collects point cloud data within a target area at a fixed scanning cycle. The 3D spatial coordinates and sampling time of each lidar point are recorded to obtain timestamped raw point cloud data. The target area refers to the working area of ​​the underground roadway. All lidar points located within the integration time domain preceding the reference time are extracted from the raw point cloud data. An original point cloud set is constructed based on the 3D spatial coordinates and sampling time of each lidar point. When constructing the original point cloud set, one lidar point corresponds to one point cloud data point, and the 3D spatial coordinates of each lidar point are the original measurement coordinates of the corresponding point cloud data point. The original point cloud set includes the original measurement coordinates and sampling time of each point cloud data point. Each point cloud data point in the original point cloud set has its own independent sampling time, and one sampling time corresponds to only one point cloud data point. The reference time is a preset baseline time point, referring to the time used for obstacle analysis, usually the current time. The integration time domain refers to a specific, continuous time period used for data accumulation, typically one frame scanning cycle of the lidar or one control cycle of the solid-state 3D lidar acquisition system. Therefore, the original point cloud set contains all timestamped original point cloud data within a fixed time length (integral time domain) shifted forward from the reference time. Within a single frame scan cycle of the LiDAR, the sampling times of different point cloud data points are continuously distributed. Since different points have different sampling times, when the radar or platform moves within the scan cycle, the sensor poses corresponding to each point are different, and direct combination will result in distortion. Therefore, motion compensation (anti-distortion) needs to be performed using the sampling time of each point to map all points to a unified reference time.

[0032] For the constructed original point cloud set, a continuous motion model of the lidar or mobile platform relative to the roadway environment is constructed. Each point cloud data point in the original point cloud set is mapped from the anti-distortion point at the sampling time to the reference time to obtain the anti-distortion point cloud set. The continuous motion model can be any one of the following: static extrinsic parameter model, rigid body continuous motion model, uniform motion model, constant angular velocity motion model, or Lie group continuous motion model. When the solid-state 3D lidar is fixedly installed at the monitoring position in the underground roadway, the continuous motion model degenerates into a fixed observation model or a rigid body motion model that includes small vibration compensation. When the solid-state 3D lidar is fixedly mounted on an underground mobile platform, the continuous motion model is used to describe the continuous change of the pose of the mobile platform relative to the roadway environment over time.

[0033] The anti-distortion point mapping satisfies: ; In the formula, For point cloud data point indexing, Indicates the first Sampling time of each point cloud data point Indicates reference time. Indicates the first The spatial coordinates of a point cloud data point after anti-distortion mapping Indicates the first The original measured coordinates of each point cloud data point; Indicates the sensor at the first The pose matrix relative to the global coordinate system at each sampling time. This represents the pose matrix of the sensor relative to the global coordinate system at the reference time; This represents the inverse of the matrix; the pose matrix is ​​provided by the high-frequency inertial measurement unit (IMU), odometry, and simultaneous localization and mapping (SLAM) system. This is a general calculation formula for anti-distortion point mapping.

[0034] When a solid-state 3D lidar moves at a constant speed (or uniform velocity) during the scanning cycle, the anti-distortion point mapping satisfies: ; ; In the formula, Represents the natural exponential function. Indicates from the first The time difference between each sampling time and the reference time For continuous motion parameters, This represents the Lie algebra matrix corresponding to the continuous motion parameters, which include the instantaneous angular velocity and instantaneous linear velocity of the sensor during the scanning cycle.

[0035] In S3, a local coordinate system for the roadway is established based on the roadway extension direction curve. The roadway extension direction curve is a continuous curve in three-dimensional space, which describes the direction of the roadway in space, including the roadway centerline and the local traffic axis of the roadway. The curve representing the direction of the tunnel extension is: ; In the formula, A curve indicating the direction of the tunnel's extension. The arc length parameter indicates the distance from the curve. The cumulative distance measured along the curve from the previous fixed reference point (e.g., tunnel entrance, tunnel cross-section, etc.). Represents curve In arc length The three-dimensional coordinates of the curve point are equivalent to the cumulative distance from the reference point on the curve along the direction of the tunnel extension. The three-dimensional coordinates of the curve points in the local coordinate system; This indicates that the arc length can be any real number. Belongs to, For the set of real numbers, The separator indicates the curve. By satisfying conditional The corresponding curve points constitute, Represents a set.

[0036] For any point cloud data point in the anti-distortion point cloud set, determine its corresponding parameters on the curve along the roadway extension direction: ; In the formula, since one sampling time corresponds to one point cloud data point, therefore For point cloud data point indexing, For the first The corresponding parameters of each point cloud data point on the curve extending along the tunnel direction are equivalent to the distance on the centerline. The arc length parameter of the most recent point cloud data point; exist Establish a local orthogonal basis at the corresponding curve points: ; ; ; In the formula, , , These represent the curves in the direction of tunnel extension in arc length. The vertical unit direction vector, the horizontal unit direction vector, and the vertical unit direction vector at the location; The curve indicating the direction of tunnel extension is in The normal vector at that location.

[0037] Longitudinal direction is usually taken as the curve in the direction of tunnel extension, with an arc length of The tangential direction at a point, the horizontal direction refers to the direction perpendicular to the longitudinal direction and located in the horizontal plane, and is usually defined as the direction pointing to the left when facing the direction of advancement of the tunnel; the height direction is usually taken as upward.

[0038] In a locally orthogonal basis, three unit direction vectors have a magnitude of 1 and are pairwise orthogonal. The tunnel reference surface is a reference surface constructed based on historical point cloud data or tunnel design cross-sections under barrier-free access conditions, representing the ideal floor, sidewalls, and roof structure of the tunnel.

[0039] In S4, abnormal occupied points are extracted based on spatial deviation, and an abnormal point cloud set is constructed. : ; In the formula, This indicates that the point cloud data points are located within the baseline envelope model of the alleyway passage space. For adaptive tolerance. Abnormal occupancy points include point cloud data points located within the baseline envelope model and point cloud data points whose Euclidean distance to the roadway baseline surface exceeds the corresponding adaptive tolerance. Adaptive tolerance From the reference surface of the tunnel The adaptive weights are determined by the local roughness, statistical variance of historical point clouds, or preset noise tolerance, and are determined by the local density of point clouds, temporal persistence, reflection intensity, or spatial connectivity.

[0040] Adaptive weighting of anomalous occupancy points when calculating the salience defect quantity of obstacles. It is determined by the local density of the point cloud, temporal persistence, reflection intensity, neighborhood consistency, or spatial connectivity.

[0041] In S5, the effective coverage represents the number of point cloud data points in the accumulated anti-distortion point cloud set within the integration time, and the residual temporal variable represents the spatial inconsistency between point cloud data points at different sampling times after motion distortion removal processing.

[0042] The local unit of the roadway is jointly determined by the local coordinate system of the roadway and the reference envelope model. Specifically, in the local coordinate system of the roadway, the longitudinal coordinate is used as the reference envelope model. Horizontal coordinates and height coordinates For discrete dimensions, the baseline envelope model The defined passage space is divided into multiple spatial sub-regions, each of which serves as a local unit of the alleyway.

[0043] For the A partial unit of the alleyway , can be represented as: ; in, These are the division scales in the vertical, horizontal, and vertical directions, respectively. Candidate integration duration. The inverse distortion point cloud within is assigned to the corresponding local roadway unit based on its local 3D coordinates, falling into the first... The number of effective observation points in a local unit of a roadway is denoted as .

[0044] Residual variables in residual time series variables From the partial unit that fell into the alleyway The spatial residuals are formed by the point cloud data points within the tunnel, and these spatial residuals can be selected from the deviation of the point cloud data points relative to the tunnel reference surface. The deviation of point cloud data points from the reference envelope boundary Significant defect quantity of obstacles at abnormally occupied points Or baseline envelope intrusion amount One or more of the following. Sampling time variables From falling into the same lane local unit The sampling times of the point cloud data points within the data are constituted.

[0045] This can be further clarified as follows: ; ; in, or , Desirable , Any one of them or a weighted combination thereof; For abnormal occupied points or It can also be defined as the normalized combined residual.

[0046] The residual temporal variables are calculated by normalizing the correlation between the residual variables and the sampling time variables within the same local unit of the same roadway. This characterizes whether the point cloud residuals still change with the sampling time after the anti-distortion processing, thereby reflecting the degree of residual motion distortion, temporal misalignment or boundary ghosting within the integration time.

[0047] The residual time-series variable of the candidate integration duration can be represented in continuous form. Let the candidate integration duration be... Corresponding spatial integration domain for: ; In spatial location Define the residual function in the local neighborhood. and relative sampling time variables Then the residual time series variables can be expressed as: ; in, Indicates the duration of candidate integration The corresponding residual time series variable is in continuous form. The larger the value, the stronger the correlation between the point cloud residual after anti-distortion processing and the sampling time. That is, the more obvious the residual motion distortion or time series misalignment. The continuous form is convenient for theoretical analysis, while the discrete form is convenient for engineering implementation. This is the spatial location weight, used to weight the importance of different areas. It can be set to a constant or related to the access security level. The residual function represents the spatial location. Nearby, Points Duration The spatial residuals of the point cloud data points obtained within the data. To prevent the denominator from being zero, a pre-set positive number is usually chosen (e.g., a very small positive value). ), This indicates the original sampling time of the point cloud data points. Indicates reference time. This represents a spatial integral infinitesimal element. During discrete computation, it will be... Divided into multiple local units of roadways And constituted by the residual samples within each unit and the samples at the sampling time. and This allows us to obtain a set of local units based on the roadway. The discrete form of the residual time series variables.

[0048] In the discrete form, the coverage saturation scale parameter is a preset positive integer, representing the average number of observation points required to achieve 95% coverage for a single roadway local unit. Its value can be calibrated based on point cloud density and unit scale. When calculating the residual time-series variables in the discrete form, the residual variables... The preferred value is the deviation of the point cloud data points from the roadway reference surface. For identified abnormal occupied points, it can also be replaced by the amount of significant obstacle defects or the amount of reference envelope intrusion.

[0049] The discrete form of the balance relationship between effective coverage and residual time series variables is as follows: ; In the formula, The value represents the change between two adjacent candidate integration times. The optimal integration time is the smallest candidate integration time that satisfies this balance relationship. The discrete candidate set consists of a series of positive integration times that increase with a fixed step size or an adaptive step size.

[0050] In S6, when generating obstacle detection results for underground roadways, the risk level of the obstacle needs to be assessed and output based on at least one of the following factors: the intrusion amount of the obstacle point cloud cluster's passage envelope, its height, its distance from the passage axis, and its duration of time.

[0051] Continuous motion models can employ static extrinsic parameter models, micro-vibration compensation models, uniform velocity models, constant angular velocity models, short-time polynomial models, spline trajectory models, or Lie group continuous-time models. When a solid-state 3D lidar is fixedly installed on the roadway sidewall, roof, support, or monitoring column, the continuous motion model can be simplified to a fixed observation model. When there is slight vibration in the lidar installation structure, the small vibration amount can be estimated through inertial measurement units, stable regions of background point clouds, or short-time rigid body registration. When a solid-state 3D lidar is mounted on an underground mobile platform, the pose of the mobile platform can be... The results are obtained from wheeled odometers, inertial measurement units, tunnel positioning systems, visual inertial odometers, laser point cloud matching results, or combinations thereof; the tunnel centerline or local traffic axis can be determined from tunnel design drawings, pre-built mapping results, track centerlines, road centerlines, tunnel boundary fitting results, or mobile platform planning paths; the tunnel reference surface can be constructed from historical background point clouds, calibration point clouds under barrier-free access conditions, tunnel design cross-sections, local plane fitting, surface fitting, voxel maps, or occupied grid maps; the tunnel's traversable envelope can be determined based on the mining vehicle's external dimensions. The dimensions of the inspection robot, the safe distance for personnel passage, the roadway clearance, the track clearance, pipeline layout requirements, or operating procedures are determined; the effective coverage can be constructed based on the longitudinal slice coverage rate of the roadway, the lateral-height section coverage entropy, the three-dimensional voxel occupancy saturation, the stability of the number of obstacle candidate area points, or multi-scale spatial coverage indicators; residual temporal variables can be constructed based on the correlation between spatial deviation residuals, the residuals of significant obstacle defects, the residuals of passage envelope intrusion and sampling time, conditional mutual information, temporal residual trends, or the degree of local boundary ghosting; obstacle points Cloud clustering can employ Euclidean clustering, DBSCAN density clustering, voxel connectivity clustering, region growing, graph cut segmentation, or learning-based point cloud instance segmentation methods; risk levels can be comprehensively assessed based on factors such as obstacle intrusion depth, obstacle height, lateral occupancy width, longitudinal duration, distance from the traffic axis, mobile platform speed, vehicle braking distance, and obstacle duration; detection results can be output to underground vehicle autonomous driving systems, inspection robot obstacle avoidance systems, roadway safety monitoring platforms, dispatch center alarm systems, or mine digital twin platforms.

[0052] To verify the detection effectiveness of a solid-state 3D lidar-based obstacle handling method for underground roadways in a fixed monitoring scenario, a field experiment was conducted. The experiment included various common underground obstacles such as falling rocks, coal gangue piles, tools and equipment, support materials, and simulated sagging pipelines. These obstacles were placed in the central area of ​​the roadway, the edge of the passageway, near the track, and near the sidewalls.

[0053] A fixed solid-state 3D lidar was deployed in a test underground roadway. This test roadway was approximately straight, with a length of about 60m, an effective width of about 4.2m, and an effective height of about 3.1m. The roadway floor exhibited local undulations, and the sidewalls and roof contained support structures, anchor mesh, pipelines, and locally rough surfaces. The solid-state 3D lidar was fixedly mounted on a monitoring column on the roadway sidewall at a height of approximately 1.45m, facing the passageway area, monitoring the passable space within a range of approximately 5m to 25m ahead. To minimize installation attitude errors, the lidar coordinate system was calibrated against the local roadway coordinate system before the experiment, and a reference surface for the roadway was established using background point clouds collected under unobstructed passage conditions.

[0054] During the experiment, the system continuously acquired timestamped 3D point clouds within the tunnel. For each reference time, the point cloud within the integral time domain preceding the reference time was selected to establish a fixed observation model. Considering that vibrations of underground equipment may cause minor attitude disturbances to the radar, a short-time rigid body micro-vibration compensation term was also introduced in this experiment to uniformly map each sampling point within the integral time domain to the reference time.

[0055] Subsequently, the system establishes a local coordinate system based on the roadway centerline, converting the point cloud into longitudinal, lateral, and height coordinates. Based on pre-calibrated roadway floor, sidewalls, roof, track surface, and fixed support structures, a roadway reference surface is constructed. Simultaneously, a passable envelope is constructed according to the passage requirements of mining vehicles and personnel. In this experiment, the passable envelope retains a safe passage width from both sidewalls in the lateral direction, excludes space occupied by the roof, pipelines, and support structures in the height direction, and incorporates the external dimensions and safety margins of vehicles or inspection platforms into the constraints.

[0056] For points located within the passable envelope but significantly deviating from the roadway reference surface, the system extracts them as anomalous occupancy points and further calculates the significant defect quantity of the obstacle and the intrusion quantity into the passable envelope. When anomalous occupancy points form a connected region in space and continuously appear at multiple reference times, the system classifies them as candidate obstacle point clouds. Then, the system clusters the anomalous occupancy points to obtain obstacle point cloud clusters and outputs the obstacle center coordinates, 3D dimensions, lateral occupancy width, height occupancy range, contour boundary, passable envelope intrusion quantity, and risk level.

[0057] The experiment used detection success rate, false detection rate, false negative rate, obstacle center position error, size error, passage envelope intrusion error, and risk level consistency rate as evaluation indicators. Among them, obstacle center position error is the three-dimensional distance between the detected output center point and the measured center point; size error is the average absolute error between the detected output bounding box size and the measured size; risk level consistency rate is the proportion of the system output risk level that matches the manually labeled risk level. The experimental results under fixed monitoring scenarios are shown in Table 1.

[0058] Table 1. Experimental results under fixed monitoring scenarios;

[0059] ;

[0060] The success rate, false detection rate, missed detection rate and error results in Table 1 are the average statistical results of no less than 3 repeated experiments under the same obstacle arrangement conditions. The true values ​​of the obstacle center position and size are obtained by manual measurement and verification.

[0061] As shown in Table 1, under fixed installation conditions, this invention can effectively identify abnormally occupied points within the passable envelope of the roadway and cluster them into obstacle point cloud clusters. For obstacles with relatively large volumes, such as coal gangue piles and fallen rocks, the detection success rate is high. For slender structures such as drooping pipeline simulations, the detection difficulty is relatively high due to the smaller number of point clouds and susceptibility to local occlusion, but good detection stability is still maintained. Experimental results show that this invention can output complete three-dimensional obstacle detection results in fixed roadway monitoring scenarios and can effectively determine whether obstacles intrude into the main passage area.

[0062] To verify the detection performance of a solid-state 3D LiDAR-based obstacle handling method in underground roadways under mobile conditions, a solid-state 3D LiDAR was fixedly mounted on the top of an underground inspection robot at a height of approximately 1.05m. The radar's main field of view faced the passageway in front of the platform, and the point cloud output frequency was preferably set to 10Hz. The mobile platform traveled or inspected along the longitudinal direction of the roadway at speeds of 0.4m / s, 0.8m / s, and 1.2m / s to simulate low-speed inspection, medium-speed passage, and relatively fast-moving detection conditions. Data was collected at least three times under each speed condition, and the average value was taken. The mobile platform was equipped with a wheeled odometer, an inertial measurement unit, and a roadway positioning module. The platform's pose was also corrected using point cloud matching results from adjacent time points.

[0063] Because the mobile platform is constantly moving during point cloud acquisition, directly accumulating the point cloud data in the integration time domain will result in points at different sampling times corresponding to different platform poses. This will cause stretching, ghosting, or spatial offset of the tunnel sidewalls, floor, tracks, and obstacle edges. Especially when the mobile platform is moving at high speeds or there are local obstructions in the tunnel, directly accumulating the point cloud will lead to unclear obstacle outlines, thus affecting the output of obstacle center points, dimensions, and risk levels.

[0064] Therefore, this embodiment establishes a continuous motion model of the mobile platform relative to the tunnel environment based on the mobile platform's odometer, inertial measurement unit, tunnel positioning system, and point cloud matching results. For each sampling point in the integral time domain, the system anti-distorts and maps the point to a unified reference time according to the platform pose corresponding to its sampling time. After completing the anti-distortion, the system then performs tunnel local coordinate transformation, tunnel reference surface construction, traversable envelope construction, abnormal occupancy point extraction, obstacle clustering, and risk level output on the anti-distorted point cloud.

[0065] In the experiment, three processing methods were set up to compare the effects of continuous anti-distortion:

[0066] Direct accumulation method: without anti-distortion, the point cloud in the integration time domain is directly accumulated; conventional inter-frame registration method: adjacent point cloud frames are registered frame by frame and then accumulated; a method for handling obstacles in underground roadways based on solid-state three-dimensional lidar (the method of this invention): a continuous motion model is established, and the points at each sampling time in the integration time domain are uniformly anti-distorted to the reference time.

[0067] Evaluation metrics include residual temporal variables, detection success rate, center position error, size error, and risk level consistency rate. The larger the residual temporal variables, the more obvious the temporal misalignment still exists in the point cloud after distortion correction. The experimental results in the moving scene are shown in Table 2.

[0068] Table 2. Experimental results under movement;

[0069] ;

[0070] As shown in Table 2, when the mobile platform speed is low, the direct accumulation method can still achieve a certain detection effect, but the obstacle edges already exhibit some degree of stretching and ghosting. As the platform speed increases, the residual temporal variables of the direct accumulation method increase significantly, the obstacle detection success rate decreases, and both the center position error and size error increase significantly. Conventional inter-frame registration methods can alleviate temporal misalignment to some extent, but under conditions of repetitive textures, dust interference, local occlusion, and incomplete obstacle point clouds in underground roadways, registration errors and cumulative drift may still occur.

[0071] In contrast, the method of this invention, by establishing a continuous motion model of the mobile platform relative to the tunnel environment and mapping the point distortion at each sampling time in the integral time domain to a unified reference time, can significantly reduce residual temporal variables and improve the consistency of obstacle edges and contours. Even under relatively fast inspection conditions of 1.2 m / s, the method of this invention can still maintain a high detection success rate and risk level consistency rate, indicating that it is suitable for obstacle detection in underground mobile platform inspection scenarios.

[0072] To illustrate the combined impact of integration duration on point cloud coverage integrity and temporal mismatch, reference data is obtained using an underground mobile platform inspection scenario as an example. Specifically, a solid-state 3D LiDAR is fixedly mounted above and in front of the underground inspection robot, at a height of approximately 1.05m, with the radar's main field of view facing the passageway in front of the platform. The inspection robot travels longitudinally along the tunnel, with an average speed of approximately 0.8m / s. Obstacles such as falling rocks, coal gangue piles, tools, and support materials are placed within the tunnel, in the central area, the edge of the passageway, and near the track.

[0073] During the experiment, the system selected 0.2s, 0.4s, 0.6s, 0.8s, and 1.0s as candidate integration durations. For each integration duration, the same point cloud data source and the same ground truth obstacle values ​​were used for repeated verification. The detection success rate, center position error, size error, and boundary matching rate were statistically analyzed to evaluate the impact of integration duration on detection stability and boundary accuracy. For each candidate integration duration, the system established a continuous motion model based on the inspection robot's odometry, inertial measurement unit, and point cloud matching results, and mapped the anti-distortion point cloud in the corresponding integration time domain to a unified reference time. Subsequently, the effective coverage C(T) and residual temporal variable M(T) were calculated on the anti-distorted point cloud, and obstacle segmentation, clustering, and detection result statistics were performed.

[0074] The effective coverage C(T) is calculated based on the number of effective observation points in the longitudinal slice and transverse-height section of the roadway, the degree of coverage saturation, and the coverage integrity of the candidate obstacle area. The residual temporal variable M(T) is calculated based on the correlation between the residuals of the spatial deviation after distortion, the amount of significant obstacle defects, or the intrusion of the passage envelope and the sampling time. The detection results use the manually measured center position, size, and contour boundary of the obstacle as a reference. The experimental results corresponding to different candidate integration times in the mobile platform inspection scenario are shown in Table 3.

[0075] Table 3. Experimental results for different candidate integration times;

[0076] ;

[0077] As shown in Table 3, under the aforementioned mobile platform inspection conditions, when the integration time is 0.2s, although the residual temporal variables are low, the insufficient number of accumulated point clouds results in incomplete coverage of the passable space in the alley and the candidate structure of obstacles, leading to a low detection success rate and contour boundary matching rate. With the integration time increasing to 0.4s and 0.6s, the effective coverage significantly improves, the obstacle contours become more complete, and both the center position error and size error decrease. When the integration time continues to increase to 0.8s and 1.0s, the increase in effective coverage becomes smaller, but the residual temporal variables increase significantly. This indicates that within a longer integration time domain, platform motion model errors, local occlusion changes, and point cloud temporal differences begin to accumulate, causing slight stretching or edge ghosting of the obstacle contours.

[0078] Calculate the following on the candidate integral duration set respectively and The discrete values ​​are obtained, and their derivatives are approximated by the difference between adjacent sampling points:

[0079] ;

[0080] ;

[0081] In the formula, Indicates the candidate integral duration index. , They represent the first in the candidate integral duration set, respectively. , The equilibrium point is obtained by performing linear interpolation within the interval where the two discrete marginal curves change from "greater than" to "less than". The calculations obtained in this experiment are as follows: This integration time ensures the complete coverage of candidate obstacle structures and passable lane space while avoiding a significant increase in residual temporal mismatch caused by excessively long integration times. Therefore, this invention achieves a better balance between point cloud coverage integrity and temporal mismatch control, providing a stable point cloud foundation for subsequent obstacle segmentation, clustering, and risk level output.

[0082] After determining the optimal integration time, the system only uses the anti-distortion point cloud corresponding to the optimal integration time to perform obstacle detection. First, the system adaptively determines the threshold for significant obstacle defects and the threshold for passage envelope intrusion based on the noise scale of the roadway reference surface, local point cloud density, roadway cross-sectional dimensions, historical background statistics, and current observation confidence.

[0083] To improve segmentation stability in complex downhole environments, the system can perform continuous temporal filtering of obstacle candidate points. If an anomaly-occupied region appears only briefly at a single reference time and lacks spatial connectivity or neighborhood consistency, it can be eliminated as dust, water mist, reflection anomalies, or temporary noise. If an anomaly-occupied region exists at multiple consecutive reference times and its position and contour are stable, it is retained as an obstacle candidate region.

[0084] Subsequently, the system clusters the set of obstacle candidate points. Clustering methods can include Euclidean clustering, density clustering, spatial connectivity analysis, or voxel connectivity clustering. In one implementation, obstacle candidate points are first divided into a three-dimensional voxel grid, and then connected components are formed based on the spatial connectivity relationships between adjacent voxels. For each connected component, if its number of points, spatial size, and duration meet preset conditions, it is output as an obstacle point cloud cluster.

[0085] For each obstacle point cloud cluster, the system calculates the following detection results: three-dimensional coordinates of the obstacle center point; longitudinal occupancy length of the obstacle; lateral occupancy width of the obstacle; height occupancy range of the obstacle; three-dimensional bounding box; contour boundary; intrusion amount of the passage envelope; distance from the roadway centerline or local passage axis; risk level; alarm information. The obstacle output results are shown in Table 4.

[0086] The three-dimensional coordinates of the obstacle's center point can be represented by the geometric center, weighted center, or bounding box center of the obstacle's point cloud cluster. The longitudinal occupancy length is the maximum occupancy range of the obstacle in the longitudinal direction of the roadway; the lateral occupancy width is the maximum occupancy range of the obstacle in the lateral direction of the roadway; the height occupancy range is the interval between the minimum and maximum height of the obstacle in the height direction; the three-dimensional bounding box can be an axis-aligned bounding box, a bounding box in the local coordinate system of the roadway, or a minimum bounding box.

[0087] Risk levels can be determined comprehensively based on factors such as whether an obstacle encroaches on the main passageway, the extent of encroachment into the passageway envelope, the obstacle's height, lateral occupancy width, distance from the roadway centerline or local passageway axis, and the duration of the obstacle's presence. For example, obstacles located near the roadway centerline that have a significant depth of encroachment into the main passageway envelope and a large height or lateral width can be classified as high-risk; obstacles located at the edge of the passageway envelope that are small in size and have a weak impact on the main passageway can be classified as low-risk or indicate a risk level.

[0088] Risk levels can be divided into three levels:

[0089] Level 1 risk, or low risk or warning risk: The obstacle is located at the edge of the passage envelope, the intrusion is small, and the impact on the main passage path is limited;

[0090] Level 2 risk, or medium risk: The obstacle partially encroaches on the passageway, which may affect the passage of vehicles, mobile platforms or personnel;

[0091] Level 3 risk, or high risk: If an obstacle significantly encroaches on the main traffic path, or if its height, width, or duration reaches dangerous conditions, an alarm should be sounded immediately or a vehicle should be stopped to avoid the obstacle.

[0092] Table 4. Obstacle output results;

[0093] ;

[0094] As shown in Table 4, this invention can not only determine whether there are obstacles in the roadway, but also output the location, size, occupancy range, and risk level of the obstacles in the local coordinate system of the roadway. For the same roadway cross-section, obstacles at different locations and heights have different impacts on traffic safety. For example, although the coal gangue pile located in the central area of ​​the roadway may not be the tallest, it is judged as high-risk because it directly intrudes into the main traffic path; although the lateral width of the simulated drooping pipeline is small, its intrusion into the height envelope of vehicles or mobile platforms also makes it judged as high-risk. It can be seen that the risk level output is not based solely on the size of the obstacle, but is comprehensively determined by combining the intrusion amount of the traffic envelope, the location of the obstacle, and the actual traffic demand.

[0095] To verify the adaptability of this method in complex downhole conditions, experiments were conducted in various environments, including conventional roadways, dust / water mist, uneven floor surfaces, interference from fixed structures, localized water accumulation / abnormal reflections, motion distortion of the mobile platform, localized obstructions, and various types of obstacles. The effective width of the experimental roadway was approximately 4.2 m, the effective height was approximately 3.1 m, and the test distance ranged from 3 to 15 m. A 3D LiDAR was mounted on top of the inspection mobile robot at a height of 1.05 m, with the scanning direction facing the direction of vehicle or platform movement. The mobile platform speed was set to 0.4 m / s, 0.8 m / s, and 1.2 m / s. Obstacles included falling rocks, gangue, toolboxes, support materials, pipelines, vehicle parts, and simulated personnel, with obstacle heights ranging from approximately 0.1 to 1.8 m, lateral widths from approximately 0.1 to 1.5 m, and longitudinal lengths from approximately 0.2 to 2.0 m.

[0096] Under normal tunnel conditions, with normal lighting, no significant dust or water mist in the air, a relatively flat floor, and no or only slight edge obstruction from obstacles, baseline detection results are obtained. Under dust / water mist conditions, a localized low-visibility area is created within 3 to 12 meters in front of the radar using dust or spray methods, causing localized sparseness, noise points, short-term echo loss, and blurred boundaries in the point cloud, reducing the visible point cloud of obstacles by approximately 10% to 30%. Under uneven floor conditions, the test area is designed with localized undulations of approximately 0.05 to 0.20 meters in the presence of gravel, slope variations, or floor height differences, causing slight attitude disturbances to the moving platform as it passes through.

[0097] Under the fixed structure interference condition, the experimental area contains fixed structures such as anchor bolts, pipelines, supports, tunnel wall protrusions, or equipment bases. Some of these fixed structures are close to or connected to the boundary of the obstacle to be detected in the point cloud, used to verify the method's ability to distinguish non-target structures. Under the local water accumulation / reflection anomaly condition, water accumulation or high reflectivity areas with an area of ​​approximately 0.5 to 2.0 m² and a depth of approximately 0.01 to 0.05 m are set on the tunnel floor, causing local voids, false reflection points, or abnormal echo intensity in the point cloud. Under the moving platform motion distortion condition, the platform speed is set to 0.4 to 1.2 m / s. Due to the time difference in radar scanning, the original point cloud may show boundary trailing, spatial misalignment, and dimensional deviation. This method reduces this impact through motion compensation and temporal alignment.

[0098] In partial obstruction mode, the obstruction is located between the radar and the target obstacle, reducing the visible area of ​​the target obstacle by approximately 20% to 50%. Obstruction locations include the lower part, sides, and local corners of the obstacle. In multi-shaped obstacle mode, targets of different shapes, such as regular boxes, irregular rocks, slender pipelines, sheet-like support materials, and simulated personnel, are used to verify the adaptability of this method to obstacles of different sizes, shapes, and occupancy methods.

[0099] Under the aforementioned complex operating conditions, the system still performs processing according to the method of this invention. First, continuous anti-distortion is performed on the time-stamped point cloud; then, a reference surface and a passable envelope are constructed in the local coordinate system of the roadway; next, abnormal occupied points are extracted based on the significant defect quantity of obstacles and the intrusion quantity of the passable envelope; finally, obstacle detection results are output through clustering and risk assessment.

[0100] This experiment uses detection success rate, false detection rate, false negative rate, center position error, size error, and risk level consistency rate as evaluation indicators. Specifically, the false detection rate evaluates the proportion of non-obstacle areas incorrectly output as obstacles by the system; the false negative rate evaluates the proportion of real obstacles not output by the system; and the risk level consistency rate evaluates the consistency between the risk level output by the system and the human safety assessment level.

[0101] For each type of complex working condition, no less than three rounds of independent data collection were conducted. Each round of data collection included multiple obstacle placement locations and different obstacle types. For local dust, water mist, abnormal reflection, and occlusion working conditions, the actual visible area and occlusion area of ​​the obstacle were recorded simultaneously during manual verification to avoid misjudging the failure to detect due to complete invisibility as a missed detection by the algorithm. The experimental results for complex working conditions are shown in Figure 5.

[0102] Table 5. Experimental results under complex working conditions;

[0103] ;

[0104] As can be seen from Table 5, under normal roadway conditions, the present invention has a high detection success rate and risk level consistency rate; under difficult conditions such as dust, water mist, abnormal reflection and partial obstruction, the detection success rate decreases and the false detection rate and missed detection rate increase, but overall it still maintains good obstacle detection stability.

[0105] In dusty and water mist conditions, discrete noise points or local echo gaps may appear in point clouds. This invention constrains anomalous occupancy points through local density, temporal persistence, and spatial connectivity, reducing the misclassification of instantaneous noise points as obstacles. In cases of uneven floor surfaces, directly using a fixed height threshold can easily misclassify floor undulations as obstacles; this invention, by constructing a roadway reference surface and introducing local roughness tolerance, can distinguish normal floor undulations from genuine anomalous occupancy. In cases of fixed structure interference, although fixed facilities such as tracks, pipelines, and supports form obvious structures in the point cloud, the system can reduce the probability of false detection because they can be incorporated into the roadway reference surface or fixed background structures.

[0106] In situations with localized water accumulation or abnormal reflections, the detection error increases due to potential voids or anomalous reflections in the point cloud echoes. This invention addresses this by using historical background statistics, neighborhood consistency, and multi-reference time persistence to suppress false obstacles caused by some reflection anomalies. In cases of partial occlusion, the obstacle point cloud is incomplete, leading to increased size estimation errors. However, through an optimal integration duration adaptive mechanism and spatial connectivity clustering, the system can still output information about the main obstacle regions.

[0107] The experimental results above demonstrate that this invention can adapt to various complex working conditions in underground roadways. Its stability primarily stems from the following aspects: First, continuous anti-distortion reduces spatial misalignment caused by mobile platform movement, radar vibration, and scanning timing differences; second, it combines obstacle detection with the actual passage direction and cross-sectional structure of the roadway using a local coordinate system; third, it reduces false detections caused by rough undulations of fixed structures and reference surfaces through constraints from the reference surface and traversable envelope; fourth, it adaptively determines the optimal integration time by balancing the effective coverage and residual temporal variables, taking into account both point cloud structural integrity and temporal mismatch control; and fifth, it achieves a comprehensive output of obstacle location, size, boundary, and risk level through obstacle saliency defect quantity, passage envelope intrusion quantity, and spatial connectivity clustering.

[0108] Figure 1 This is an overall flowchart of a method for handling obstacles in underground roadways based on solid-state 3D lidar, as shown below. Figure 1 As shown, a method for handling obstacles in underground roadways based on solid-state 3D lidar includes: acquiring original point cloud data with timestamps within the target area; performing motion distortion removal processing (anti-distortion processing) on ​​the original point cloud data to construct an anti-distortion point cloud set; establishing a local coordinate system for the roadway based on the roadway extension direction curve, and transforming the anti-distortion point cloud set into the local coordinate system to obtain the local 3D coordinates of the point cloud data points in the anti-distortion point cloud set; constructing a reference envelope model of the roadway passage space in the local coordinate system, and calculating the spatial deviation of each point cloud data point; extracting abnormal occupied points based on the spatial deviation, and calculating the abnormal... The system calculates the significant defect quantity of obstacles at the occupied points and the intrusion quantity of the baseline envelope; it sets candidate integration durations, calculates the effective coverage and residual temporal variables of the candidate integration durations based on the local 3D coordinates of the point cloud data points and the baseline envelope model, adaptively determines the integration length of the candidate integration durations through the balance relationship between the effective coverage and residual temporal variables, and adaptively determines the optimal integration duration; based on the significant defect quantity of obstacles at the abnormal occupied points and the intrusion quantity of the baseline envelope, it performs clustering and segmentation on all abnormal occupied points within the optimal integration duration, outputs the obstacle point cloud, solves the data information of the obstacle point cloud, and generates the obstacle detection results for the underground roadway.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for handling obstacles in underground roadways based on solid-state three-dimensional lidar, characterized in that, include: S1. Obtain the original point cloud data with timestamps within the target area, perform motion distortion removal processing on the original point cloud data, and construct an anti-distortion point cloud set; S2. Establish a local coordinate system for the roadway based on the roadway extension direction curve, and transform the anti-distortion point cloud set to the local coordinate system for the roadway to obtain the local three-dimensional coordinates of the point cloud data points in the anti-distortion point cloud set; S3. Construct a reference envelope model of the passage space in the local coordinate system of the passage, and calculate the spatial deviation of each point cloud data point; S4. Extract the abnormal occupied points based on the spatial deviation calculated in S3, and calculate the significant defect amount of the obstacle and the intrusion amount of the reference envelope at the abnormal occupied points; S5. Set the candidate integration duration. Based on the local three-dimensional coordinates of the point cloud data points obtained in S2 and the baseline envelope model constructed in S3, calculate the effective coverage and residual time series variables of the candidate integration duration. Adaptively determine the candidate integration duration that satisfies the balance relationship between the effective coverage and residual time series variables. The smallest candidate integration duration that satisfies the balance relationship is taken as the optimal integration duration. S6. Based on the significant defect quantity of the obstacle and the baseline envelope intrusion quantity of the abnormally occupied points calculated in S4, cluster and segment all abnormally occupied points within the optimal integration time, output the obstacle point cloud, solve the data information of the obstacle point cloud, and generate the obstacle detection results of the underground roadway. The effective coverage of candidate integration duration is: ； In the formula, Indicates the candidate integral duration. express Effective coverage, Indicates the longitudinal integration interval of the tunnel. It is a natural exponential function. To represent the differential, Represents the integral. This represents the differential element when integrating along the longitudinal direction of the tunnel. Indicates will vertical coordinate As an integration variable, it is used for integration calculation along the longitudinal integration interval of the roadway; Indicates that in a given Two-dimensional cross section of the reference envelope model, Indicates in The differential element when performing double integration with respect to the horizontal and vertical directions. Let be the probability density function. This indicates that the integration time is When, the vertical coordinate is The point cloud data points are in The conditional probability density function within; This represents the logarithmic function with base e. Construct a balance relationship between effective coverage and residual time series variables, select a positive integral duration that satisfies the balance relationship between effective coverage and residual time series variables, construct a discrete candidate set, and select the minimum value in the discrete candidate set as the optimal integral duration; The balance between effective coverage and residual time series variables is as follows: ； Optimal integration time for: ； In the formula, express The residual time-series variables, To find the minimum value of the function, The integral duration index represents the index of the discrete candidate set. Each point duration.

2. The method for handling obstacles in underground roadways based on solid-state three-dimensional lidar according to claim 1, characterized in that, In S1, the solid-state 3D lidar collects the original point cloud data with timestamps within the target area, extracts all the original point cloud data in the integration time domain before the reference time, and constructs the original point cloud set in the target integration time domain. The original point cloud set includes the original measurement coordinates and sampling time of each point cloud data point. For the original point cloud set in the target integral time domain, a continuous motion model is constructed to map each point cloud data point in the original point cloud set from the sampling time to the reference time in reverse distortion, thus obtaining the reverse distortion point cloud set. The anti-distortion mapping satisfies: ； In the formula, For point cloud data point indexing, Indicates the first Sampling time of each point cloud data point Indicates reference time. Indicates the first The spatial coordinates of a point cloud data point after anti-distortion mapping Indicates the first The original measured coordinates of each point cloud data point; Indicates the sensor at the first The pose matrix relative to the global coordinate system at each sampling time. This represents the pose matrix of the sensor relative to the global coordinate system at the reference time.

3. The method for handling obstacles in underground roadways based on solid-state three-dimensional lidar according to claim 2, characterized in that, In S2, the curve representing the direction of tunnel extension is: ； In the formula, A curve indicating the direction of the tunnel's extension. For arc length parameters, Represents curve In arc length The three-dimensional coordinates of the curve point at that location. For the set of real numbers, Belongs to; The local 3D coordinates of each point cloud data point in the anti-distortion point cloud set are: ； ； ； ； ； In the formula, Indicates the first Local 3D coordinates of a point cloud data point express The vertical coordinate, express The horizontal coordinate, express The height coordinates, Indicates the first Each point cloud data point in The corresponding arc length parameter, Represents the transpose of a vector or matrix. , They represent In arc length The horizontal unit direction vector and the vertical unit direction vector at the location; Represents the Euclidean norm. Indicates the search for arc length parameter Make the vector and The Euclidean distance between them is the smallest.

4. The method for handling obstacles in underground roadways based on solid-state three-dimensional lidar according to claim 3, characterized in that, The baseline envelope model for the passage space in S3 is as follows: ； In the formula, The baseline envelope model representing the passage space of the alleyway. This represents the coordinates of a spatial point in the local coordinate system of the tunnel. express The vertical coordinate, express The horizontal coordinate, express The height coordinates, express The horizontal left boundary function, express The horizontal right boundary function, express The lower boundary function of the height, express The height of the upper boundary function; Extract the reference surface of the tunnel and the reference envelope boundary of the reference envelope model, and calculate the spatial deviation of each point cloud data point. The spatial deviation includes the deviation of the point cloud data point relative to the reference surface of the tunnel and the deviation of the point cloud data point relative to the reference envelope boundary.

5. The method for handling obstacles in underground roadways based on solid-state three-dimensional lidar according to claim 4, characterized in that, The deviation of the point cloud data points relative to the roadway reference surface is: ； In the formula, Indicates the first The deviation of each point cloud data point from the roadway reference surface. Indicates the reference surface of the tunnel. express arrive European distance, This is a function for calculating Euclidean distance; The deviation of the point cloud data points from the reference envelope boundary is: ； In the formula, Indicates the first The deviation of each point cloud data point from the baseline envelope boundary. Indicates the reference envelope boundary. express arrive The signed distance, This is a function for calculating signed distance; When, it indicates Located inside the baseline envelope model, When, it indicates Located on the boundary of the reference envelope, When, it indicates It is located outside the baseline envelope model.

6. The method for handling obstacles in underground roadways based on solid-state three-dimensional lidar according to claim 5, characterized in that, In S4, all anomalous occupied points are extracted from the anti-distortion point cloud set, an anomalous point cloud set is constructed, and the feature quantities of the anomalous occupied points are calculated. The feature quantities include the obstacle salience defect quantity and the baseline envelope intrusion quantity. The significant defect quantity of the obstacle is: ； In the formula, Indicates the index of the abnormal occupying point. Indicates the first Significant defect quantity of obstacles at each abnormally occupied point. Indicates the first Local three-dimensional coordinates of the abnormal occupied points; express Compared to The indicator function, when lie in Inside The value is 1, when lie in The outside, The value is 0; The function for finding the maximum value. Indicates the reference surface of the tunnel at Adaptive tolerance at the location, Indicates the first Adaptive weights for each abnormally occupied point; The baseline envelope intrusion is: ； In the formula, Indicates the first The baseline envelope intrusion of each anomalous occupancy point express arrive Euclidean distance.

7. The method for handling obstacles in underground roadways based on solid-state three-dimensional lidar according to claim 6, characterized in that, The discrete form is: ； In the formula, Indicates the local unit index of the roadway. This indicates that the integration time is At that time, local units of the tunnel The number of valid observation points in the data; To cover saturation scale parameters, This indicates the local units of all roadways. Summation, Indicates a local unit of the roadway The weights; Indicates a local unit of the roadway Normalized coverage Indicates the weighted total coverage. This represents the sum of the weights of all local units in the roadway; The discrete form of the residual time series variable of the candidate integration duration is: ； In the formula, express The residual time-series variables, This represents a set of local units of a roadway. Indicates a local unit of the roadway The residual variable, Indicates a local unit of the roadway The sampling time variable, For covariance, Indicates a local unit in a given roadway Under the conditions, and The square of the covariance; For variance, , These represent local units in a given roadway. Under the conditions, , The variance; It is a preset positive number.

8. The method for handling obstacles in underground roadways based on solid-state three-dimensional lidar according to claim 7, characterized in that, In S6, based on the significant defect quantity of obstacles at the abnormal occupancy points and the baseline envelope intrusion quantity, all abnormal occupancy points within the optimal integration time are obtained, and an obstacle point cloud set is constructed. : ； In the formula, The threshold for the significant defect quantity of the obstacle. The threshold for the intrusion amount into the passage envelope; Cluster the obstacle point cloud set to obtain at least one obstacle point cloud cluster. Calculate the 3D center position, minimum bounding box size, lateral occupancy width, height occupancy range, longitudinal occupancy length, contour boundary, and distance relative to the roadway traffic axis or planned traffic path for each obstacle point cloud cluster. Calculate the feature dimensions of the corresponding obstacle based on the 3D center position, minimum bounding box size, lateral occupancy width, height occupancy range, longitudinal occupancy length, contour boundary, and distance relative to the roadway traffic axis or planned traffic path of the obstacle point cloud cluster. Clustering and segmenting of obstacle point cloud sets includes spatial connectivity analysis, Euclidean clustering, density clustering, or voxel connectivity clustering of obstacle point cloud sets; The characteristic dimensions of an obstacle include the amount of intrusion into the passage envelope, the obstacle height, the distance between the obstacle and the passage axis, and the time duration of the obstacle.

Images