Subway tunnel space safety detection method and device based on 3D point cloud

Abstract

The application discloses a subway tunnel space safety detection method and device based on a 3D point cloud. The subway tunnel space safety detection method is that a laser device is installed on a mobile chassis to move in a subway tunnel space along a preset moving direction, the laser device collects three-dimensional point cloud data of an inner wall of the subway tunnel through a spiral scanning mode, the detected point cloud after being uniformly pretreated forms a target detection point cloud, and deformation, intrusion and foreign matter detection of the subway tunnel space are performed based on the target detection point cloud. The laser device comprises at least three laser radars which are uniformly arranged in the same transverse section of the mobile chassis in a circumferential direction or at least two laser radars which are arranged in a collinear longitudinal direction at equal intervals along an axis of the mobile chassis. The application realizes blind area-free scanning of the subway tunnel space.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for transportation facilities, and in particular to a method and device for spatial safety detection in subway tunnels based on 3D point clouds. Background Technology

[0002] As a crucial component of urban transportation, the safety of subway tunnels directly impacts the stability of subway operations and passenger safety. Currently, subway tunnel safety inspections primarily rely on manual patrols and traditional contact-based inspection methods. These methods suffer from low efficiency, strong subjectivity, and limited coverage, making it difficult to detect potential safety hazards such as tunnel deformation, clearance encroachment, and foreign objects in real time and comprehensively.

[0003] In recent years, 3D laser scanning technology has provided a new solution for tunnel inspection. Laser point clouds do not rely on external lighting, have strong anti-interference performance, and can directly and accurately obtain distance information of objects. They can be directly used to determine the safe distance in tunnel space. At the same time, there is a certain relationship between laser reflection intensity and the surface material of the object. By observing changes in radiation intensity information, structural defects such as leakage can be further identified.

[0004] However, during mobile scanning, blind spots may occur when a single radar scans, affecting the scanning quality. Furthermore, ensuring the quality of 3D laser point cloud data (such as density uniformity and coordinate consistency) and efficiently and accurately identifying various specific types of defects remains a technical challenge in practical applications. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and defects of the prior art and provide a method and device for spatial safety detection of subway tunnels based on 3D point clouds, which is used for fully automatic and high-precision detection of tunnel deformation, encroachment or foreign objects, significantly improving work efficiency and data quality.

[0006] One aspect of the present invention provides a method for space safety detection in subway tunnels based on 3D point clouds. A laser device is mounted on a mobile chassis and moves along a preset direction within the subway tunnel space. The laser device collects three-dimensional point cloud data of the inner wall of the subway tunnel using a helical scanning method. The scanned detection point cloud is pre-processed to homogenize and form a target detection point cloud. Deformation, encroachment, and foreign object detection in the subway tunnel space are performed based on the target detection point cloud. The laser device includes at least three lidars uniformly arranged circumferentially within the same transverse cross-section of the mobile chassis, or at least two lidars arranged equidistantly along the longitudinal axis of the mobile chassis. The laser device outputs the detection point cloud, including:

[0007] After collecting point cloud data with the spiral scanning lines of at least three lidars in the same transverse section maintaining a constant longitudinal spacing along the direction of travel, the geometric model of the subway tunnel is reconstructed based on the scanned point cloud. Through panoramic fusion and unified mapping to the coordinate system of the measurement center, multi-angle data fusion of the point cloud is achieved, forming a 360-degree closed point cloud envelope loop without blind spots, resulting in a panoramic, unobstructed detection point cloud.

[0008] When at least two lidars arranged in a collinear longitudinal interval scan, a phase nesting algorithm is used to ensure that the spiral scanning line of the latter lidar seamlessly fills the scanning gap of the former lidar, so that at least two lidars can jointly scan the same section of the subway tunnel to obtain a high-precision, high-density detection point cloud.

[0009] Preferably, when at least three lidars scan within the same transverse section, the main control device controls the lidars to reconstruct the tunnel geometric model based on the scanned point cloud using a built-in coordinate transformation matrix, and maps the point cloud data of at least three lidars to the coordinate system of the measurement center in a unified manner, forming a 360-degree closed point cloud envelope loop without blind spots.

[0010] Preferably, after obtaining a 360-degree closed point cloud envelope loop with no blind spots from at least three initial scans of the same transverse section, the same environmental features scanned by the arch radar and the side wall radar are simultaneously extracted within the overlapping field of view of the transition zone between the arch and the side wall of the point cloud. These features include at least the joints of the pipe segments and / or bolt holes. The coordinates of the same environmental features are then determined to be deviated. If so, the deviation is calculated, and the extrinsic parameter matrix of the lidar is adjusted according to the deviation to calibrate the lidar coordinates.

[0011] Preferably, during the data acquisition process of at least three lidars within the same transverse cross section, the main control device locks the phase difference of the lidars through a hardware synchronization signal and monitors the vehicle speed of the moving chassis in real time; based on the real-time monitored vehicle speed, it automatically adjusts the rotation speed of the lidars to ensure that the spacing between the spiral scanning lines of at least three lidars along the driving direction remains constant.

[0012] Preferably, when at least two lidars arranged longitudinally and collinearly perform scanning, a time-difference phase nesting algorithm is used. Based on the physical distance between each lidar in the longitudinal direction and the real-time vehicle speed, the time interval between two adjacent lidars passing through the same physical cross-section in the tunnel is determined and coupled with the rotation period of the lidars. This ensures that the spiral scan line of the subsequent lidar seamlessly fills the scanning gap of the previous lidar, achieving phase nesting, and allowing at least two lidars to jointly scan the same cross-section of the tunnel. The time interval is the sum of any integer number of rotation periods and half a rotation period of the lidar.

[0013] Preferably, the detection point cloud is homogenized through pose compensation and / or voxel mesh downsampling:

[0014] The pose-compensated homogenization process includes: determining the pose of the mobile chassis based on the three-dimensional pose and displacement changes of the mobile chassis detected by the sensor; performing motion distortion correction on each frame of the detection point cloud based on the pose and timestamp, determining the instantaneous pose of each point in each frame of the point cloud using an interpolation algorithm, and transforming the detection point cloud from the local coordinate system of the lidar to the global coordinate system to make the longitudinal density of the detection point cloud consistent.

[0015] The homogenization process based on voxel grid downsampling divides the point cloud space into a uniform three-dimensional voxel grid. Within each voxel, its centroid or a single point represents all points within the three-dimensional voxel grid, thereby achieving spatial homogenization of the detected point cloud density.

[0016] Preferably, deformation of subway tunnel space based on target detection point cloud includes:

[0017] The target detection point cloud is used as the point cloud to be registered and is registered with the reference point cloud using a hierarchical registration strategy of first coarse registration and then fine registration.

[0018] The registered tunnel point cloud is sliced ​​along the tunnel's central axis at a preset fixed interval to generate multiple continuous tunnel cross-sectional point cloud slices.

[0019] For each tunnel cross-section point cloud slice, preprocessing is performed, and a statistical outlier removal algorithm is used to filter out floating noise points;

[0020] The least squares method was used to fit an ellipse to the tunnel cross-section point cloud after filtering out floating noise points, and the ellipse parameters were solved.

[0021] The eccentricity of the ellipse is calculated based on the ellipse parameters, and the deformation is determined based on the eccentricity.

[0022] The coarse registration includes:

[0023] Based on the ISS key point algorithm, ISS key points are detected in the reference point cloud and the point cloud to be registered, and ISS key points are selected. For each ISS key point, an FPFH feature descriptor is calculated, and the change of normal vector in the neighborhood of the ISS key point is characterized by the point feature histogram.

[0024] The kd-tree is used to perform nearest neighbor search in the feature space. For each ISS key point in the point cloud to be registered, the candidate point closest to the FPFH feature descriptor is found in the reference point cloud. The nearest neighbor distance ratio test strategy is adopted to retain matching points whose ratio of nearest neighbor distance to second nearest neighbor distance is less than a preset threshold.

[0025] Based on the sample consistency initial registration strategy, three pairs of matching points are randomly sampled to calculate the transformation matrix, and the fit of other matching points under the transformation matrix is ​​evaluated. After multiple iterations, the transformation matrix with the most matching points is selected as the coarse registration result, and the initial rotation matrix and translation vector are obtained. Based on the coarse registration result, the point cloud to be registered is initially aligned to the reference point cloud.

[0026] The fine registration is based on a robust iterative nearest point (ICP) algorithm with refined iterative optimization, including:

[0027] In each iteration, a nearest neighbor search strategy based on dynamic threshold is combined with a kd-tree accelerated search algorithm to search for the nearest neighbor point in the reference point cloud for each point in the initial registration point cloud to be registered.

[0028] Assign a weight to each point and the point pair formed by its nearest neighbor, and perform weighted singular value decomposition based on the weights to obtain the rigid body transformation matrix with the minimum weighted squared distance.

[0029] Through multiple iterations, until the convergence condition is met, the fine registration is completed; the convergence condition includes reaching the required number of iterations or the difference between the rigid body transformation matrices of two adjacent iterations being less than a preset threshold.

[0030] Preferably, the encroachment detection of subway tunnels based on target detection point clouds includes:

[0031] The target detection point cloud is aligned with the tunnel standard clearance three-dimensional model constructed based on the subway tunnel design parameters in a spatial coordinate system. The tunnel standard clearance three-dimensional model represents the clearance envelope surface that allows trains to pass safely.

[0032] For each measured point in the target detection point cloud, the nearest neighbor point is found on the surface of the tunnel standard clearance 3D model based on the nearest neighbor search algorithm, and the spatial position relationship of the measured point relative to the tunnel standard clearance 3D surface is calculated; if the coordinates of the measured point are determined to be inside the net envelope surface of the tunnel standard clearance 3D model according to the spatial position relationship, the measured point is determined to be an encroaching point, otherwise it is a non-encroaching point.

[0033] Output the intrusion limit determination status of all measured points, mark the intrusion limit points in the target detection point cloud and output the display, and use cluster analysis to aggregate adjacent intrusion limit points into intrusion limit regions, calculate the center coordinates, range and maximum intrusion limit value of the intrusion limit regions, and generate an intrusion limit detection report.

[0034] Preferably, foreign object detection in subway tunnels based on target detection point clouds includes:

[0035] After normalizing the target detection point cloud and the original tunnel point cloud without foreign objects, the data is input into a pre-trained point cloud registration neural network to extract the topological information of the geometric structure of the target detection point cloud and the original tunnel point cloud, thus obtaining structural features. By learning the three-dimensional coordinate data of the original tunnel point cloud, the positional distribution features of the original tunnel point cloud in the global space are extracted.

[0036] By fusing structural features and location distribution features, the optimal rigid body transformation matrix of the target detection point cloud relative to the original tunnel point cloud is predicted by regression, and point cloud registration is performed based on the optimal rigid body transformation matrix.

[0037] The nearest neighbor residual distance from each point in the registration point cloud to the surface of the original tunnel point cloud is compared with the error threshold; based on the comparison result, it is determined whether it is a foreign object point.

[0038] Cluster and extract all foreign object points to form a set of foreign object targets and generate a detection report output and / or graphically visualize the location and / or geometric contour of the foreign object points.

[0039] In another aspect, the present invention provides a subway tunnel space safety detection device based on 3D point clouds. The subway tunnel space safety detection method based on 3D point clouds performs subway tunnel space safety detection, comprising: a mobile chassis and a laser device mounted on the mobile chassis. The mobile chassis moves within the subway tunnel space along a preset moving direction. The laser device collects three-dimensional point cloud data of the inner wall of the subway tunnel through a spiral scanning method. The scanned detection point cloud is pre-processed to homogenize and form a target detection point cloud. Deformation, encroachment, and foreign object detection of the subway tunnel space are performed based on the target detection point cloud. The laser device includes at least three laser radars uniformly arranged circumferentially within the same transverse cross section of the mobile chassis, or at least two laser radars arranged equidistantly along the longitudinal axis of the mobile chassis.

[0040] The present invention is based on scanning with at least two lidars arranged longitudinally or at least three lidars arranged laterally in a linear configuration. This enables scanning without blind spots and, through lidar scanning control, effectively acquires target detection point cloud data for tunnel space safety inspection. It achieves uniform scanning without blind spots, resolves the contradiction between high-speed detection and high-density coverage, and provides a new detection method for subway tunnel space safety inspection. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the first embodiment of the subway tunnel space safety detection device based on 3D point cloud of the present invention.

[0042] Figure 2 This is a schematic diagram of the second embodiment of the subway tunnel space safety detection device based on 3D point cloud of the present invention. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0044] In terms of data processing, motion compensation and voxel downsampling are used to achieve point cloud homogenization, and a layered strategy is used to complete high-precision registration. In particular, this invention uses a direct 3D calculation method based on nearest neighbor search for intrusion detection, avoiding projection errors; and uses a point cloud registration method based on deep learning (bi-branch feature extraction) for foreign object detection, solving the registration problem in weak texture environments.

[0045] In an exemplary embodiment of this application, the 3D point cloud-based subway tunnel space safety detection method involves installing a laser device on a mobile chassis and moving it along a preset direction within the subway tunnel space. The laser device collects three-dimensional point cloud data of the inner wall of the subway tunnel through a spiral scanning method. The scanned detection point cloud is pre-processed to homogenize and form a target detection point cloud. Based on the target detection point cloud, deformation, encroachment, and foreign object detection are performed in the subway tunnel space. See also... Figure 1 as well as Figure 2 As shown, the laser device described in this application includes three lidars, consisting of a first lidar 101, a second lidar 102, and a third lidar 103. The three lidars are arranged in an equilateral triangle within the same transverse section of the mobile chassis. Figure 1 As shown, or arranged at equal intervals along the longitudinal direction of the mobile chassis axis, such as... Figure 2 As shown, in specific implementations, when arranged on the same cross section, there can be more than three, and they can be evenly arranged along the circumference of the same cross section to achieve the present invention. When arranged longitudinally, there can be at least two, not limited to three, and there can be more.

[0046] For ease of description or brevity, the specification of this application uses an example of arranging three lidars horizontally and vertically to illustrate in detail how the technical solution of this application is implemented. In this layout of three lidars, the lidar device outputs a detection point cloud, including:

[0047] Point cloud data is collected by using three lidars within the same transverse cross-section with their spiral scanning lines kept at a constant longitudinal spacing along the direction of travel. Based on the scanned point clouds, geometric models of the subway tunnel's arch, sidewalls, and roadbed are reconstructed. Through panoramic fusion and unified mapping to the measurement center coordinate system, multi-angle data fusion of the point cloud in the transition area between the arch and sidewalls is achieved, forming a 360-degree closed, blind-spot-free point cloud envelope, ultimately resulting in a panoramic, unobstructed detection point cloud. When scanning with three lidars arranged collinearly at longitudinal intervals, a phase nesting algorithm is used to seamlessly fill the scanning gaps of the previous lidar with the spiral scanning lines of the subsequent lidars, enabling the three lidars to jointly scan the same cross-section of the subway tunnel, resulting in a high-precision, high-density detection point cloud.

[0048] In one embodiment, the mobile chassis 400 adopts a high-strength aluminum alloy frame structure and is equipped with a four-wheel suspension system to ensure stability when driving on uneven road surfaces in tunnels. It is equipped with a drive unit 300, including two high-torque brushless DC motors that drive the two rear wheels respectively. Differential speed control is used to achieve steering of the mobile chassis. The laser device includes three lidars with a horizontal field of view. It has a maximum measurement radius of 120 meters and is equipped with a main control device of 200, which is a ruggedized industrial computer that communicates with the lidar and camera via a gigabit Ethernet interface, communicates with the drive controller via a CAN bus, and communicates with the industrial control computer.

[0049] In existing technologies, mobile measurement systems typically employ a single lidar for helical scanning. When the vehicle travels at high speeds, the distance between individual helical scanning lines (i.e., the pitch) increases, leading to "blind spots" that easily miss tunnel gaps or traces of leakage. To address the conflict between "high efficiency (high vehicle speed)" and "high precision (high density)," this invention proposes a multi-radar cooperative detection technology and a preset control strategy. The main control device 200 utilizes a built-in control algorithm to control the radar scanning action based on the predicted control strategy, achieving blind-spot-free scanning.

[0050] According to the preset control strategy, in one embodiment, when three lidars within the same transverse section are used for scanning, the main control device 200 executes a three-dimensional envelope collaborative control and self-calibration strategy to achieve high-precision full-section scanning. Specifically, the main control device can control the top lidar to reconstruct the tunnel arch geometric model based on the coordinate transformation matrix of the built-in triangular geometric parameters, and the side lidars to reconstruct the tunnel sidewalls and roadbed geometric models. The point cloud data of the three lidars are uniformly mapped to the measurement center coordinate system to form a 360-degree closed point cloud envelope loop without blind spots.

[0051] If more lidars are used for data scanning, the required coordinate transformation matrix can be formed based on the rectangular or polygonal geometric parameters formed by the arrangement of the lidars. This allows multiple lidars to be controlled to reconstruct models of different locations on the tunnel cross section. Then, the point cloud data from multiple lidars are uniformly mapped to the coordinate system of the measurement center, forming a 360-degree closed point cloud envelope loop without blind spots.

[0052] Furthermore, according to a preset control strategy, in one embodiment, after the initial scanning of three lidars within the same transverse cross-section to obtain a 360-degree closed point cloud envelope without blind spots, the same environmental features scanned simultaneously within the overlapping field of view of the point cloud, including at least the pipe segment joints and / or bolt holes, are extracted and compared. It is then determined whether the coordinates of the same environmental features will cause slight displacement of the radar due to the vibration of the moving chassis, resulting in a deviation in the coordinates of the same feature point between the two radars. If so, then calculate the deviation. According to the deviation Adjusting the extrinsic parameter matrix of the lidar enables millisecond-level automatic calibration and closed-loop correction.

[0053] Taking three lidars as an example, in the equilateral triangular arrangement of the three lidars in this application, the scanning space of the lidars can be divided into three logical sectors: a top sector and two side sectors. The main control device has pre-set parameters based on triangle geometry (such as the height of the triangle's vertices). Bottom width Side tilt angle The coordinate transformation matrix of the top second lidar 102 is used to reconstruct a high-precision geometric model of the tunnel arch. The data of the first lidar 101 and the third lidar 103 on both sides are used to reconstruct the tunnel sidewall and roadbed model. Finally, the three local point clouds are mapped to the measurement center coordinate system in real time through panoramic fusion. Based on spatial envelope mapping, a 360-degree closed envelope loop without blind spots is formed, which effectively solves the problem of scanning blind spots of a single radar in complex curved tunnels.

[0054] In the triangular layout of the lidar, the second lidar 102 at the top and the first lidar 101 and the third lidar 103 on the side have significant field-of-view overlap in the 10 o'clock and 2 o'clock directions of the tunnel cross section (i.e., the transition area between the arch and the side wall). This application realizes real-time accuracy verification of the overlapping area by using the dynamic self-calibration algorithm for the double overlap area, eliminating overlapping data, and using the interparametric nature of the triangular structure to ensure extremely high consistency of point cloud stitching in high-speed mobile operations.

[0055] According to a preset control strategy, in one embodiment, to prevent the laser beams of the three radars from interfering with each other in space (crosstalk) and to ensure that the spiral scan lines are evenly distributed on the tunnel wall, based on an interleaved phase and velocity coupling control strategy, the main control device can establish a dynamic coupling model of travel speed-radar rotation speed-phase offset. During the data acquisition process of the three lidars within the same transverse section, the main control device forcibly locks the phase difference of the lidars through a hardware synchronization signal. For example, the phase of the second lidar 102 at the top is set to... The first lidar on the left, 101, is The third laser auxiliary radar on the right is 103. The system monitors the vehicle speed of the mobile chassis in real time. Based on this real-time speed monitoring, it automatically adjusts the rotation speed of the lidar to maintain a constant spacing between the spiral scan lines of the three lidars along the driving direction. Specifically, speed coupling is achieved using the following formula:

[0056] ,in ;

[0057] in, Indicates the speed of the vehicle on the mobile chassis. For the number of lidar units, Indicates the radar rotation speed. This indicates the spacing of the helical scan lines of the lidar along the driving direction.

[0058] The main control equipment monitors vehicle speed in real time, and automatically adjusts the radar speed when the vehicle speed changes. Ensure that the spacing of the three sets of interlaced spiral scan lines in the longitudinal direction of the tunnel remains constant (e.g., maintained at a distance of 100 mm). (within), thus improving the detection speed by three times while still obtaining dense and uniform point cloud data.

[0059] In detection schemes using collinear LiDAR arrays, the challenge lies in spatiotemporal synchronization. To address this, this application proposes a time-difference phase nesting algorithm for scanning control. According to a preset control strategy, in one embodiment, when three LiDARs arranged collinearly with longitudinal spacing scan, the time-difference phase nesting algorithm is used. Based on the physical distance between each LiDAR in the longitudinal direction and the real-time vehicle speed, the time interval between two adjacent LiDARs passing through the same physical cross-section within the tunnel is determined and coupled with the LiDAR's rotation period. This ensures that the spiral scan line of the subsequent LiDAR seamlessly fills the scanning gap of the preceding LiDAR, achieving phase nesting. The three LiDARs then jointly scan the same cross-section of the tunnel. The time interval is the sum of any integer number of LiDAR rotation periods and half a rotation period.

[0060] Because the three radars arranged longitudinally are positioned one in front of the other, after one radar scans a certain area, the next radar needs a certain amount of time to scan before scanning a certain area. Only when the two radars are synchronized, i.e., have the same phase, can the scanning density be increased. This problem is solved by using a time-difference phase nesting algorithm to control the scanning of the second radar to be exactly in the middle of the scanning line of the first radar.

[0061] Specifically, the main control equipment is set to monitor vehicle speed in real time. The distance between the front and rear radars is 1 meter, and the vehicle speed is... If the time difference is 1 second, then the rotation speed of the lidar is... (i.e., 0.1 seconds per revolution), then 1 second is exactly 10 whole cycles, with the phase remaining unchanged; the main control equipment fine-tunes the vehicle speed or phase so that the time difference corresponds to " "one cycle", If is any integer, then the subsequent radar will scan exactly in the middle of the scan line of the previous radar.

[0062] The formula used to calculate the time difference between adjacent radars passing through the same physical cross section is as follows:

[0063] ;

[0064] (Time difference) is the time interval between two adjacent lidars passing through the same physical cross section inside the tunnel. (Physical spacing) refers to the physical installation distance of each lidar in the direction of travel of the mobile chassis; (Traveling speed) is the real-time speed at which the current mobile chassis is traveling along the tunnel.

[0065] To achieve the best point cloud encryption effect, the vehicle speed was fine-tuned. Or radar phase, causing time difference The following relationship must be satisfied:

[0066] ;

[0067] For the rotation period of the lidar (e.g.) At rotational speed, ), It can be any integer.

[0068] When the time difference corresponds to "an integer number of cycles + half a cycle", the scan line of the subsequent radar will precisely fall at the center of the spiral gap left by the previous radar. Thus, through the coupled control logic between vehicle speed and radar rotation cycle, a nested phase scanning is achieved, ensuring that the scan line of the subsequent radar fills the scanning gap of the previous radar, thereby achieving uniform coverage without blind spots. Ultimately, this application, based on the control coupled with physical spacing and vehicle speed, makes the longitudinally arranged radar array resemble a multi-headed sewing machine, weaving an extremely dense point cloud network while moving at high speed.

[0069] In an optional embodiment, the detected point cloud is homogenized through pose compensation or voxel mesh downsampling:

[0070] The pose-compensated homogenization process includes: determining the pose of the mobile chassis based on the three-dimensional pose and displacement changes of the mobile chassis detected by the sensor; performing motion distortion correction on each frame of the detection point cloud based on the pose and timestamp, determining the instantaneous pose of each point in each frame of the point cloud using an interpolation algorithm, and transforming the detection point cloud from the local coordinate system of the lidar to the global coordinate system to make the longitudinal density of the detection point cloud consistent.

[0071] During mobile scanning, speed fluctuations in the mobile device can lead to uneven distribution of point clouds longitudinally within the tunnel. Motion data is collected in real time by an IMU (Inertial Measurement Unit) and an odometer built into the mobile chassis. The IMU outputs triaxial acceleration and angular velocity at a frequency of 100Hz, while the odometer outputs displacement increments at a frequency of 50Hz. The data processing unit then uses a fusion algorithm based on extended Kalman filtering to estimate the precise pose (position and attitude) of the mobile device in real time. For each frame of point cloud acquired by the LiDAR, motion distortion correction is performed based on the pose of the mobile device corresponding to its timestamp. Specifically, for each point within a frame of point cloud, based on its relative acquisition time within that frame, the instantaneous pose obtained through interpolation is used to transform it from the local coordinate system of the LiDAR to the global coordinate system, effectively compensating for longitudinal stretching or compression of the point cloud caused by vehicle acceleration and deceleration.

[0072] In the embodiments of this application, the uniformization processing based on voxel grid downsampling involves dividing the point cloud space into a uniform three-dimensional voxel grid. Within each voxel, its centroid or a single point represents all points within the three-dimensional voxel grid, achieving spatial uniformity of the high-density original detection point cloud density. Specifically, the following steps are employed:

[0073] Step 1: Create a 3D voxel mesh;

[0074] Calculate the bounding box of a point cloud: First, find the minimum and maximum coordinates of the entire point cloud in 3D space. This forms a cuboid region that can encompass all points.

[0075] Define voxel size: Specifies the voxel dimension, which is the side length of a cube voxel. For example, set the voxel dimension to 0.01 meters (1 centimeter).

[0076] Mesh division: Based on the size of the bounding box and the set voxel size, the entire bounding box space is divided into countless neatly arranged tiny cubes with a side length of 1cm; each small cube is a voxel.

[0077] Step 2: Assign point clouds to voxels;

[0078] Traverse each point in the original detection point cloud; based on the 3D coordinates of each point... Calculate the voxel to which it belongs and assign it a voxel index, discarding empty voxels. The voxel index calculation method is as follows:

[0079] The voxel index _i = floor( (point coordinates _i - minimum coordinate value _i) / voxel size); where i represents Three dimensions, floor is the floor function.

[0080] Step 3: Generate representative points;

[0081] Iterate through each non-empty voxel, and for each voxel containing at least one point, perform an aggregation operation. The method is as follows: 1) Calculate the centroid: add up the coordinates of all points within the voxel, then divide by the number of points to obtain an average coordinate point, which is the representative point of the voxel; 2) Representative point coordinates = (coordinates of point 1 + coordinates of point 2 + ... + coordinates of point n) / n.

[0082] Step 4: Output the new point cloud;

[0083] Collect the representative points generated by all non-empty voxels to form a new point cloud output, resulting in a simplified point cloud with a uniform distribution.

[0084] In one embodiment, deformation of the subway tunnel space based on the target detection point cloud includes:

[0085] The target detection point cloud is used as the point cloud to be registered and the reference point cloud. A hierarchical registration strategy of first coarse registration and then fine registration is adopted for registration. The complete tunnel point cloud formed after registration is sliced ​​along the tunnel central axis at a preset fixed interval to generate multiple continuous tunnel cross-section point cloud slices. Each tunnel cross-section point cloud slice is preprocessed, and a statistical outlier removal algorithm is used to filter out floating noise points. The tunnel cross-section point cloud with floating noise points filtered out is ellipse fitted using the least squares method to solve for the ellipse parameters. The ellipse eccentricity is calculated based on the ellipse parameters, and the deformation is determined based on the ellipse eccentricity.

[0086] Specifically, the registered complete tunnel point cloud is cut along the tunnel's designed central axis at fixed intervals of 0.5 meters, generating hundreds of continuous cross-sectional point cloud slices; outlier removal is performed on each slice, and the number of neighboring points is calculated. The standard deviation factor was set to 1.5 to remove floating noise points. For the denoised sliced ​​point cloud, ellipse fitting was performed using the least squares method, based on the general quadratic equation of an ellipse. and satisfy the constraints. The ellipse parameters are obtained by solving a system of linear equations, and then converted into the coordinates of the ellipse center. Long half shaft short half shaft and rotation angle Calculate the eccentricity of the ellipse For an ideal circle, Based on tunnel design specifications and extensive experimental data, a deformation alarm threshold of 0.15 was set. When the eccentricity of a certain slice... If the cross-section is deemed to have a significant risk of deformation, its mileage location is recorded and highlighted in the 3D model.

[0087] This application adopts a layered registration strategy, which combines coarse and fine registration to ensure high-precision and robust point cloud alignment even in complex tunnel environments.

[0088] In this embodiment of the application, the goal of the coarse registration is to quickly provide a near-correct initial transformation and avoid the fine registration from getting trapped in local optima, including:

[0089] First, based on the ISS key point algorithm, ISS key point detection is performed on the reference point cloud and the point cloud to be registered, and key points with high repeatability are extracted and ISS key points are selected.

[0090] The ISS keypoint algorithm evaluates the saliency of each point based on the variation of its feature values ​​within its local neighborhood. For a given point... Set a local support radius (e.g., 0.2m), calculate the covariance matrix of all points within the spherical neighborhood and calculate the eigenvalues ​​{ }( By setting a threshold (e.g.) and The corner or boundary points that change significantly in each direction are selected. These corner or boundary points are usually located at stable structures such as segment joints and bolt holes, and are used as key points of the ISS.

[0091] Secondly, a 33-dimensional FPFH feature descriptor is calculated for each ISS keypoint, and the neighborhood (radius) of the ISS keypoint is characterized by the point feature histogram (SPFH). The changes in the normal vector within the point cloud exhibit good invariance to rotation, translation, and density changes.

[0092] Finally, a kd-tree is used to perform nearest neighbor search in the feature space to find the candidate point in the reference point cloud that is closest to the FPFH feature descriptor for each ISS key point in the point cloud to be registered. The nearest neighbor distance ratio (NNDR) test strategy is adopted to eliminate a large number of mismatches contained in the initial matching and retain matching points whose ratio of nearest neighbor distance to second nearest neighbor distance is less than a preset threshold (such as 0.8) to improve the uniqueness of matching.

[0093] Based on the Sample Consistency Initial Registration (SAC-IA) strategy, three pairs of matching points are randomly sampled to calculate the transformation matrix, and the matching degree of other matching points under the transformation matrix is ​​evaluated. After multiple iterations, the transformation matrix with the most matching points (interior points) is selected as the coarse registration result, and the initial rotation matrix and translation vector are obtained. This process can effectively resist the interference of mismatches. Based on the coarse registration result, the point cloud to be registered is initially aligned to the reference point cloud, and the expected accuracy can reach 5-10 cm.

[0094] In this embodiment of the application, the fine registration is performed based on the robust Iterative Closest Point (ICP) algorithm with fine iterative optimization, including:

[0095] In each iteration, a dynamic threshold nearest neighbor search strategy, combined with a kd-tree accelerated search algorithm, is used to search for the nearest neighbor in the reference point cloud for each point in the initial registration point cloud. To improve efficiency and ensure accuracy, a maximum corresponding distance threshold is dynamically set when using the kd-tree accelerated search. Its initial value can be set to 0.1m, and it gradually decreases as the number of iterations increases.

[0096] Since not all point pairs are equally reliable, to reduce the impact of noise and outliers (such as mobile workers and temporary facilities), point pairs are formed for each point and its nearest neighbor. Assign a weight; where the weight is calculated based on the distance between the points. , or normal vector consistency , , Let be the normal vector of the point pair.

[0097] Based on all valid point pairs and their weights, a weighted least squares problem is constructed and solved using weighted singular value decomposition (SVD) to obtain the rigid body transformation matrix with the minimum weighted sum of squared distances. Through multiple iterations, until the convergence condition is met, fine registration is completed. The convergence condition includes reaching the required number of iterations or the difference between the rigid body transformation matrices of two adjacent iterations (expressed as a rotation angle variation). Measurement of translational shift ,like radian, The accuracy (in meters) is less than the preset threshold. After fine registration, the two point clouds can be aligned to millimeter-level precision (typically 2-3 mm), providing a precise coordinate basis for subsequent deformation detection.

[0098] In one embodiment of this application, encroachment detection of a subway tunnel based on a target detection point cloud includes:

[0099] The target detection point cloud is aligned with the standard tunnel clearance 3D model constructed based on the subway tunnel design parameters, with the surface normal vector pointing outwards. This standard tunnel clearance 3D model represents the clearance envelope surface that allows safe train passage. For each measured point... Based on the nearest neighbor search algorithm, the nearest neighbor point is found on the surface of the standard tunnel clearance model. The system calculates the spatial positional relationship of the measured point relative to the three-dimensional surface of the tunnel standard clearance. If the coordinates of the measured point are determined to be inside the net envelope of the three-dimensional model of the tunnel standard clearance based on the spatial positional relationship, then the measured point is determined to be an encroachment point. Finally, the encroachment determination results of all measured points are output and three-dimensional visualization is performed. At the same time, a density clustering algorithm is used to aggregate adjacent encroachment points into encroachment regions, calculate and extract the center coordinates, range and maximum encroachment depth of each region, and automatically generate a structured inspection report.

[0100] In a standard 3D tunnel clearance model constructed based on the design parameters of a subway tunnel, each measured point in the target detection point cloud is traversed using a nearest neighbor search algorithm. nearest neighbor (i.e., the closest point on the surface of the tunnel clearance model to the measured point), and obtain its unit normal. Calculate the direction vector With unit normal direction The dot product yields the directed distance. If there is a directional distance And its absolute value is greater than the set safety tolerance threshold. Then the measured point is determined to be the intrusion point (i.e., point). If the point is located inside the standard tunnel clearance model, it is considered a non-encroaching point, i.e., a normal point. The encroachment judgment status of all measured points is output, and the encroachment points are marked in the target detection point cloud (i.e., in the 3D point cloud data, all points judged as encroaching are specially marked, for example, assigned different colors, such as red) for output visualization. At the same time, cluster analysis is used to aggregate adjacent encroaching points into encroaching regions, calculate the center coordinates, range and maximum encroachment value of the encroaching region, and generate an encroachment detection report.

[0101] In this application, after the inspection is completed, a 3D point cloud model marked with encroachment points is visually displayed using visualization software. Users can clearly see which areas within the tunnel have encroachment points and can further query the specific location and depth of the encroachment points, i.e., the distance. The value of .

[0102] In order to accelerate the search process, when traversing each measured point in the target detection point cloud based on the nearest neighbor search algorithm, a spatial index structure can be constructed based on the preprocessed target detection point cloud, such as a kd-tree or an octree, for nearest neighbor search; and an adjacency graph can be constructed using the Euclidean distance calculation results for later use.

[0103] In this application, the three-dimensional model of the tunnel clearance can be a fine mesh model or a high-density point cloud model, which accurately represents the allowable clearance range of the tunnel.

[0104] In one embodiment of this application, foreign object detection in a subway tunnel based on a target detection point cloud includes:

[0105] After normalizing the target detection point cloud and the original tunnel point cloud without foreign objects, the data is input into a pre-trained point cloud registration neural network to extract the topological information of the geometric structure of the target detection point cloud and the original tunnel point cloud, thus obtaining structural features. By learning the three-dimensional coordinate data of the original tunnel point cloud, the positional distribution features of the original tunnel point cloud in the global space are extracted.

[0106] By fusing structural features and location distribution features, the optimal rigid body transformation matrix of the target detection point cloud relative to the original tunnel point cloud is predicted by regression, and point cloud registration is performed based on the optimal rigid body transformation matrix.

[0107] The nearest neighbor residual distance from each point in the registration point cloud to the surface of the original tunnel point cloud is compared with the error threshold; based on the comparison results, it is determined whether it is a foreign object point.

[0108] Cluster and extract all foreign object points to form a set of foreign object targets and generate a detection report output and / or graphically visualize the location and / or geometric contour of the foreign object points.

[0109] This application utilizes deep neural networks to extract deep features through deep learning for high-precision registration of foreign objects, achieving robust foreign object recognition and solving the problem that traditional rigid registration algorithms are prone to failure in environments such as tunnels with repetitive textures and simple features.

[0110] In this application, the original tunnel point cloud is retrieved from the system database as the original tunnel data of the section in a non-objective state, and used as a comparison benchmark. Before inputting into the deep neural network, the two sets of point cloud data undergo normalization processing, mapping the point cloud coordinate values ​​to a standard unit space (e.g., the [0,1] interval) to eliminate absolute coordinate differences caused by different acquisition locations. Then, the two pre-processed sets of point clouds are input into a pre-trained point cloud registration neural network.

[0111] The point cloud registration neural network described in this application includes two branches:

[0112] The Structural Feature Extraction branch is used to capture geometric topological information, including local geometric structures such as tunnel walls, segment joints, and pipe supports, using Graph Convolutional Networks (GCNs) or point cloud convolution operators. The Coordinate Feature Extraction branch is used to directly process the raw 3D coordinate data and learn the positional distribution characteristics of the point cloud in the global space.

[0113] After high-precision registration is completed, the nearest neighbor residual distance (Registration Error) from each point in the real-time point cloud to the surface of the original tunnel data is calculated, and compared with the set foreign object detection threshold. (For example, 5cm) Comparison, if the registration error distance at a certain point is less than This is determined to be an inherent structure of the tunnel (background). If the registration error distance at a certain point is greater than... This means that the point does not have a corresponding structure in the original data, and it is determined to be a foreign object point. All points determined to be foreign objects are clustered and extracted to generate a set of foreign object targets.

[0114] Specifically, intuitive inspection reports can be generated in the human-computer interaction interface, such as displaying the original tunnel structure in green or gray and highlighting the marked foreign object area in red. The visualization results not only show the presence of foreign objects, but also accurately reflect the specific location (such as the track area and sidewall area) and geometric outline of foreign objects in the tunnel cross section, assisting operators in quickly formulating cleanup plans.

[0115] Another aspect of this invention provides a subway tunnel space safety detection device based on 3D point clouds. The device performs subway tunnel space safety detection according to a 3D point cloud-based method, including: a mobile chassis 400 and a laser device mounted on the mobile chassis. The mobile chassis moves within the subway tunnel space along a preset direction. The laser device collects three-dimensional point cloud data of the inner wall of the subway tunnel through a helical scanning method. The scanned detection point cloud is pre-processed to homogenize and form a target detection point cloud. Based on the target detection point cloud, deformation, encroachment, and foreign object detection are performed within the subway tunnel space. The laser device includes at least three lidars, evenly distributed circumferentially within the same cross-section, such as... Figure 1 As shown, an embodiment consisting of a first lidar 101, a second lidar 102, and a third lidar 103 is illustrated. The three lidars are arranged in an equilateral triangle within the same transverse section of the mobile chassis, as shown in the diagram. Figure 1 As shown, or at least two lidar sensors, are arranged at equal intervals along the longitudinal direction of the mobile chassis axis, such as... Figure 2 As shown, a structure employing three lidar sensors arranged equidistantly along the longitudinal axis of a mobile chassis is illustrated.

[0116] In some embodiments, the mobile chassis 400 is also equipped with at least one global shutter CCD industrial camera, mounted next to the lidar, for acquiring tunnel surface texture, and is equipped with a set of high-brightness LED fill lights to provide uniform illumination for the CCD camera in the dark tunnel environment.

[0117] The detection device in this embodiment scans within the tunnel, transmitting data to the main control device in real time. After preprocessing the point cloud data, the main control device initiates deformation, encroachment, and foreign object detection algorithms in parallel. The results of each detection algorithm are integrated into a comprehensive report, which is displayed in a visual interface with a 3D tunnel model as the background. The report highlights the deformation cross-section (e.g., red), encroachment area (e.g., yellow), and foreign objects (e.g., purple) in different colors, and lists detailed mileage, dimensions, and risk level information. It supports one-click generation and export of the detection report, achieving efficient, accurate, and automated detection of subway tunnel space safety, and providing reliable technical support for operational safety.

[0118] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention.

[0119] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the invention.

[0120] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A spatial safety detection method for subway tunnels based on 3D point clouds, characterized in that, A laser device is installed on a mobile chassis and moves along a preset direction within the subway tunnel space. The laser device collects three-dimensional point cloud data of the inner wall of the subway tunnel through a spiral scanning method. The scanned detection point cloud is homogenized and preprocessed to form a target detection point cloud. Based on the target detection point cloud, deformation, encroachment, and foreign object detection in the subway tunnel space are performed. The laser device includes three lidars that are uniformly arranged in the same transverse section of the mobile chassis along the circumference, or at least two lidars that are collinear and equidistantly spaced along the axis of the mobile chassis. When three lidars scan within the same cross-section, the laser device outputs a detection point cloud, including: After collecting point cloud data with the spiral scanning lines of three lidars maintaining a constant longitudinal spacing along the driving direction within the same transverse cross-section, the lidars are controlled to reconstruct the geometric model of the subway tunnel based on the scanned point cloud using a built-in coordinate transformation matrix. Through panoramic fusion and unified mapping to the measurement center coordinate system, multi-angle data fusion of the point cloud is achieved, forming a 360-degree closed, blind-angle-free point cloud envelope, resulting in a panoramic, unobstructed detection point cloud. The three lidars within the same transverse cross-section consist of three lidars arranged in an equilateral triangular pattern. The scanning space of the three lidars is divided into a top sector and two... In the flank area, the geometric model of the subway tunnel includes the geometric models of the tunnel's arch, sidewalls, and roadbed. After obtaining a 360-degree closed point cloud envelope with no blind spots through the initial scan, the same environmental features scanned by the radar are simultaneously extracted within the overlapping area of ​​the point cloud's field of view. It is then determined whether the coordinates of the same environmental features have deviated. If so, the deviation is calculated, and the extrinsic parameter matrix of the lidar is adjusted according to the deviation to calibrate the lidar coordinates. Based on the real-time monitored vehicle speed, the lidar rotation speed is automatically adjusted to ensure that the spacing between the spiral scan lines of the three lidars along the driving direction remains constant. When at least two lidars arranged longitudinally and collinearly perform scanning, a time-difference phase nesting algorithm is used for phase nesting processing. The time interval between two adjacent lidars passing through the same physical cross section in the tunnel is determined based on the physical distance between each lidar in the longitudinal direction and the real-time vehicle speed. This time interval is coupled with the rotation period of the lidars, so that the spiral scanning line of the latter lidar seamlessly fills the scanning gap of the former lidar, thus achieving phase nesting. This enables at least two lidars to jointly scan the same cross section of the subway tunnel, resulting in a high-precision, high-density detection point cloud.

2. The subway tunnel spatial safety detection method based on 3D point cloud as described in claim 1, characterized in that, The coordinate transformation matrix includes a coordinate transformation matrix based on the geometric parameters of triangles. Among the three laser scanning radars, the top laser radar scanning the top sector reconstructs the geometric model of the tunnel arch, and the two laser radars scanning the side wing areas reconstruct the model of the tunnel sidewalls and roadbed.

3. The subway tunnel spatial safety detection method based on 3D point cloud as described in claim 1, characterized in that, The same environmental features include at least segment joints and / or bolt holes.

4. The subway tunnel spatial safety detection method based on 3D point cloud as described in claim 1, characterized in that, During the data acquisition process of the three lidars within the same transverse cross section, the main control equipment locks the phase difference of the lidars through hardware synchronization signals and monitors the vehicle speed of the mobile chassis in real time.

5. The subway tunnel spatial safety detection method based on 3D point cloud as described in claim 1, characterized in that, When at least two lidars arranged collinearly and longitudinally spaced perform scanning, the time interval is the sum of any integer number of rotation cycles and half a rotation cycle of the lidars.

6. The subway tunnel spatial safety detection method based on 3D point cloud as described in claim 1, characterized in that, The detection point cloud is homogenized through pose compensation and / or voxel grid downsampling: Homogenization based on pose compensation includes: determining the pose of the moving chassis based on the three-dimensional pose and displacement changes detected by the sensor; performing motion distortion correction on each frame of the detection point cloud based on the pose and timestamp, determining the instantaneous pose of each point in each frame of the point cloud using an interpolation algorithm, and transforming the detection point cloud from the local coordinate system of the lidar to the global coordinate system to make the longitudinal density of the detection point cloud consistent; Homogenization based on voxel grid downsampling involves dividing the point cloud space into a uniform three-dimensional voxel grid, and using its centroid point or a single point within each voxel to represent all points in the three-dimensional voxel grid, thereby achieving spatial homogenization of the detection point cloud density.

7. The subway tunnel spatial safety detection method based on 3D point cloud as described in claim 1, characterized in that, Deformation detection in subway tunnel space based on target detection point cloud includes: The target detection point cloud is used as the point cloud to be registered and is registered with the reference point cloud using a hierarchical registration strategy of first coarse registration and then fine registration. The registered tunnel point cloud is sliced ​​along the tunnel's central axis at a preset fixed interval to generate multiple continuous tunnel cross-sectional point cloud slices. For each tunnel cross-section point cloud slice, preprocessing is performed, and a statistical outlier point cloud removal algorithm is used to filter out floating noise points; The least squares method was used to fit an ellipse to the tunnel cross-section point cloud after filtering out floating noise points, and the ellipse parameters were solved. Calculate the ellipse eccentricity based on the ellipse parameters, and determine whether there is deformation based on the ellipse eccentricity. The coarse registration includes: Based on the ISS key point algorithm, ISS key points are detected in the reference point cloud and the point cloud to be registered, and ISS key points are selected. For each ISS key point, an FPFH feature descriptor is calculated, and the change of normal vector in the neighborhood of the ISS key point is characterized by the point feature histogram. The kd-tree is used to perform nearest neighbor search in the feature space. For each ISS key point in the point cloud to be registered, the candidate point closest to the FPFH feature descriptor is found in the reference point cloud. The nearest neighbor distance ratio test strategy is adopted to retain matching points whose ratio of nearest neighbor distance to second nearest neighbor distance is less than 0.

8. Based on the sample consistency initial registration strategy, three pairs of matching points are randomly sampled to calculate the transformation matrix, and the fit of other matching points under the transformation matrix is ​​evaluated. After multiple iterations, the transformation matrix with the most matching points is selected as the coarse registration result, and the initial rotation matrix and translation vector are obtained. Based on the coarse registration preparation result, the point cloud to be registered is initially aligned to the reference point cloud. The fine registration is based on a robust iterative nearest point (ICP) algorithm with refined iterative optimization, including: In each iteration, a nearest neighbor search strategy based on dynamic threshold is combined with a kd-tree accelerated search algorithm to search for the nearest neighbor point in the reference point cloud for each point in the initial registration point cloud to be registered. Assign a weight to each point and the point pair formed by its nearest neighbor, and perform weighted singular value decomposition based on the weights to obtain the rigid body transformation matrix with the minimum weighted squared distance. Through multiple iterations, until the convergence condition is met, the fine registration is completed; the convergence condition includes reaching the required number of iterations or the difference between the rigid body transformation matrices of two adjacent iterations being less than a preset threshold.

8. The subway tunnel spatial safety detection method based on 3D point cloud as described in claim 1, characterized in that, Entry detection in subway tunnels based on target detection point clouds includes: The target detection point cloud is aligned with the tunnel standard clearance three-dimensional model constructed based on the subway tunnel design parameters in a spatial coordinate system. The tunnel standard clearance three-dimensional model represents the clearance envelope surface that allows trains to pass safely. For each measured point in the target detection point cloud, the nearest neighbor point is found on the surface of the tunnel standard clearance 3D model based on the nearest neighbor search algorithm, and the spatial position relationship of the measured point relative to the tunnel standard clearance 3D surface is calculated; if the coordinates of the measured point are determined to be inside the net envelope surface of the tunnel standard clearance 3D model according to the spatial position relationship, the measured point is determined to be an encroaching point, otherwise it is a non-encroaching point. Output the intrusion limit determination status of all measured points, mark the intrusion limit points in the target detection point cloud and output the display, and use cluster analysis to aggregate adjacent intrusion limit points into intrusion limit regions, calculate the center coordinates, range and maximum intrusion limit value of the intrusion limit regions, and generate an intrusion limit detection report.

9. The subway tunnel spatial safety detection method based on 3D point cloud as described in claim 1, characterized in that, Foreign object detection in subway tunnels based on target detection point clouds includes: After normalizing the target detection point cloud and the original tunnel point cloud without foreign objects, the data is input into a pre-trained point cloud registration neural network to extract the topological information of the geometric structure of the target detection point cloud and the original tunnel point cloud, thus obtaining structural features. By learning the three-dimensional coordinate data of the original tunnel point cloud, the positional distribution features of the original tunnel point cloud in the global space are extracted. By fusing structural features and location distribution features, the optimal rigid body transformation matrix of the target detection point cloud relative to the original tunnel point cloud is predicted by regression, and point cloud registration is performed based on the optimal rigid body transformation matrix. The nearest neighbor residual distance from each point in the registered point cloud to the surface of the original tunnel point cloud is compared with the error threshold; based on the comparison results, it is determined whether it is a foreign object point. All foreign object points are clustered and extracted to form a set of foreign object targets, generating a detection report output and / or a graphical output showing the location and / or geometric outline of the foreign object points.

10. A subway tunnel spatial safety detection device based on 3D point clouds, characterized in that, The method for spatial safety detection of subway tunnels based on 3D point clouds according to any one of claims 1-9 includes: The mobile chassis and the laser device mounted on the mobile chassis move along a preset direction within the subway tunnel space. The laser device collects three-dimensional point cloud data of the inner wall of the subway tunnel through a spiral scanning method. The scanned detection point cloud is homogenized and preprocessed to form a target detection point cloud. Based on the target detection point cloud, deformation, encroachment, and foreign object detection in the subway tunnel space are performed. The laser device includes three lidars uniformly arranged in the same transverse section of the mobile chassis along the circumference, or at least two lidars arranged at equal intervals in the longitudinal direction along the axis of the mobile chassis.

Images