Urban drainage pipe network internal environment high-precision three-dimensional reconstruction method and system

Abstract

This invention relates to a high-precision 3D reconstruction method and system for the internal environment of urban drainage pipe networks, comprising the following steps: S2: acquiring a task parameter file and establishing a computable sensor relationship for the multiple sensors of a patrol robot to obtain a calibration parameter set; S3: the patrol robot patrols and synchronously collects multi-sensor data at a unified timestamp to obtain raw data packets; S4: preprocessing the raw data packets to obtain a cleaned synchronous dataset; S5: constructing state variables and recursively estimating them to obtain a continuous trajectory; S6: based on the continuous trajectory, performing point cloud distortion removal and local mapping to obtain several local sub-maps and a global initial point cloud; S7: performing loop closure detection and global consistency optimization to obtain a globally optimized trajectory and a globally consistent point cloud; S8: reconstructing the pipeline centerline and cross-section sequence, and obtaining a high-precision 3D model of the pipe network through meshing and local refinement. This invention effectively improves the modeling accuracy and reliability of the internal environment of urban drainage pipe networks.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional reconstruction of pipelines, and in particular to a high-precision three-dimensional reconstruction method and system based on the internal environment of urban drainage pipe networks. Background Technology

[0002] Urban drainage networks bear critical functions such as rainwater and sewage transport and flood control. These pipelines operate in a constantly changing environment of moisture, corrosion, and sediment erosion, making them prone to problems such as cracks and leaks, misaligned joints, structural deformation, siltation and blockage, and foreign object intrusion. To ensure the safe operation of the drainage system, maintenance units need to accurately understand the internal spatial morphology and location of defects within the pipelines, and to conduct regular inspections, hazard identification, and repair decisions. This places higher demands on the accuracy and measurability of three-dimensional digital reconstruction of the pipeline's internal environment.

[0003] Current pipeline inspection methods primarily rely on CCTV / periscopes, which provide intuitive video, but the results depend heavily on manual interpretation and lack a stable three-dimensional scale benchmark. This makes it difficult to quantify deformation, misalignment height, and siltation volume with high precision. Furthermore, two-dimensional images are prone to missed detections and misjudgments under conditions of limited viewing angle, insufficient lighting, and specular reflection. Single-sensor-based 3D reconstruction schemes (such as those relying solely on visual SfM / SLAM or laser scanning) also face significant challenges within pipeline networks. Pipelines often exhibit characteristics such as long, narrow lengths, repetitive or poor textures, abrupt changes in bends and manhole structures, and water reflection and mist scattering. These features lead to feature matching degradation, pose drift accumulation, point cloud distortion, and splicing misalignment, ultimately resulting in inaccurate scale, local fragmentation, or global inconsistencies in the model, making it difficult to meet the needs of project acceptance and data storage management. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks, effectively improving the modeling accuracy and reliability of the internal environment of urban drainage pipe networks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks includes the following steps: S1: Obtain the task parameter file, establish a computable sensor relationship for the patrol robot's multiple sensors, and obtain the calibration parameter set; S2: The patrol robot patrols based on the task parameter file and calibration parameter set and synchronously collects multi-sensor data according to a unified timestamp to obtain raw data packets; S3: Preprocess the original data packets to obtain the cleaned synchronization dataset; S4: Based on the cleaned synchronous dataset, construct the state variables and recursively estimate them to obtain the continuous trajectory; S5: Based on the continuous trajectory, point cloud distortion removal and local mapping are performed to obtain several local sub-maps and the global initial point cloud; S6: Perform loop closure detection and global consistency optimization on the obtained local sub-graphs and the global initial point cloud to obtain the global optimized trajectory and the global consistent point cloud; S7: Based on the local optimization trajectory and the global consistent point, the pipeline centerline and cross-section sequence are reconstructed, and a high-precision three-dimensional model of the pipeline network is obtained through meshing and local refinement.

[0006] Furthermore, the task parameter file is obtained, and a computable sensor relationship is established for the patrol robot's multiple sensors to obtain the calibration parameter set, as detailed below: Obtain and parse the task parameter file to clarify the boundary conditions and data specifications for pipeline inspection and 3D reconstruction. The task parameter file includes: the starting and ending well numbers and estimated mileage of the target pipe section, the pipe type and cross-sectional form and pipe diameter range, the allowable travel speed and stopping strategy, the sensor sampling frequency and triggering method, the lighting brightness and exposure strategy, the data storage format and naming rules, the target accuracy index, and the coordinate system and output requirements. Establish computable relationships among multiple sensors. After the task parameters are determined, establish computable sensor relationships among the multiple sensors carried by the patrol robot, so that the observations of each sensor can be fused and calculated under a unified coordinate chain and a unified time reference. After the computable sensor relationship framework is determined, calibration and quality verification are performed, and a calibration parameter set is output. The calibration content typically includes: camera intrinsic parameters and distortion parameters, extrinsic parameters between multiple cameras, LiDAR-camera extrinsic parameters, LiDAR and IMU extrinsic parameters and time offset, odometry scale and slip model parameters. After calibration, consistency verification is required. If the threshold in the task parameter file is not met, the process is to adjust the installation rigidity, synchronize the configuration, or reacquire calibration data. Finally, the verification results are solidified into a calibration parameter set and stored in a versioned manner.

[0007] Furthermore, a multi-sensor computable relationship is established, as follows: Define the robot base coordinate system B, the engineering coordinate system W, the LiDAR coordinate system L, and the coordinate systems of each camera C. i IMU coordinate system I, odometer coordinate system O; a unified time axis is achieved through network time synchronization, mapping the device timestamps of each observation to a unified time, and aligning them according to the unified time during fusion; Then, a space extrinsic parameter link was established: solving the extrinsic parameters of the LiDAR base. Camera base external parameters IMU base external parameters This allows any sensor observation to be mapped to the base coordinate system via fixed extrinsic parameters. LiDAR base external parameters The scene is calibrated using a planar plate. An error function is established between the geometric elements in the LiDAR point cloud and the geometric model of the calibration object in the base system. The plane is represented in the B-frame as π=(n,d). Then, the LiDAR point p L Transform to B series: ; in, Let be the coordinates of the point in the robot base coordinate system; These are the original coordinates of the point in the LiDAR coordinate system; For each camera, the three-dimensional coordinates P of the calibration plate corner points are used. B Projected onto the image plane, camera model: ; Where u is the pixel coordinate, representing the projected position of the 3D point in the image; π(.) is the camera projection function, projecting the 3D point onto the 2D image plane; K i D is the intrinsic parameter matrix of the i-th camera; i Let be the distortion parameter vector of the i-th camera, containing radial and tangential distortion coefficients; Joint calibration using motion data: The IMU provides an angular velocity sequence, and the LiDAR provides relative motion constraints, ensuring that the rotational increments of both are consistent under the same motion. ; in, This is the rotation matrix from the LiDAR coordinate system to the base coordinate system; Let be the relative rotation matrix of the LiDAR coordinate system within the time interval [t, t+Δt]. This is the relative rotation matrix of the base coordinate system within the same time interval; When scanning a single frame of LiDAR data, let the reference time be the scan start time t0, and the unified time for the j-th point during the scan be t. j The pose TWB(t) estimated from the continuous trajectory is used to transform the point to the world pose corresponding to the reference time, thereby removing motion distortion. ; Where, p W,j Let j be the coordinates of the j-th point in the world coordinate system; For time t j The pose transformation matrix of the base relative to the world coordinate system; p L,j Let t be the original coordinates of the j-th point in the LiDAR coordinate system; j This is the timestamp for the data collection at point j; At the same time, based on the mission's requirements for scale stability, the odometer scale coefficient was incorporated. With roller skating correction parameters This allows the mileage constraint to be used directly in subsequent state estimation.

[0008] Furthermore, the patrol robot conducts patrols based on the task parameter file and calibration parameter set, and synchronously collects multi-sensor data at a unified timestamp to obtain raw data packets, as follows: The patrol robot reads the mission parameter file and calibration parameter set Θ to complete the data acquisition configuration and execution constraint loading for this patrol. This includes writing the patrol path information from the mission parameters into the motion control module; writing the acquisition parameters into the sensor driver; and loading the coordinate chain and time mapping model so that the system can attach a unified timestamp and necessary metadata to each observation during the acquisition phase. The robot enters the pipeline for patrol according to the speed curve and attitude stabilization strategy specified by the task parameters, and all sensor data are marked and collected synchronously with a unified time axis t. The raw observations from multiple sensors throughout the patrol process are encapsulated using a unified timestamp to form a structured raw data packet D. raw At the same time, complete metadata is saved to ensure traceability and reproducibility.

[0009] Further preprocessing is performed as follows: The original data packets output by S2 are subjected to time base unification and integrity verification. Based on the time mapping model given by correction S1, the timestamps of each sensor device, including LiDAR, camera, IMU, and odometer, are converted into a unified timestamp t, and the multi-stream timeline is reconstructed by sorting them according to the unified time. Subsequently, data integrity checks are performed: packet loss rate, timestamp jump, sampling frequency drift, camera triggering asynchrony, and IMU saturation segment issues are statistically analyzed for each data stream. Based on time alignment, various observations are cleaned and normalized. For IMU data, zero-bias initial value estimation, peak removal, and low-pass filtering are performed, and abnormal saturated samples are removed or invalidated. For odometry data, zero-point reset, pulse anomaly detection, and short-time slip segment identification are performed, and a usable incremental odometer sequence is output. For camera images, blur detection, overexposure and underexposure detection, and strong glare ratio estimation are performed, severely unusable frames are removed, and distortion correction and brightness normalization are performed to improve the stability of subsequent features. For LiDAR point clouds, distance threshold clipping, outlier removal, and intensity anomaly suppression are performed, and downsampling is performed as needed to balance accuracy and computational load. The cleaned synchronization dataset is obtained.

[0010] Furthermore, based on the cleaned synchronous dataset, state variables are constructed and recursively estimated to obtain continuous trajectories, as follows: Based on the cleaned synchronized dataset D, with key moments t on a unified time axis kThe goal of establishing a recursive estimation framework is to obtain the continuous pose trajectory of the robot base in the engineering frame. At time t k Define the state vector: ; Among them, R WB,k For posture, p WB,k For position, v WB,k For speed, b g,k ,b a,k These are the gyroscope and the zero bias accelerator, respectively; k s λ is the odometer scaling factor. k Slip ratio; IMU measurement model: ; Where, ω m ω is the gyroscope measurement; b is the true angular velocity; g For zero bias of the gyroscope; n g To measure noise for a gyroscope; a m For acceleration, a is the actual acceleration; b is the acceleration. a For accelerometer zero bias; n a To measure noise for accelerometers; The kinematic equations for continuous time, expressed in W-frame, with gravity g W Given: ; in, R is the time derivative of the rotation matrix of the base relative to the world coordinate system; WB This is the rotation matrix of the base relative to the world coordinate system; Indicates an antisymmetric matrix operator; g is the time derivative of the velocity of the base in the world coordinate system; W v is the gravity vector in the world coordinate system. WB The velocity of the base in the world coordinate system; p represents the position-time derivative of the base in the world coordinate system. WB This refers to the position of the base in the world coordinate system. Zero bias employs random walk: ; in, n is the time rate of change of the gyroscope's zero bias; bg The noise is zero-bias random walk noise of the gyroscope, which follows a Gaussian distribution; n is the time rate of change of the accelerometer's zero bias. ba This refers to the accelerometer zero-bias random walk noise. During recursion, IMU high-frequency data is used in [t]k ,t k+1 Integrating upwards forms a prediction of the state, resulting in a high-frequency skeleton of the continuous trajectory; To effectively incorporate the high-frequency IMU into the keyframe recursion, in each segment [t] k ,t k+1 Pre-integrating the IMU yields the relative motion quantities (ΔR, Δv, Δp): ; Where, ΔR k For the time interval [t] k ,t k+1 The relative rotation matrix within ]; It is the cumulative product from time k to k+1; The exponential mapping of rotation matrices transforms antisymmetric matrices into rotation matrices; ω m,j b is the gyroscope measurement at time j; g,k Δt is the gyroscope zero-bias estimate at time k; j Let Δv be the j-th sampling interval; k ΔR represents the change in relative velocity over a time interval. k→j Let a be the relative rotation matrix from time k to time j; m,j b is the accelerometer measurement at time j; a,k Δp is the accelerometer zero bias estimate at time k. k Δv represents the change in relative position over a time interval. k→j This represents the change in relative velocity from time k to time j; The state prediction from k→k+1 can be expressed in discrete form using pre-integration: ; in, R is the prediction rotation matrix at time k+1; WB,k Let be the estimated rotation matrix at time k; The predicted velocity at time k+1; v WB,k The estimated velocity at time k; p is the predicted position at time k+1; WB,k The estimated position at time k; Simultaneously, the covariance is recursively calculated: ; Among them, F k G k To linearize the Jacobian, Q k The covariance of IMU noise and zero-bias random walk noise; Let be the prediction covariance matrix at time k+1; At each critical moment tk The predicted state is updated using the cleaned sensor observations to obtain a continuous trajectory.

[0011] Furthermore, at each critical moment t k The predicted state is updated using the cleaned sensor observations to obtain a continuous trajectory, as follows: Register the current scan before distortion correction with the previous local map to obtain the relative pose observation. residual Written in logarithmic form as a Lie group: ; in, Let be the transformation matrix from the world coordinate system to the base coordinate system at time k; Let be the transformation matrix from the world coordinate system to the base coordinate system at time k+1; And minimize residuals Used for EKF updates; The odometer gives the increment Δs meas,k Establish a measurement model: ; Among them, z O,k For the odometer in the time period [t] k ,t k+1 The measured observation value of h(x) k ) is the prediction function for odometer observations; k s λ is the odometer scale factor; k p is the slip ratio parameter. WB,k+1 Let p be the position of the base in the world coordinate system at time k+1; WB,k Let k be the position of the base in the world coordinate system at time k. It is the Euclidean norm; Corresponding residuals: ; Where, r O For odometer observation residuals; Reduce the weight of this factor when slippage risk is detected; Relative pose is obtained using keyframe feature tracking. Constraints are formed by mapping extrinsic parameters to the base: ; in, This refers to a relative transformation from the perspective of the base coordinate system; Keyframe estimation is obtained Then, high-frequency attitude is obtained by IMU integration between two keyframes, thereby outputting a continuous trajectory. .

[0012] Furthermore, based on the continuous trajectory, point cloud distortion correction and local mapping are performed to obtain several local sub-images and the global initial point cloud, as detailed below: The original continuous trajectory scan output in step S4 is subjected to distortion correction processing to eliminate the geometric distortion of the point cloud caused by robot motion. For each frame of LiDAR scan, the acquisition time of the j-th laser point is t. j The original point coordinates are p L,j The extrinsic parameters established through S1 Continuous trajectory interpolation transforms each point from the LiDAR coordinate system at its acquisition time to the world coordinate system corresponding to a unified reference time. The distortion correction formula is: ; That is, first transform the points to the world frame, and then unify them to the coordinate system of the reference time. For time periods when the trajectory quality is below the threshold, a distortion removal strategy is adopted. After distortion correction, the continuous point cloud sequence is divided into several local sub-images, each covering a certain spatial range. Fine point cloud registration and geometric optimization are performed within each sub-image. For the distorted point cloud sequence within each sub-image, frame-by-frame registration is used for local optimization: the first frame of the sub-image is used as the reference coordinate system, and the subsequent frames are transformed through point cloud registration. All point clouds are accumulated into the sub-image coordinate system to form a local dense map. After the local subgraphs are constructed, all subgraphs are stitched together according to their positions in the global coordinate system to form a global initial point cloud covering the entire inspection section.

[0013] Subgraph stitching utilizes the global pose information provided by the S4 continuous trajectory: the reference coordinate system of each subgraph is linked to the coordinate system at the corresponding time step. Transform to the world coordinate system, that is: ; This unifies all local subgraphs into the same global coordinate frame, resulting in the final output of a global initial point cloud. It covers the entire pipeline inspection route.

[0014] Furthermore, loop closure detection and global consistency optimization are performed on several obtained local sub-graphs and the initial global point cloud to obtain the globally optimized trajectory and the globally consistent point cloud, as detailed below: Loop closure detection is performed on the local subgraph set output by S5 and the global initial point cloud to identify the spatial locations that the robot repeatedly passes through during the inspection process, providing closure constraints for subsequent global consistency optimization; For candidate loops that pass geometric verification, precise inter-subgraph registration is performed to obtain the relative transformation of the candidate loops. and its covariance estimation To construct a reliable set of closed constraints for global optimization; Based on the detected loop closure constraints, a global pose graph is constructed and optimized to eliminate the cumulative drift in the S4 continuous trajectory and obtain a globally consistent optimized trajectory. The pose map uses keyframe poses. For each node, there are three types of edge constraints: odometry constraints between adjacent keyframes, loop closure constraints, and absolute position constraints; the global optimization objective function combines the Mahalanobis distance of all constraints. ; Where, r ij For the residual, Σ ij To address the covariance, a manifold optimization method is used to solve the problem on SE(3) to obtain a globally consistent set of keyframe poses. The optimized keyframe poses are then used to reconstruct continuous trajectories through spline interpolation or IMU reintegration to form a globally optimized trajectory. ; Utilizing global trajectory optimization The original LiDAR scans are distorted and globally registered to generate a geometrically globally consistent final point cloud model. .

[0015] Furthermore, based on the local optimized trajectory and the globally consistent point, the pipeline centerline and cross-section sequence are reconstructed, and through meshing and local refinement, a high-precision 3D model of the pipeline network is obtained, as follows: Based on global optimization trajectory Consistent with global points Pipeline centerline extraction and cross-section sequence reconstruction are performed to convert discrete point clouds into parameterized pipeline geometric models. Based on the reconstructed centerline and cross-section sequence, a three-dimensional mesh model of the pipeline is constructed to form a continuous surface representation with correct topological relationships. The meshing adopts a cross-section-based parametric method: mesh nodes are set along the centerline at uniform arc length intervals Δs, and sampling points in the circumferential direction are generated at each node position according to the cross-section parameters to form a regular mesh topology in cylindrical coordinates; corresponding points of adjacent cross-sections are connected by the Delaunay triangulation meshing method to generate a triangular mesh surface. Based on the meshing, local refinement and accuracy optimization are performed. Local refinement adopts a mesh optimization method with point cloud constraints: the mesh vertices are used as adjustable parameters, and the original point cloud is used as geometric constraints to construct an objective function that minimizes the distance from a point to the mesh surface. ; Where, p i M represents a point in a point cloud. jThe mesh is composed of surface patches; the positions of the mesh vertices are adjusted through iterative optimization to ensure that the mesh surface best fits the geometry of the original point cloud, while maintaining the topological continuity and smoothness of the mesh; the final high-precision 3D model M of the pipeline network is generated. final .

[0016] A high-precision 3D reconstruction system based on the internal environment of urban drainage pipe networks includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of a high-precision 3D reconstruction method based on the internal environment of urban drainage pipe networks as described above. The present invention has the following beneficial effects: 1. This invention establishes a multi-sensor computable spatiotemporal relationship, enabling the collected data to be directly fused under the same time reference, avoiding systematic errors caused by asynchrony, and partially adopting recursive state estimation, so that stable and continuous trajectories can still be obtained in narrow, low-texture, and repetitive network scenarios. 2. This invention incorporates local subgraphs and global initial point clouds into a unified graph optimization framework through loop closure detection and global consistency optimization, which significantly suppresses long-distance cumulative drift and ensures geometric self-consistency between the global point cloud and the trajectory. 3. Based on globally optimized trajectories and globally consistent point clouds, this invention reconstructs centerlines and cross-sectional sequences, and then combines meshing and local refinement to output high-precision 3D models, effectively improving the modeling accuracy in pipeline network operation and maintenance applications such as inspection and verification, defect location, siltation volume estimation, and deformation assessment, and improving detection reliability. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1 In this embodiment, a high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks is provided, which includes the following steps: S1: Obtain the task parameter file, establish a computable sensor relationship for the patrol robot's multiple sensors, and obtain the calibration parameter set; S2: The patrol robot patrols based on the task parameter file and calibration parameter set and synchronously collects multi-sensor data according to a unified timestamp to obtain raw data packets; S3: Preprocess the original data packets to obtain the cleaned synchronization dataset; S4: Based on the cleaned synchronous dataset, construct the state variables and recursively estimate them to obtain the continuous trajectory; S5: Based on the continuous trajectory, point cloud distortion removal and local mapping are performed to obtain several local sub-maps and the global initial point cloud; S6: Perform loop closure detection and global consistency optimization on the obtained local sub-graphs and the global initial point cloud to obtain the global optimized trajectory and the global consistent point cloud; S7: Based on the local optimization trajectory and the global consistent point, the pipeline centerline and cross-section sequence are reconstructed, and a high-precision three-dimensional model of the pipeline network is obtained through meshing and local refinement.

[0019] In this embodiment, a task parameter file is obtained, and a computable sensor relationship is established among the multiple sensors of the patrol robot to obtain a calibration parameter set, as detailed below: Obtain and parse the task parameter file to clarify the boundary conditions and data specifications for pipeline inspection and 3D reconstruction. The task parameter file includes: the starting and ending well numbers and estimated mileage of the target pipe section, the pipe type and cross-sectional form (circular pipe / box culvert / ellipse, etc.) and pipe diameter range, the allowable travel speed and stopping strategy, the sensor sampling frequency and triggering method, the lighting brightness and exposure strategy, the data storage format and naming rules, the target accuracy indicators (relative accuracy / absolute accuracy / coverage threshold), and the coordinate system and output requirements (whether it needs to be aligned with GIS coordinates). Establish computable relationships among multiple sensors (time synchronization + spatial extrinsic parameters + coordinate chain). After the task parameters are determined, establish computable sensor relationships among the multiple sensors carried by the patrol robot, so that the observations of each sensor can be fused and calculated under a unified coordinate chain and a unified time reference. After the computable sensor relationship framework is determined, calibration and quality verification are performed, and a calibration parameter set is output. The calibration content typically includes: camera intrinsic parameters and distortion parameters, extrinsic parameters between multiple cameras, LiDAR-camera extrinsic parameters, LiDAR and IMU extrinsic parameters and time offset, odometry scale and slip model parameters. After calibration, consistency verification is required, including camera reprojection error threshold check, residual check from LiDAR point to calibration plane / cylinder, short-distance round-trip mileage closure error check, and time synchronization jitter statistics. If the thresholds in the task parameter file are not met, the process is to return to adjust the installation rigidity, synchronize the configuration, or reacquire calibration data. Finally, the verification results are solidified into a calibration parameter set and stored in a versioned manner.

[0020] In this embodiment, a computable relationship between multiple sensors is established, as detailed below: Define the robot base coordinate system B, the engineering coordinate system W, the LiDAR coordinate system L, and the coordinate systems of each camera C. i IMU coordinate system I, odometer coordinate system O; a unified time axis is achieved through network time synchronization, mapping the device timestamps of each observation to a unified time, and aligning them according to the unified time during fusion; Then, a space extrinsic parameter link was established: solving the extrinsic parameters of the LiDAR base. Camera base external parameters IMU base external parameters This allows any sensor observation to be mapped to the base coordinate system via fixed extrinsic parameters. LiDAR base external parameters The scene is calibrated using a planar plate. An error function is established between the geometric elements in the LiDAR point cloud and the geometric model of the calibration object in the base system. The plane is represented in the B-frame as π=(n,d). Then, the LiDAR point p L Transform to B series: ; in, Let be the coordinates of the point in the robot base coordinate system; These are the original coordinates of the point in the LiDAR coordinate system; For each camera, the three-dimensional coordinates P of the calibration plate corner points are used. B Projected onto the image plane, camera model: ; Where u is the pixel coordinate, representing the projected position of the 3D point in the image; π(.) is the camera projection function, projecting the 3D point onto the 2D image plane; K i D is the intrinsic parameter matrix of the i-th camera; i Let be the distortion parameter vector of the i-th camera, containing radial and tangential distortion coefficients; Joint calibration using motion data: The IMU provides an angular velocity sequence, and the LiDAR provides relative motion constraints, ensuring that the rotational increments of both are consistent under the same motion. ; in, This is the rotation matrix from the LiDAR coordinate system to the base coordinate system; Let be the relative rotation matrix of the LiDAR coordinate system within the time interval [t, t+Δt]. This is the relative rotation matrix of the base coordinate system within the same time interval; When performing a LiDAR scan of a frame (or a loop), let the reference time be the scan start time t0, and the unified time for the j-th point during the scan be t. j The pose TWB(t) estimated from the continuous trajectory is used to transform the point to the world pose corresponding to the reference time, thereby removing motion distortion. ; Where, p W,j Let j be the coordinates of the j-th point in the world coordinate system; For time t jThe pose transformation matrix of the base relative to the world coordinate system; p L,j Let t be the original coordinates of the j-th point in the LiDAR coordinate system; j This is the timestamp for the data collection at point j; At the same time, based on the mission's requirements for scale stability, the odometer scale coefficient was incorporated. With roller skating correction parameters This allows the mileage constraint to be used directly in subsequent state estimation.

[0021] In this embodiment, the patrol robot patrols based on the task parameter file and calibration parameter set, and synchronously collects multi-sensor data at a unified timestamp to obtain the raw data packet, as detailed below: The patrol robot reads the mission parameter file and calibration parameter set Θ to complete the data acquisition configuration and execution constraint loading for this patrol. This includes writing the patrol path information (start and end well numbers, estimated mileage, bifurcation / turn / wellhead event points, allowed U-turn or round-trip strategies) from the mission parameters into the motion control module; writing the acquisition parameters (LiDAR rotation speed / frame rate, camera frame rate and trigger mode, IMU sampling rate, odometer output frequency, illumination brightness and exposure / gain range) into the sensor driver; and loading the coordinate chain and time mapping model (αs, βs or Δts for each sensor, and LiDAR base extrinsic parameters). Camera base external parameters IMU base external parameters (External parameters), enabling the system to attach a unified timestamp and necessary metadata to each observation during the data acquisition phase; The robot enters the pipeline for patrol according to the speed curve and attitude stabilization strategy specified by the task parameters, and all sensor data are marked and collected synchronously with a unified time axis t. The raw observations from multiple sensors throughout the patrol process are encapsulated using a unified timestamp to form a structured raw data packet D. raw Simultaneously, complete metadata is preserved to ensure traceability and reproducibility. The original data package includes: LiDAR raw point stream (including (x,y,z), intensity, point-level / frame-level timestamps), raw image frames from each camera and their exposure parameters, raw IMU measurements (angular velocity / acceleration and timestamps), odometry increments / wheel speed data, and patrol event sequences (timestamps and mileage positions for wellhead / turn / fork / stop, etc.); metadata includes the mission parameter file version number, calibration parameter set Θ version number, synchronization status statistics (PTP deviation / triggered frame loss rate), and acquisition process logs (packet loss rate, abnormal frame count, temperature, and power status).

[0022] In this embodiment, the preprocessing is as follows: The original data packets Draw={PL(ts),Ii(ts),U(ts),O(ts),events} output by S2 are subjected to time base unification and integrity verification. Based on the time mapping model given by correction S1, the timestamps of each sensor device, including LiDAR, camera, IMU, and odometer, are converted into a unified timestamp t, and the multi-stream timeline is reconstructed by sorting them according to the unified time. Subsequently, data integrity checks are performed: packet loss rate, timestamp jump, sampling frequency drift, camera triggering asynchrony (multi-view time difference exceeding threshold), and IMU saturation segment (acceleration / angular velocity exceeding range) issues are statistically analyzed for each data stream. Based on time alignment, various observations are cleaned and normalized. For IMU data, zero-bias initial value estimation (stationary segment estimation or temperature drift model initialization), peak removal and low-pass filtering are performed, and abnormal saturation samples are removed or invalidated. For odometry data, zero-point reset, pulse anomaly detection and short-term slip segment identification (e.g., mileage jumps or inconsistencies with IMU / velocity) are performed, and a usable incremental mileage sequence is output. For camera images, blur detection (e.g., Laplacian variance), overexposure and underexposure detection, strong glare ratio estimation are performed, severely unusable frames are removed, and distortion correction and brightness normalization can be performed to improve the stability of subsequent features. For LiDAR point clouds, distance threshold clipping, outlier removal (statistical filtering / radius filtering), intensity anomaly suppression are performed, and downsampling (voxel filtering) is performed as needed to balance accuracy and computational load. The cleaned synchronization dataset is obtained.

[0023] In this embodiment, based on the cleaned synchronization dataset, state variables are constructed and recursively estimated to obtain continuous trajectories, as follows: Based on the cleaned synchronized dataset D, with key moments t on a unified time axis k The goal of establishing a recursive estimation framework is to obtain the continuous pose trajectory of the robot base in the engineering frame. At time t k Define the state vector: ; Among them, R WB,k For posture, p WB,k For position, v WB,k For speed, b g,k ,b a,k These are the gyroscope and the zero bias accelerator, respectively; k s λ is the odometer scaling factor. k Slip ratio; IMU measurement model: ; Where, ω m ω is the gyroscope measurement; b is the true angular velocity;g For zero bias of the gyroscope; n g To measure noise for a gyroscope; a m For acceleration, a is the actual acceleration; b is the acceleration. a For accelerometer zero bias; n a To measure noise for accelerometers; The kinematic equations for continuous time, expressed in W-frame, with gravity g W Given: ; in, R is the time derivative of the rotation matrix of the base relative to the world coordinate system; WB This is the rotation matrix of the base relative to the world coordinate system; The antisymmetric matrix operator transforms a 3×1 vector into a 3×3 antisymmetric matrix. g is the time derivative of the velocity of the base in the world coordinate system; W v is the gravity vector in the world coordinate system. WB The velocity of the base in the world coordinate system; p represents the position-time derivative of the base in the world coordinate system. WB This refers to the position of the base in the world coordinate system. Zero bias employs random walk: ; in, n is the time rate of change of the gyroscope's zero bias; bg The noise is zero-bias random walk noise of the gyroscope, which follows a Gaussian distribution; n is the time rate of change of the accelerometer's zero bias. ba This refers to the accelerometer zero-bias random walk noise. During recursion, IMU high-frequency data is used in [t] k ,t k+1 Integrating upwards forms a prediction of the state, resulting in a high-frequency skeleton of the continuous trajectory; To effectively incorporate the high-frequency IMU into the keyframe recursion, in each segment [t] k ,t k+1 Pre-integrating the IMU yields the relative motion quantities (ΔR, Δv, Δp): ; Where, ΔR k For the time interval [t] k ,t k+1 The relative rotation matrix within ]; It is the cumulative product from time k to k+1; The exponential mapping of rotation matrices transforms antisymmetric matrices into rotation matrices; ω m,jb is the gyroscope measurement at time j; g,k Δt is the gyroscope zero-bias estimate at time k; j Let Δv be the j-th sampling interval; k ΔR represents the change in relative velocity over a time interval. k→j Let a be the relative rotation matrix from time k to time j; m,j b is the accelerometer measurement at time j; a,k Δp is the accelerometer zero bias estimate at time k. k Δv represents the change in relative position over a time interval. k→j This represents the change in relative velocity from time k to time j; The state prediction from k→k+1 can be expressed in discrete form using pre-integration: ; in, R is the prediction rotation matrix at time k+1; WB,k Let be the estimated rotation matrix at time k; The predicted velocity at time k+1; v WB,k The estimated velocity at time k; p is the predicted position at time k+1; WB,k The estimated position at time k; Simultaneously, recursively calculate the covariance (error state, EKF form): ; Among them, F k G k To linearize the Jacobian, Q k The covariance of IMU noise and zero-bias random walk noise; Let be the prediction covariance matrix at time k+1; At each critical moment t k The predicted state is updated using the cleaned sensor observations to obtain a continuous trajectory.

[0024] In this embodiment, at each critical moment t k The predicted state is updated using the cleaned sensor observations to obtain a continuous trajectory, as follows: Register the current undistorted scan (or feature points) with the previous local map (point-to-plane / point-to-line / GICP, etc.) to obtain the relative pose observation. residual Written in logarithmic form as a Lie group: ; in, Let be the transformation matrix from the world coordinate system to the base coordinate system at time k; Let be the transformation matrix from the world coordinate system to the base coordinate system at time k+1; And minimize residuals Used for EKF updates; The odometer gives the increment Δs meas,k Establish a measurement model: ; Among them, z O,k For the odometer in the time period [t] k ,t k+1 The measured observation value of h(x) k ) is the prediction function for odometer observations; k s λ is the odometer scale factor; k p is the slip ratio parameter. WB,k+1 Let p be the position of the base in the world coordinate system at time k+1; WB,k Let k be the position of the base in the world coordinate system at time k. It is the Euclidean norm; Corresponding residuals: ; Where, r O For odometer observation residuals; When slip risk (λ) is detected k Reduce the weight of this factor when it is a high-quality or high-grade segment. Relative pose is obtained using keyframe feature tracking. Constraints are formed by mapping extrinsic parameters to the base: ; in, This refers to a relative transformation from the perspective of the base coordinate system; Keyframe estimation is obtained Then, high-frequency attitude is obtained by IMU integration between two keyframes, thereby outputting a continuous trajectory. It also outputs the trajectory confidence (covariance or residual statistics) at each time step, which can be directly used by S5 point cloud distortion removal and local mapping.

[0025] In this embodiment, point cloud distortion correction and local mapping are performed based on continuous trajectories to obtain several local sub-images and a global initial point cloud, as detailed below: The original continuous trajectory scan output in step S4 is subjected to distortion correction processing to eliminate the geometric distortion of the point cloud caused by robot motion. For each LiDAR scan frame (or each scan cycle), the acquisition time of the j-th laser point is t. j The original point coordinates are p L,j The extrinsic parameters established through S1 Continuous trajectory interpolation transforms each point from the LiDAR coordinate system at its acquisition time to the world coordinate system corresponding to a unified reference time. The distortion correction formula is: ; First, the points are transformed to the world frame, and then unified to the coordinate system of the reference time. For time periods where the trajectory quality is below the threshold (identified by Σ(t) or quality_flags), a distortion correction strategy is adopted: shortening the single scan time window, reducing the motion compensation amplitude, or directly skipping severely unreliable segments. The point cloud after distortion correction maintains geometric consistency, eliminating artifacts such as "tailing" and "bending" caused by robot turning, acceleration, deceleration, and other movements, providing a reliable geometric basis for subsequent accurate mapping. After distortion correction, the continuous point cloud sequence is segmented into several local sub-images, each covering a certain spatial range (e.g., a 10-20 meter pipe segment). Fine-grained point cloud registration and geometric optimization are performed within each sub-image. The sub-image segmentation strategy combines spatial distance, time interval, and scene changes: straight pipe segments are segmented by mileage intervals, while key areas such as bends, branches, and manholes are segmented by semantic events, ensuring that the scene within each sub-image is relatively stable and possesses sufficient geometric features for registration. For the distortion-corrected point cloud sequence within each sub-image, frame-by-frame registration (e.g., ICP, GICP, feature matching) is used for local optimization: using the first frame of the sub-image as the reference coordinate system, subsequent frames undergo relative transformations through point cloud registration, and all point clouds are accumulated into the sub-image coordinate system to form a local dense map. After the local subgraphs are constructed, all subgraphs are stitched together according to their positions in the global coordinate system to form a global initial point cloud covering the entire inspection section.

[0026] Subgraph stitching utilizes the global pose information provided by the S4 continuous trajectory: the reference coordinate system of each subgraph is linked to the coordinate system at the corresponding time step. Transform to the world coordinate system, that is: ; This unifies all local sub-maps into the same global coordinate frame. Quality control and consistency checks are performed during the stitching process: registration errors in overlapping areas of adjacent sub-maps are detected, and sub-map boundaries with geometric discontinuities or significant deviations are identified; redundant points in overlapping areas are removed or weighted fusion is performed to avoid uneven point cloud density or repeated accumulation; the reliability weights of each sub-map are adjusted based on trajectory quality markers to reduce the contribution of sub-maps corresponding to unreliable time periods. The final output is the global initial point cloud. It covers the entire pipeline inspection route, with millimeter-level local geometric accuracy and decimeter-level global positioning accuracy, and also includes quality markers for each point (source submap, timestamp, credibility and other metadata).

[0027] In this embodiment, loop closure detection and global consistency optimization are performed on the obtained local sub-graphs and the global initial point cloud to obtain the globally optimized trajectory and the globally consistent point cloud, as detailed below: Loop closure detection is performed on the local sub-graph set output by S5 and the global initial point cloud to identify the spatial locations repeatedly traversed by the robot during the inspection process, providing closure constraints for subsequent global consistency optimization. A multi-level strategy is employed for loop closure detection: at the coarse level, spatial proximity filtering is performed using the positional information of the continuous trajectory from S4 to identify keyframe pairs on the trajectory that are far apart in time but close in spatial location (such as round-trip inspections, ring networks, branch pipe revisits, etc.). Spatial distance thresholds (typically 2-5 meters) and time interval thresholds (to avoid false loops between adjacent moments) are set for candidate pair filtering. At the fine level, geometric similarity verification is performed on the selected candidate keyframe pairs: geometric feature descriptors of the corresponding local sub-graphs are extracted (such as spin images based on normal vector distribution, feature points based on curvature, cylindrical surface parameters of pipe structures, etc.), and similarity scores are calculated through feature matching. Thresholds are set to eliminate false positive candidates with excessively large geometric differences. For candidate loop closures that pass geometric verification, perform precise inter-subgraph registration (ICP, GICP, or registration based on pipeline geometric constraints) to obtain the relative transformation of the candidate loop closures. and its covariance estimation To construct a reliable set of closed constraints for global optimization; Based on the detected loop closure constraints, a global pose graph is constructed and optimized to eliminate the cumulative drift in the S4 continuous trajectory and obtain a globally consistent optimized trajectory. The pose map uses keyframe poses. For nodes, there are three types of edge constraints: odometry constraints between adjacent keyframes (from the relative pose estimation and its covariance in S4), loop closure constraints (from loop closure detection), and edge constraints. and ), and absolute position constraints (such as known wellhead coordinates or GPS signals); the global optimization objective function combines the Mahalanobis distance of all constraints: ; Where, r ij For the residual, Σ ij To address the covariance, manifold optimization methods (such as Gauss-Newton and Levenberg-Marquardt based on Lie groups) are used to solve the problem on SE(3) to obtain a globally consistent set of keyframe poses. The optimized keyframe poses are then used to reconstruct continuous trajectories through spline interpolation or IMU reintegration methods to form a globally optimized trajectory. ; Utilizing global trajectory optimization The original LiDAR scans are distorted and globally registered to generate a geometrically globally consistent final point cloud model. Specifically, this includes: replacing the initial trajectory in S5 with an optimized continuous trajectory, re-executing the point cloud distortion correction process to ensure that the spatiotemporal transformation of each laser point is based on globally consistent pose information; reconstructing local sub-images, but at this time the relative positional relationships of each sub-image in the global coordinate system have been accurately corrected through pose graph optimization; and stitching all the re-distorted sub-images together according to the optimized global pose to form a globally consistent point cloud. .

[0028] In this embodiment, based on the local optimized trajectory and the globally consistent point, the pipeline centerline and cross-section sequence are reconstructed, and a high-precision 3D model of the pipeline network is obtained through meshing and local refinement, as follows: Based on global optimization trajectory Consistent with global points Pipeline centerline extraction and cross-section sequence reconstruction are performed to convert discrete point clouds into parameterized pipeline geometric models. Centerline extraction employs a multi-step strategy: A globally optimized trajectory is used as the initial centerline estimate. Cross-sectional planes are set at fixed intervals (e.g., 0.1-0.2 meters) along the trajectory direction. Local point clouds are extracted at each cross-sectional location, and circular or elliptical cross-sections are fitted. The center coordinates of the cross-sections are obtained through RANSAC circle fitting or least-squares ellipse fitting. Connecting all cross-sectional centers forms a refined pipe centerline. The centerline is smoothed (B-spline fitting or moving average) to eliminate local noise while preserving realistic geometric features such as turns and diameter changes. During cross-section sequence reconstruction, geometric parameters such as pipe radius, ellipticity, and normal vector are calculated for each cross-sectional location, forming a pipe cross-section parameter sequence {r(s), n(s), e(s)} along the centerline, where s is the arc length parameter along the centerline. This parameterized representation not only preserves the complete geometric information of the pipe but also provides a standardized geometric description framework for subsequent applications such as diameter change detection, deformation analysis, and defect location, achieving a structured transformation from "disordered point clouds" to "ordered geometric parameters." Based on the reconstructed centerline and cross-section sequence, a 3D mesh model of the pipeline is constructed to form a continuous surface representation with correct topological relationships. Meshing employs a cross-section-based parametric method: mesh nodes are set along the centerline at uniform arc length intervals Δs, and sampling points in the circumferential direction are generated at each node location based on cross-section parameters, forming a regular mesh topology in cylindrical coordinates. For complex geometric regions such as pipe diameter changes, elbows, and tees, an adaptive mesh refinement strategy is adopted, increasing mesh density at locations with drastic curvature changes or large gradients in cross-section parameters. The Delaunay triangulation meshing method connects corresponding points of adjacent cross-sections to generate triangular mesh surfaces. Topology checks and repairs are performed during mesh construction: non-manifold edges, repeated surfaces, isolated points, and other topological errors are detected and repaired. Special processing is applied to complex connection areas such as pipe bifurcation and merging to ensure topological continuity of the mesh at these locations. The mesh is adaptively refined or simplified based on the original point cloud density, controlling model complexity while ensuring geometric accuracy. The generated preliminary mesh model possesses a complete pipeline surface topology. Based on the meshing, local refinement and accuracy optimization are performed. Local refinement adopts a mesh optimization method with point cloud constraints: the mesh vertices are used as adjustable parameters, and the original point cloud is used as geometric constraints to construct an objective function that minimizes the distance from a point to the mesh surface. ; Where, p i M represents a point in a point cloud. j The mesh is composed of facets. The positions of mesh vertices are adjusted through iterative optimization to best fit the original point cloud geometry, while maintaining the topological continuity and smoothness of the mesh. Multiple geometric constraints are introduced during the refinement process: pipe surface smoothness constraint (continuity of normal vectors between adjacent facets), circular cross-section constraint (maintaining cross-sectional roundness in straight pipe sections), and feature preservation constraint (preserving important geometric features such as pipe joints, flanges, and supports). For defect areas such as cracks, corrosion pits, and deposits, a high-resolution local reconstruction strategy is adopted. In these areas, the mesh density is increased and the point cloud details are accurately fitted, ultimately generating a high-precision 3D model M of the pipe network. final .

[0029] A high-precision three-dimensional reconstruction system based on the internal environment of urban drainage pipe network includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe network as described above.

[0030] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0031] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0032] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0033] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0034] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A high-precision three-dimensional reconstruction method for the internal environment of urban drainage pipe networks, characterized in that, Includes the following steps: S1: Obtain the task parameter file, establish a computable sensor relationship for the patrol robot's multiple sensors, and obtain the calibration parameter set; S2: The patrol robot patrols based on the task parameter file and calibration parameter set and synchronously collects multi-sensor data according to a unified timestamp to obtain raw data packets; S3: Preprocess the original data packets to obtain the cleaned synchronization dataset; S4: Based on the cleaned synchronous dataset, construct the state variables and recursively estimate them to obtain the continuous trajectory; S5: Based on the continuous trajectory, point cloud distortion removal and local mapping are performed to obtain several local sub-maps and the global initial point cloud; S6: Perform loop closure detection and global consistency optimization on the obtained local sub-graphs and the global initial point cloud to obtain the global optimized trajectory and the global consistent point cloud; S7: Based on the local optimization trajectory and the global consistent point, the pipeline centerline and cross-section sequence are reconstructed, and a high-precision three-dimensional model of the pipeline network is obtained through meshing and local refinement.

2. The high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks as described in claim 1, characterized in that, The establishment of multi-sensor computable relationships is as follows: Define the robot base coordinate system B, the engineering coordinate system W, the LiDAR coordinate system L, and the camera coordinate system C i , the IMU coordinate system I, and the odometer coordinate system O. By unifying the timeline through network time synchronization, the device timestamps of each observation are mapped to a unified time, and aligned according to the unified time during fusion; Then, a space extrinsic parameter link was established: solving the extrinsic parameters of the LiDAR base. Camera base external parameters IMU base external parameters This allows any sensor observation to be mapped to the base coordinate system via fixed extrinsic parameters. LiDAR base external parameters The scene is calibrated using a planar plate. An error function is established between the geometric elements in the LiDAR point cloud and the geometric model of the calibration object in the base system. The plane is represented in the B-frame as π=(n,d). Then, the LiDAR point p L Transform to B series: ; in, Let be the coordinates of the point in the robot base coordinate system; These are the original coordinates of the point in the LiDAR coordinate system; For each camera, the three-dimensional coordinates P of the calibration plate corner points are used. B Projected onto the image plane, camera model: ; Where u is the pixel coordinate, representing the projected position of the 3D point in the image; π(.) is the camera projection function, projecting the 3D point onto the 2D image plane; K i D is the intrinsic parameter matrix of the i-th camera; i Let be the distortion parameter vector of the i-th camera, containing radial and tangential distortion coefficients; Joint calibration using motion data: The IMU provides an angular velocity sequence, and the LiDAR provides relative motion constraints, ensuring that the rotational increments of both are consistent under the same motion. ； in, This is the rotation matrix from the LiDAR coordinate system to the base coordinate system; Let be the relative rotation matrix of the LiDAR coordinate system within the time interval [t, t+Δt]. This is the relative rotation matrix of the base coordinate system within the same time interval; When scanning a single frame of LiDAR data, let the reference time be the scan start time t0, and the unified time for the j-th point during the scan be t. j The pose TWB(t) estimated from the continuous trajectory is used to transform the point to the world pose corresponding to the reference time, thereby removing motion distortion. ； Where, p W,j Let j be the coordinates of the j-th point in the world coordinate system; For time t j The pose transformation matrix of the base relative to the world coordinate system; p L,j Let t be the original coordinates of the j-th point in the LiDAR coordinate system; j This is the timestamp for the data collection at point j; At the same time, based on the mission's requirements for scale stability, the odometer scale coefficient was incorporated. With roller skating correction parameters This allows the mileage constraint to be used directly in subsequent state estimation.

3. The high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks according to claim 1, characterized in that, The patrol robot patrols based on a task parameter file and a calibration parameter set, and synchronously collects multi-sensor data at a unified timestamp to obtain raw data packets, as detailed below: The patrol robot reads the mission parameter file and calibration parameter set Θ to complete the data acquisition configuration and execution constraint loading for this patrol. This includes writing the patrol path information from the mission parameters into the motion control module; writing the acquisition parameters into the sensor driver; and loading the coordinate chain and time mapping model so that the system can attach a unified timestamp and necessary metadata to each observation during the acquisition phase. The robot enters the pipeline for patrol according to the speed curve and attitude stabilization strategy specified by the task parameters, and all sensor data are marked and collected synchronously with a unified time axis t. The raw observations from multiple sensors throughout the patrol process are encapsulated using a unified timestamp to form a structured raw data packet D. raw At the same time, complete metadata is saved to ensure traceability and reproducibility.

4. The high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks according to claim 3, characterized in that, The preprocessing is as follows: The original data packets output by S2 are subjected to time base unification and integrity verification. Based on the time mapping model given by correction S1, the timestamps of each sensor device, including LiDAR, camera, IMU, and odometer, are converted into a unified timestamp t, and the multi-stream timeline is reconstructed by sorting them according to the unified time. Subsequently, data integrity checks are performed: packet loss rate, timestamp jump, sampling frequency drift, camera triggering asynchrony, and IMU saturation segment issues are statistically analyzed for each data stream. Based on time alignment, various observations are cleaned and normalized. For IMU data, zero-bias initial value estimation, peak removal and low-pass filtering are performed, and abnormal saturated samples are removed or invalidated. For odometry data, zero-point reset, pulse anomaly detection and short-time slip segment identification are performed, and usable incremental mileage sequences are output. For camera images, blur detection, overexposure and underexposure detection, strong glare ratio estimation are performed, severely unusable frames are removed, and distortion correction and brightness normalization can be performed to improve the stability of subsequent features. For LiDAR point clouds, distance threshold clipping, outlier removal, and intensity anomaly suppression are performed, and downsampling is carried out as needed to balance accuracy and computational load; thus, a cleaned synchronous dataset is obtained.

5. The high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks according to claim 1, characterized in that, Based on the cleaned synchronization dataset, state variables are constructed and recursively estimated to obtain continuous trajectories, as detailed below: Based on the cleaned synchronized dataset D, with key moments t on a unified time axis k The goal of establishing a recursive estimation framework is to obtain the continuous pose trajectory of the robot base in the engineering frame. At time t k Define the state vector: ； Among them, R WB,k For posture, p WB,k For position, v WB,k For speed, b g,k ,b a,k These are the gyroscope and the zero bias accelerator, respectively; k s λ is the odometer scaling factor. k Slip ratio; IMU measurement model: ; Where, ω m ω is the gyroscope measurement; b is the true angular velocity; g For zero bias of the gyroscope; n g To measure noise for a gyroscope; a m For acceleration, a is the actual acceleration; b is the acceleration. a For accelerometer zero bias; n a To measure noise for accelerometers; The kinematic equations for continuous time, expressed in W-frame, with gravity g W Given: ; in, R is the time derivative of the rotation matrix of the base relative to the world coordinate system; WB This is the rotation matrix of the base relative to the world coordinate system; Indicates an antisymmetric matrix operator; g is the time derivative of the velocity of the base in the world coordinate system; W v is the gravity vector in the world coordinate system. WB The velocity of the base in the world coordinate system; p represents the position-time derivative of the base in the world coordinate system. WB This refers to the position of the base in the world coordinate system. Zero bias employs random walk: ; in, n is the time rate of change of the gyroscope's zero bias; bg The noise is zero-bias random walk noise of the gyroscope, which follows a Gaussian distribution; n is the time rate of change of the accelerometer's zero bias. ba This refers to the accelerometer zero-bias random walk noise. During recursion, IMU high-frequency data is used in [t] k ,t k+1 Integrating upwards forms a prediction of the state, resulting in a high-frequency skeleton of the continuous trajectory; To effectively incorporate the high-frequency IMU into the keyframe recursion, in each segment [t] k ,t k+1 Pre-integrating the IMU yields the relative motion quantities (ΔR, Δv, Δp): ; Where, ΔR k For the time interval [t] k ,t k+1 The relative rotation matrix within ]; It is the cumulative product from time k to k+1; The exponential mapping of rotation matrices transforms antisymmetric matrices into rotation matrices; ω m,j b is the gyroscope measurement at time j; g,k Δt is the gyroscope zero-bias estimate at time k; j Let Δv be the j-th sampling interval; k ΔR represents the change in relative velocity over a time interval. k→j Let a be the relative rotation matrix from time k to time j; m,j b is the accelerometer measurement at time j; a,k Δp is the accelerometer zero bias estimate at time k. k Δv represents the change in relative position over a time interval. k→j This represents the change in relative velocity from time k to time j; The state prediction from k→k+1 can be expressed in discrete form using pre-integration: ; in, R is the prediction rotation matrix at time k+1; WB,k Let be the estimated rotation matrix at time k; The predicted velocity at time k+1; v WB,k The estimated velocity at time k; p is the predicted position at time k+1; WB,k The estimated position at time k; Simultaneously, the covariance is recursively calculated: ; Among them, F k G k To linearize the Jacobian, Q k The covariance of IMU noise and zero-bias random walk noise; Let be the prediction covariance matrix at time k+1; At each critical moment t k The predicted state is updated using the cleaned sensor observations to obtain a continuous trajectory.

6. The high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks according to claim 5, characterized in that, At each critical moment t k The predicted state is updated using the cleaned sensor observations to obtain a continuous trajectory, as follows: Register the current scan before distortion correction with the previous local map to obtain the relative pose observation. residual Written in logarithmic form as a Lie group: ； in, Let be the transformation matrix from the world coordinate system to the base coordinate system at time k; Let be the transformation matrix from the world coordinate system to the base coordinate system at time k+1; And minimize residuals Used for EKF updates; The odometer gives the increment Δs meas,k Establish a measurement model: ； Among them, z O,k For the odometer in the time period [t] k ,t k+1 The measured observation value of h(x) k ) is the prediction function for odometer observations; k s λ is the odometer scale factor; k p is the slip ratio parameter. WB,k+1 Let p be the position of the base in the world coordinate system at time k+1; WB,k Let k be the position of the base in the world coordinate system at time k. It is the Euclidean norm; Corresponding residuals: ； Where, r O For odometer observation residuals; Reduce the weight of this factor when slippage risk is detected; Relative pose is obtained using keyframe feature tracking. Constraints are formed by mapping extrinsic parameters to the base: ； in, This refers to a relative transformation from the perspective of the base coordinate system; Keyframe estimation is obtained Then, high-frequency attitude is obtained by IMU integration between two keyframes, thereby outputting a continuous trajectory. .

7. The high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks according to claim 6, characterized in that, The point cloud distortion correction and local mapping based on continuous trajectories are performed to obtain several local sub-maps and a global initial point cloud, as detailed below: The original continuous trajectory scan output in step S4 is subjected to distortion correction processing to eliminate the geometric distortion of the point cloud caused by robot motion. For each frame of LiDAR scan, the acquisition time of the j-th laser point is t. j The original point coordinates are p L,j The extrinsic parameters established through S1 Continuous trajectory interpolation transforms each point from the LiDAR coordinate system at its acquisition time to the world coordinate system corresponding to a unified reference time. The distortion correction formula is: ; That is, first transform the points to the world frame, and then unify them to the coordinate system of the reference time. For time periods when the trajectory quality is below the threshold, a distortion removal strategy is adopted. After distortion correction, the continuous point cloud sequence is divided into several local sub-images, each covering a certain spatial range. Fine point cloud registration and geometric optimization are performed within each sub-image. For the distorted point cloud sequence within each sub-image, frame-by-frame registration is used for local optimization: the first frame of the sub-image is used as the reference coordinate system, and the subsequent frames are transformed through point cloud registration. All point clouds are accumulated into the sub-image coordinate system to form a local dense map. After the local subgraphs are constructed, all subgraphs are spliced ​​together according to their positions in the global coordinate system to form a global initial point cloud covering the entire inspection section. Subgraph stitching utilizes the global pose information provided by the S4 continuous trajectory: the reference coordinate system of each subgraph is linked to the coordinate system at the corresponding time step. Transform to the world coordinate system, that is: ； This unifies all local subgraphs into the same global coordinate frame, resulting in the final output of a global initial point cloud. It covers the entire pipeline inspection route.

8. The high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks according to claim 7, characterized in that, The process of performing loop closure detection and global consistency optimization on the obtained local sub-graphs and the global initial point cloud to obtain the globally optimized trajectory and the globally consistent point cloud is as follows: Loop closure detection is performed on the local subgraph set output by S5 and the global initial point cloud to identify the spatial locations that the robot repeatedly passes through during the inspection process, providing closure constraints for subsequent global consistency optimization; For candidate loops that pass geometric verification, precise inter-subgraph registration is performed to obtain the relative transformation of the candidate loops. and its covariance estimation To construct a reliable set of closed constraints for global optimization; Based on the detected loop closure constraints, a global pose graph is constructed and optimized to eliminate the cumulative drift in the S4 continuous trajectory and obtain a globally consistent optimized trajectory. The pose map uses keyframe poses. For each node, there are three types of edge constraints: odometry constraints between adjacent keyframes, loop closure constraints, and absolute position constraints; the global optimization objective function combines the Mahalanobis distance of all constraints. ； Where, r ij For the residual, Σ ij To address the covariance, a manifold optimization method is used to solve the problem on SE(3) to obtain a globally consistent set of keyframe poses. The optimized keyframe poses are then used to reconstruct continuous trajectories through spline interpolation or IMU reintegration to form a globally optimized trajectory. ; Utilizing global trajectory optimization The original LiDAR scans are distorted and globally registered to generate a geometrically globally consistent final point cloud model. .

9. The high-precision three-dimensional reconstruction method based on the internal environment of urban drainage pipe networks according to claim 8, characterized in that, Based on the local optimization trajectory and globally consistent points, the pipeline centerline and cross-section sequence are reconstructed, and a high-precision 3D model of the pipeline network is obtained through meshing and local refinement, as detailed below: Based on global optimization trajectory Consistent with global points Pipeline centerline extraction and cross-section sequence reconstruction are performed to convert discrete point clouds into parameterized pipeline geometric models. Based on the reconstructed centerline and cross-section sequence, a three-dimensional mesh model of the pipeline is constructed to form a continuous surface representation with correct topological relationships. The meshing adopts a cross-section-based parametric method: mesh nodes are set along the centerline at uniform arc length intervals Δs, and sampling points in the circumferential direction are generated at each node position according to the cross-section parameters to form a regular mesh topology in cylindrical coordinates; corresponding points of adjacent cross-sections are connected by the Delaunay triangulation meshing method to generate a triangular mesh surface. Based on the meshing, local refinement and accuracy optimization are performed. Local refinement adopts a mesh optimization method with point cloud constraints: the mesh vertices are used as adjustable parameters, and the original point cloud is used as geometric constraints to construct an objective function that minimizes the distance from a point to the mesh surface. ； Where, p i M represents a point in a point cloud. j The mesh is composed of surface patches; the positions of the mesh vertices are adjusted through iterative optimization to ensure that the mesh surface best fits the geometry of the original point cloud, while maintaining the topological continuity and smoothness of the mesh; the final high-precision 3D model M of the pipeline network is generated. final .

10. A high-precision three-dimensional reconstruction system for the internal environment of urban drainage pipe networks, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the high-precision three-dimensional reconstruction method based on the internal environment of an urban drainage pipe network as described in any one of claims 1-9.

Images