A radar and photoelectric inspection map construction system

Abstract

This invention discloses an inspection map construction system based on radar and photoelectric technology. The system includes a 3D LiDAR module, a photoelectric positioning module, a data processing module, and a map reconstruction module. The 3D LiDAR module collects environmental data to construct a global 3D point cloud map and outputs the robot's initial pose. The photoelectric positioning module corrects parking deviations using cross-arranged photoelectric sensors and acquires wheel axle images through a top-view camera. The data processing module preprocesses the map, fuses multi-source data, performs Kalman filtering, outputs the optimal pose, and completes secondary pose correction by comparing the wheel axle images with a standard template. The map reconstruction module dynamically optimizes and reconstructs the inspection map. Through multi-module collaboration and hierarchical correction, it improves positioning accuracy and scene adaptability, effectively solving problems such as positioning drift and weak dynamic map update capabilities in traditional technologies, and meeting the high-precision requirements of rail transit vehicle undercarriage inspection.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for rail transit, specifically to an inspection map construction system based on radar and photoelectric technology. Background Technology

[0002] With the rapid development of the rail transit industry, the operating mileage and vehicle inventory of trains continue to grow, making train maintenance increasingly urgent. Traditional train inspections rely on manual labor, which suffers from low efficiency, high labor intensity, and the susceptibility of inspection results to the influence of personnel experience and condition, making it difficult to meet the safety requirements under high-density operation. Therefore, intelligent inspection robots are gradually being applied in the rail transit field. Undercarriage inspection, due to its enclosed environment, dense equipment, and narrow space, places extremely high demands on the robot's positioning accuracy and map adaptation capabilities.

[0003] Existing map-building technologies for inspection robots mainly rely on single-sensor or general multi-sensor fusion solutions, which have the following shortcomings: Low localization progress: Traditional 2D LiDAR map building is highly dependent on fixed environmental references, but the environment of rail transit sections contains temporary obstacles and dynamic changes in equipment, which easily leads to positioning drift and cannot meet the centimeter-level alignment requirements of undercarriage inspection; Poor scene adaptability: General robot map building solutions are not optimized for the positional relationship between trains and robots, ignoring personalized features such as differences in train car length and minor adjustments to the train head position, resulting in alignment deviations between the robot and the inspection points, affecting the accuracy of fault detection; Low efficiency of multi-sensor data fusion: Although some solutions combine laser and vision sensors, they lack collaborative strategies for rail transit scenarios. Problems such as laser point cloud noise and visual image reflection can easily lead to the accumulation of fusion errors, affecting map reliability.

[0004] Therefore, in order to solve the problems existing in the prior art, this invention proposes an inspection map construction system based on radar and photoelectric technology. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an inspection map construction system based on radar and photoelectric technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A patrol map construction system based on radar and photoelectric technology, comprising:

[0008] The 3D LiDAR module is used to collect 3D point cloud data, inertial measurement data and odometer data of the environment, construct a global 3D point cloud map based on the collected data and output the robot's initial pose;

[0009] The photoelectric positioning module includes a cross-arranged photoelectric sensor and a top-view camera. The photoelectric sensor is used to detect the position of the train's front end and correct the robot's parking deviation; the top-view camera is used to acquire images of the train's wheel axles.

[0010] The data processing module is communicatively connected to the photoelectric positioning module and the 3D lidar module. It preprocesses the 3D point cloud map, combines the inertial measurement data, mileage data, and initial pose, and optimizes the robot's optimal pose through Kalman filtering, then outputs the real-time positioning result. It calculates the position deviation based on the matching result between the wheel axle image and the standard wheel axle template, and performs secondary correction on the robot's pose relative to the train based on the position deviation to obtain the pose correction result.

[0011] The map reconstruction module reconstructs and generates the inspection map based on the 3D point cloud map and pose correction results.

[0012] As a further improvement of the present invention, the three-dimensional lidar module includes a map acquisition and construction submodule, which is used to acquire three-dimensional lidar data, inertial measurement data and odometer data through robot closed-loop path scanning, generate a three-dimensional point cloud map composed of multiple frames of motion data, perform rasterization and noise reduction processing on the three-dimensional point cloud map, and set the initial positioning value of the robot relative to the coordinate system of the three-dimensional point cloud map.

[0013] As a further improvement of the present invention, the data processing module includes a data acquisition and matching submodule. The data acquisition and matching submodule is used to obtain the robot's initial pose by analyzing the inertial measurement data and odometry data based on the initial positioning value, and to match the current frame of the lidar with the preprocessed three-dimensional point cloud map frame to output the robot's six-degree-of-freedom pose. The six-degree-of-freedom pose includes X-axis position, Y-axis position, Z-axis position, yaw angle, pitch angle, and roll angle. The six-degree-of-freedom pose is used as an observation value and fused with the inertial measurement data and odometry data to output the optimal pose of the robot when it reaches the inspection target position.

[0014] As a further improvement of the present invention, the secondary pose correction in the data processing module includes: capturing wheel and axle images of each carriage of the train using the upward-looking camera, extracting wheel and axle features to analyze the actual position of the wheel and axle, comparing the actual position of the wheel and axle with the position of a preset standard wheel and axle template, calculating the wheel and axle position deviation, which is the pose deviation of the robot relative to the current carriage, adjusting the robot's pose relative to the train according to the wheel and axle position deviation, and outputting the pose correction result after secondary correction. The pose correction result includes the six-degree-of-freedom correction value of the robot relative to the current carriage, the relative coordinate relationship between the robot and the train detection points, and pose calibration marks for map reconstruction. The pose calibration marks are used to associate the coordinate mapping relationship between the global three-dimensional point cloud map and the local features of the train, and are output to the map reconstruction module.

[0015] As a further improvement of the present invention, the map reconstruction module includes: a coordinate mapping submodule, used to establish a dynamic mapping relationship between the global three-dimensional point cloud map coordinate system and the train local feature coordinate system based on the pose calibration mark; a pose correction submodule, used to inversely superimpose the six-degree-of-freedom correction value onto the robot's historical positioning trajectory in the global three-dimensional point cloud map to correct the cumulative error of the robot's pose in the global map; and a feature update submodule, used to update the coordinate information of the train inspection points in the global three-dimensional point cloud map according to the relative coordinate relationship between the robot and the train inspection points combined with the dynamic mapping relationship and complete the dynamic optimization and reconstruction of the inspection map.

[0016] As a further improvement of the present invention, the photoelectric positioning module corrects parking deviation by means of: during the process of the robot traveling towards the preset theoretical parking point based on the optimal pose, the photoelectric sensors arranged in a cross pattern scan the train head area in real time, identify the position of the train head through the preset train head edge feature model, and predict the parking deviation that will occur when reaching the theoretical parking point based on the current driving speed, the distance to the theoretical parking point and historical driving error data, adjust the robot's driving trajectory according to the predicted parking deviation, and after reaching the theoretical parking point, detect the deviation between the actual train head position and the theoretical position a second time through the photoelectric sensors, and make compensation adjustments in combination with the predicted parking deviation to complete the correction of parking deviation.

[0017] As a further improvement of the present invention, the map reconstruction module further includes: calling the pose calibration marker to parse the coordinates of the marker points in the global 3D point cloud map coordinate system and the coordinates of the marker points in the train local feature coordinate system; solving the transformation matrix of the two coordinate systems using the least squares method to establish a dynamic mapping relationship, which is updated in real time as the train position changes; extracting the robot's historical positioning trajectory data from the global 3D point cloud map; superimposing the six degrees of freedom correction values ​​in reverse order to the historical pose at the corresponding time; calculating the correction amount for each historical pose; wherein the correction amounts for the X-axis, Y-axis, and Z-axis are allocated to the historical trajectory segments using linear interpolation. The heading angle, pitch angle, and roll angle corrections are compensated for rotational errors using spherical linear interpolation. Based on the relative coordinate relationship between the robot and the train inspection points, the local coordinates of the inspection points are converted into absolute coordinates in the global 3D point cloud map coordinate system using a transformation matrix. The converted coordinates are compared with the original coordinates of the inspection points stored in the map. If the deviation exceeds a preset threshold, the map data is updated with the converted coordinates, and the coordinate update timestamp and the corresponding pose calibration mark version are recorded. The corrected historical trajectory data and the updated inspection point coordinate information are integrated to generate an inspection map that includes global environmental features, robot trajectory, and train dynamic features.

[0018] As a further improvement of the present invention, the map reconstruction module further includes: after establishing the dynamic mapping relationship, optimizing the transformation matrix through a random sampling consistency algorithm, eliminating abnormal mapping data caused by marker point identification errors, and retaining transformation matrix parameters with confidence levels higher than a preset threshold; after generating the inspection map, calculating the matching degree between the corrected historical trajectory and the static features in the global three-dimensional point cloud map; if the matching degree is lower than a preset threshold, re-establishing the coordinate mapping relationship and subsequent steps until the matching degree is higher than the preset threshold, wherein the matching degree is calculated by the mean Euclidean distance of the overlapping area of ​​the point cloud.

[0019] As a further improvement of the present invention, the map construction method of the system includes the following steps: collecting environmental data through closed-loop path scanning, constructing an initial global three-dimensional point cloud map and performing rasterization and noise reduction preprocessing, and setting the initial positioning value of the robot; based on the initial positioning value, fusing inertial measurement data, mileage data and lidar matching results, outputting the optimal pose of the robot through Kalman filtering, controlling the robot to autonomously navigate to the target track, the photoelectric positioning module predicts the parking deviation and adjusts the trajectory in advance during the robot's movement, and after arriving at the target track, the position of the train head is detected by the cross-type photoelectric sensor to complete the correction of the parking deviation; capturing train wheel axle images, calculating the pose deviation by comparing the wheel axle images with a standard template, and outputting the pose correction result including six-degree-of-freedom correction values, relative coordinate relationships and pose calibration marks; establishing coordinate mapping based on the pose correction result, correcting the global pose error and updating the coordinates of the detected points to complete the dynamic reconstruction and storage of the inspection map.

[0020] The beneficial effects of this invention are:

[0021] (1) Improve positioning accuracy and stability. By combining the three-dimensional lidar module and the Kalman filter algorithm, inertial measurement data and mileage data are integrated to achieve the robot's global optimal pose output, and the positioning error can be controlled within the centimeter level. At the same time, the cross-type photoelectric sensor of the photoelectric positioning module accurately corrects the robot's parking deviation through a correction mechanism, avoiding the problem of false triggering of single-point detection. On this basis, the wheel axle secondary correction further eliminates the local deviation caused by the difference in carriage length through the six-degree-of-freedom correction value, and finally achieves millimeter-level alignment between the robot and the inspection points of the train, effectively preventing the robotic arm from colliding with the equipment under the train and ensuring the safety of the inspection.

[0022] (2) Enhanced adaptability to complex scenarios: The 3D LiDAR module reduces its dependence on fixed environmental references through closed-loop path scanning and point cloud preprocessing technology, solving the positioning drift problem of traditional 2D LiDAR in scenarios with temporary obstacles in the depot and dynamic changes in equipment. At the same time, the dynamic mapping relationship between the global coordinate system and the train's local coordinate system established by the pose calibration mark and coordinate mapping submodule can be updated in real time as the train stops, adapting to the personalized characteristics of different train formations and breaking through the strong dependence of general map construction schemes on static environments.

[0023] (3) Improve anti-interference capability and improve map construction reliability. By layering and fusing multi-sensor data through the data processing module, noise and environmental interference are effectively filtered out, solving the measurement deviation problem of a single sensor in a complex vehicle under environment and improving the robustness of map construction. Attached Figure Description

[0024] Figure 1 This is a system block diagram of an inspection map construction system based on radar and photoelectric technology according to the present invention.

[0025] Figure 2 This is a schematic diagram of the map reconstruction module of the system of the present invention;

[0026] Figure 3 This is a flowchart of the photoelectric positioning module deviation correction process of the system of the present invention;

[0027] Figure 4 This is a flowchart of the map reconstruction module of the system of the present invention;

[0028] Figure 5 This is a flowchart of the inspection map construction method of the inspection map construction system based on radar and photoelectric technology according to the present invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0030] This embodiment proposes a patrol map construction system based on radar and photoelectric technology, such as... Figure 1 As shown, it includes: a 3D LiDAR module, an optoelectronic positioning module, a data processing module, and a map reconstruction module.

[0031] The 3D LiDAR module is used to collect 3D point cloud data, inertial measurement data and odometer data of the environment, construct a global 3D point cloud map based on the collected data and output the robot's initial pose;

[0032] The 3D LiDAR module is the core unit of the system for acquiring environmental spatial information. It is mainly responsible for collecting 3D point cloud data, inertial measurement data and mileage data of the inspection environment, and constructing a global 3D point cloud map based on these data, while outputting the robot's initial pose.

[0033] 3D point cloud data is acquired by scanning the surrounding environment with laser beams emitted by LiDAR, which can accurately reflect the 3D coordinates and shape characteristics of static objects in the environment; inertial measurement data is collected by an integrated inertial measurement unit to capture the changes in acceleration and angular velocity during the robot's movement, and to help determine the robot's motion state; odometer data records the distance and direction information traveled by the robot, providing a basic reference for pose calculation.

[0034] During map building, the module collects data through the robot's closed-loop path scanning. This means the robot starts from the starting point, travels along a preset path, and eventually returns to the starting point, ensuring the collected data forms a complete closed loop. Based on this closed-loop data, the module fuses and stitches together point cloud data collected during multiple frames of movement to generate a global 3D point cloud map. The initial pose serves as the robot's starting reference in this map coordinate system, providing coordinate reference for subsequent localization and navigation.

[0035] Specifically, such as Figures 1 to 5 As shown, the three-dimensional LiDAR module includes a map acquisition and construction submodule, which is used to acquire three-dimensional LiDAR data, inertial measurement data and odometry data through robot closed-loop path scanning, generate a three-dimensional point cloud map composed of multiple frames of motion data, perform rasterization and noise reduction processing on the three-dimensional point cloud map, and set the initial positioning value of the robot relative to the coordinate system of the three-dimensional point cloud map.

[0036] The map acquisition and construction submodule completes environmental data collection through the robot's closed-loop path scanning. Closed-loop path scanning refers to the robot starting from a preset starting point, traveling along a planned inspection path, covering the target track and its surrounding environment, and then returning to the starting point, forming a complete closed trajectory. The core advantage of this path design is that by ensuring the trajectory overlaps end to end, it provides an error calibration benchmark for subsequent data fusion, reducing the cumulative errors caused by long-distance travel.

[0037] During the scanning process, the submodule simultaneously collects three types of key data: 3D LiDAR data, inertial measurement data, and odometry data. The 3D LiDAR data uses high-frequency laser pulses to scan the environment, generating point cloud data containing spatial coordinates and reflection intensity, accurately capturing the shape and position of static features such as tracks, walls, and equipment. The inertial measurement data records the robot's acceleration and angular velocity changes in real time, reflecting the robot's dynamic posture. The odometry data records the distance and direction traveled by the robot, assisting in tracking its trajectory. These three types of data are synchronized spatiotemporally through a timestamp synchronization mechanism, ensuring that each frame of point cloud data matches the corresponding motion state parameters, providing a consistent foundation for subsequent map stitching.

[0038] Based on the collected multi-frame motion data, the submodule performs fusion and generation of a 3D point cloud map. Multi-frame data fusion is achieved through coordinate transformation: each frame of point cloud data is transformed to the global coordinate system based on the robot's current pose, and then a point cloud registration algorithm is used to match features in overlapping areas of adjacent frames, eliminating inter-frame deviations. For repetitive structures commonly found in rail transit scenarios, such as track spacing and equipment supports, the submodule identifies stable feature points through feature extraction algorithms. These feature points are used as a benchmark to optimize inter-frame registration accuracy, ensuring that a continuous and complete global point cloud map is formed after stitching together multiple frames of data.

[0039] To improve map usability and subsequent positioning efficiency, the submodule performs rasterization and noise reduction on the generated 3D point cloud map.

[0040] Rasterization divides a three-dimensional space into regular grids of fixed dimensions, with each grid serving as an independent spatial unit to store point cloud data within that region. This processing method reduces data redundancy, accelerates point cloud matching efficiency in subsequent localization processes, and enhances the identifiability of environmental features through feature statistics of grid units, such as point cloud density and average reflectance intensity.

[0041] The noise reduction process filters out noise introduced during data acquisition. Noise primarily originates from laser reflection interference, environmental interference, and sensor errors. The submodule identifies and removes isolated outliers in a single frame through multi-frame comparative analysis; it removes false point clouds caused by reflections by using a reflection intensity threshold; and it filters discrete points that do not conform to physical morphology through spatial continuity verification. After noise reduction, the map retains only stable static feature point clouds from the environment, providing a reliable feature benchmark for subsequent positioning.

[0042] After generating the map, the map acquisition and construction submodule needs to set the robot's initial positioning values ​​relative to the 3D point cloud map coordinate system. These initial positioning values ​​are the robot's reference coordinates and attitude parameters on the map, directly determining the starting point accuracy for subsequent navigation and positioning.

[0043] During the setup process, the submodule first uses the starting point of the closed-loop path as a physical reference point. By matching the point cloud data collected at this point by the LiDAR with surrounding environmental features, such as right angles of walls and edges of fixed equipment, the three-dimensional coordinates of this point in the map coordinate system are determined. At the same time, combined with the attitude data of the inertial measurement unit at the starting point, the robot's initial heading angle, pitch angle, and roll angle are determined to ensure that the initial attitude is consistent with the actual physical state.

[0044] In rail transit scenarios, the initial positioning values ​​are calibrated in conjunction with fixed scene features such as track number and track spacing. For example, by identifying the edge features of the rails on both sides of the track, the X and Y axis parameters of the initial coordinates are fine-tuned to ensure that the map coordinate system is consistent with the actual track direction, laying the foundation for the robot to travel accurately along the track in the future.

[0045] Through the above process, the 3D point cloud map generated by the map acquisition and construction submodule can not only fully reflect the spatial characteristics of the inspection environment, but also effectively offset environmental interference and equipment errors through closed-loop path design, multi-source data fusion and fine preprocessing, providing high-precision basic map support for the positioning and navigation of the entire system.

[0046] The photoelectric positioning module includes a cross-arranged photoelectric sensor and a top-view camera. The photoelectric sensor is used to detect the position of the train's front end and correct the robot's parking deviation; the top-view camera is used to acquire images of the train's wheel axles.

[0047] The photoelectric positioning module adopts a multi-sensor collaborative design, including cross-arranged photoelectric sensors and an upward-looking camera, which is mainly used to achieve precise alignment between the robot and the train.

[0048] A series of cross-arranged photoelectric sensors are mounted on the front of the robot in an X-shaped structure. Their core function is to detect the position of the train's front end and correct the robot's parking deviation. This cross-arrangement expands the detection range, reduces false judgments caused by single sensors being obstructed or reflecting light, and improves the stability of front end edge recognition. As the robot approaches the train, the sensors scan the front end area in real time, determining the relative positional relationship by identifying the front end edge features, thus providing a basis for adjusting the parking position.

[0049] An upward-facing camera is mounted on top of the robot, with its lens pointing upwards towards the bottom of the train, to acquire images of the wheel axles of each carriage. As a fixed feature of the train, the wheel axles have a stable correspondence with the carriages. By capturing images of the wheel axles, their actual position information can be extracted, providing a crucial reference for subsequent pose correction.

[0050] Specifically, such as Figures 1 to 5 As shown, the photoelectric positioning module corrects parking deviation by: during the robot's journey towards the preset theoretical parking point based on the optimal pose, cross-arranged photoelectric sensors scan the train's front area in real time, identify the front position through a preset front edge feature model, predict the parking deviation that will occur when reaching the theoretical parking point based on the current driving speed, the distance to the theoretical parking point, and historical driving error data, adjust the robot's driving trajectory according to the predicted parking deviation, and after reaching the theoretical parking point, use photoelectric sensors to detect the deviation between the actual front position and the theoretical position a second time, and combine the predicted parking deviation to make compensation adjustments to complete the correction of the parking deviation.

[0051] The cross-arranged photoelectric sensors are the core sensing units for parking deviation correction. They are distributed at the front of the robot in an X-shaped structure, forming a multi-angle, full-coverage detection area. This layout reduces blind spots caused by single-direction sensors due to occlusion at the edge of the train's front, light reflection, or stray light from the environment, thus improving the stability of the train's front position detection. As the robot moves towards the preset theoretical parking point based on its optimal pose, the sensors continuously scan the train's front area with high-frequency pulse signals. When the emitted detection beam encounters the edge of the train's front, it is reflected. The sensors receive the reflected signals and convert them into electrical signals, forming contour perception data of the train's front area.

[0052] To accurately identify the position of the train's front end, the module has a built-in preset edge feature model of the train's front end. This model is constructed based on the typical geometric features of the train's front end and includes key parameters such as edge straightness, corner angles, and contour gradient changes. By comparing the contour data collected by sensors in real time with the model features, the actual edge position of the train's front end can be quickly located, distinguishing the front end from the surrounding environment, such as the boundaries of equipment and walls next to the track, ensuring the accuracy of the train's front end position identification.

[0053] Based on the real-time position of the robot's front end, the module predicts parking deviation by combining the robot's current driving status parameters. The prediction process integrates three types of key data: the current driving speed is obtained through feedback from the odometer and drive motor, reflecting the robot's speed; the distance to the theoretical parking point is calculated using LiDAR positioning data and theoretical coordinates, quantifying the remaining driving distance; and historical driving error data is a statistical analysis of the robot's past parking deviations under similar conditions on the same track, including typical error patterns at different speeds and distances.

[0054] The module fuses three types of data using an error prediction model to calculate the potential positional deviation the robot may experience when it travels along the current trajectory to the theoretical parking point. For example, if historical data shows that the robot is prone to premature parking errors due to inertia at a certain speed range, and the current distance to the theoretical parking point is relatively close, the model will predict the corresponding amount of premature deviation. Based on the predicted parking deviation, the module sends trajectory adjustment commands to the robot's drive system: if a lateral deviation is predicted, lateral displacement compensation is achieved by adjusting the speed difference between the left and right wheels; if a longitudinal deviation is predicted, the longitudinal position is adjusted by fine-tuning the driving speed, ensuring that the robot has eliminated most of the deviation before reaching the theoretical parking point.

[0055] After the robot arrives at the theoretical parking point, the photoelectric positioning module initiates a secondary detection process. Using cross-type photoelectric sensors, it re-scans the front area of ​​the vehicle to obtain deviation data between the actual front position and the theoretical front position corresponding to the parking point. The core purpose of this secondary detection is to correct the difference between the predicted deviation and the actual scenario—due to uncontrollable factors such as changes in ground friction and slight bumps during driving, the predicted deviation may differ from the actual deviation; the secondary detection can capture this real-time difference.

[0056] The module fuses the actual deviation obtained from the secondary detection with the previously predicted parking deviation to calculate the final compensation adjustment. If the actual deviation and the predicted deviation are in the same direction, the compensation is the sum of the two; if they are in opposite directions, the difference between the two is used as the compensation to avoid over-adjustment. The adjustment command is executed through the drive system, and the robot performs minute displacements or posture rotations until the deviation between the actual front position and the theoretical position is reduced to within a preset threshold, completing the final correction of the parking deviation.

[0057] Through a continuous process of real-time scanning and recognition, dynamic prediction and pre-adjustment, and secondary detection and compensation, the photoelectric positioning module effectively offsets the cumulative errors, environmental interference, and uncontrollable factors during the robot's movement, ensuring that the deviation between the robot's final stopping position and the theoretical stopping point is controlled within a very small range, providing a stable initial position reference for subsequent wheel axle image acquisition, pose secondary correction, and other processes.

[0058] The data processing module is communicatively connected to the photoelectric positioning module and the 3D lidar module. It preprocesses the 3D point cloud map, combines the inertial measurement data, mileage data, and initial pose, and optimizes the robot's optimal pose through Kalman filtering, then outputs the real-time positioning result. It calculates the position deviation based on the matching result between the wheel axle image and the standard wheel axle template, and performs secondary correction on the robot's pose relative to the train based on the position deviation to obtain the pose correction result.

[0059] The data processing module is the core computing unit of the system. It maintains real-time communication with the 3D LiDAR module and the photoelectric positioning module, and is responsible for data fusion, optimization, and correction calculations. Specifically, it performs the following functions:

[0060] The data processing module preprocesses the global 3D point cloud map output by the 3D LiDAR module. It divides the point cloud data into spatial grids through rasterization, removing noise points and redundant data to improve map clarity and usability. The module then combines the preprocessed map data, inertial measurement data, odometer data, and initial pose, using a Kalman filter algorithm for data fusion optimization. Kalman filtering effectively compensates for measurement errors from a single sensor by predicting and updating sensor data, such as cumulative drift in inertial measurement data and deviations in odometer data due to ground slippage. Ultimately, it outputs the robot's optimal pose and real-time positioning results, ensuring the robot's global positioning stability. Finally, the module compares and analyzes wheel axle images acquired by the top-view camera with a pre-set standard wheel axle template. The standard wheel axle template contains the standard position and morphological features of the wheel axles. By extracting wheel axle features from the actual wheel axle images and comparing them with the template, the positional deviation of the wheel axles can be calculated. This deviation directly reflects the robot's pose deviation relative to the current vehicle compartment. Based on this deviation, the module performs secondary correction on the robot's pose, generating a pose correction result to ensure accurate alignment between the robot and the train inspection points.

[0061] Specifically, such as Figures 1 to 5 As shown, the data processing module includes a data acquisition and matching submodule. This submodule is used to analyze the robot's initial pose based on the initial positioning value using the inertial measurement data and odometry data. It then matches the current frame of the lidar with the preprocessed 3D point cloud map frame to output the robot's six-degree-of-freedom pose, which includes X-axis position, Y-axis position, Z-axis position, yaw angle, pitch angle, and roll angle. Finally, it uses the six-degree-of-freedom pose as an observation value to fuse with the inertial measurement data and odometry data, outputting the optimal pose for the robot to reach the inspection target position.

[0062] The data acquisition and matching submodule calculates the robot's initial pose based on the initial positioning values ​​set by the 3D LiDAR module, combined with inertial measurement and odometry data. The initial positioning values ​​are the robot's starting coordinates and attitude reference in the map coordinate system, while the inertial measurement and odometry data are used to refine the dynamic accuracy of the initial pose.

[0063] Inertial measurement data contains acceleration and angular velocity information of the robot's motion, reflecting attitude changes in real time during dynamic processes such as robot startup and turning. Odometry data records the cumulative distance traveled and directional deflection angle, quantifying the robot's displacement trajectory. The submodule fuses these two types of data through a kinematic model: starting from the initial positioning value, it uses odometry data to calculate the robot's displacement increment, and combines this with inertial measurement data to correct attitude deviations during motion, ultimately obtaining the initial pose containing position and attitude parameters. This calculation method supplements the static reference with dynamic data, effectively reducing errors in the initial positioning value caused by environmental interference, laying the foundation for subsequent positioning.

[0064] To obtain the robot's real-time pose on the global map, the submodule matches the current LiDAR frame with the preprocessed 3D point cloud map frame. The preprocessed 3D point cloud map frame has undergone rasterization and denoising, preserving stable static features in the environment, such as track edges and equipment supports. These features serve as map templates to provide a reference benchmark for matching.

[0065] The current frame of the LiDAR is generated through real-time scanning and contains environmental point cloud features within the robot's current field of view. The submodule uses a feature extraction algorithm to identify key features shared with the map frame, such as geometric structures like planes and corners, from the current frame. Then, a point cloud registration algorithm calculates the spatial transformation relationship between the current frame and the map frame. During registration, the submodule prioritizes matching highly stable features, such as the intersection of walls and the ground, and the fixed contours of equipment. By minimizing the spatial distance error between feature points, the submodule determines the robot's six-degree-of-freedom pose relative to the map.

[0066] The six-DOF pose provides a complete description of the robot's spatial state: the X-axis and Y-axis positions reflect the robot's coordinates on the horizontal plane, while the Z-axis position reflects changes in altitude; the yaw angle determines the robot's direction of travel, and the pitch and roll angles reflect the robot's tilt caused by uneven or bumpy ground. Accurate acquisition of these parameters ensures that the robot can clearly define its spatial position in complex environments.

[0067] The submodule uses the six-DOF pose obtained from lidar frame matching as the observation value, and fuses it with inertial measurement data and odometry data to output the optimal pose for the robot to reach the inspection target position. This fusion improves accuracy by complementing the characteristics of different sensors: lidar frame matching observations have high positioning accuracy, but are limited by scanning frequency and have time intervals in the output; inertial measurement data and odometry data have high output frequency and can reflect dynamic changes in real time, but long-term accumulation will produce drift errors.

[0068] The submodule achieves data fusion through a filtering algorithm: using LiDAR observations as a high-precision benchmark, it calibrates the recursive results of the inertial measurement unit (IMU) and odometer. When the LiDAR outputs new observations, the submodule compares them with the recursive results, calculates the deviation, and corrects subsequent recursive processes. During the interval between two observation outputs, dynamic supplementation with IMU and odometer data maintains the continuity of pose output. This fusion strategy retains the high-precision advantage of LiDAR while compensating for the insufficient observation interval with high-frequency dynamic data. The final optimal pose output possesses both high static accuracy and high dynamic response characteristics, ensuring that the robot maintains precise alignment with the target track throughout its movement.

[0069] By establishing a baseline through initial pose calculation, obtaining real-time coordinates through lidar frame matching, and optimizing dynamic accuracy through multi-source data fusion, the data acquisition and matching submodule constructs a complete positioning chain of baseline calibration, real-time observation, and dynamic optimization. This provides reliable pose support for the robot's autonomous navigation to the inspection target location, effectively adapting to the complexity of the rail transit section environment and the high-precision requirements of the inspection task.

[0070] Specifically, such as Figures 1 to 5 As shown, the secondary pose correction in the data processing module includes: capturing wheel and axle images of each train carriage using the upward-facing camera, extracting wheel and axle features to analyze the actual wheel and axle positions, comparing the actual wheel and axle positions with preset standard wheel and axle template positions, calculating the wheel and axle position deviation, which is the robot's pose deviation relative to the current carriage, adjusting the robot's pose relative to the train based on the wheel and axle position deviation, and outputting the pose correction result after secondary correction. The pose correction result includes the robot's six-degree-of-freedom correction value relative to the current carriage, the relative coordinate relationship between the robot and the train detection points, and pose calibration markers for map reconstruction. The pose calibration markers are used to associate the coordinate mapping relationship between the global 3D point cloud map and the train's local features and are output to the map reconstruction module.

[0071] The second-stage calibration begins with axle image acquisition. A top-view camera is mounted on top of the robot, its lens pointing vertically upwards at the bottom of the train. As the robot travels along the length of the train, images are captured at preset intervals, obtaining high-resolution images of the axles of each carriage. As an inherent structure of the train, the axles have a strict geometric correspondence with the carriages and are morphologically stable and not easily affected by environmental interference, making them ideal feature carriers for calibrating the relative position of the robot and the train. During the capture process, the camera adaptively adjusts its exposure parameters to cope with changes in under-train lighting, ensuring clear imaging of key structures such as axle edges and hubs, providing high-quality image data for subsequent feature extraction.

[0072] After acquiring the axle image, the data processing module extracts axle features and analyzes its actual position using image recognition algorithms. Feature extraction focuses on the key geometric elements of the axle: first, edge detection algorithms are used to identify the axle's contour boundaries, distinguishing the axle from other structures at the bottom of the vehicle; then, shape matching algorithms are used to locate feature points such as the hub center and rim edge. The coordinates of these feature points constitute the axle's feature set.

[0073] Based on the feature set, the module combines the camera's internal parameters and installation height to convert the image's planar coordinates into three-dimensional spatial coordinates, obtaining the actual position of the wheel axle in the robot's local coordinate system. This conversion process is achieved through the principle of perspective projection, restoring two-dimensional image information to three-dimensional spatial position, ensuring the accuracy of the wheel axle's actual position measurement.

[0074] To quantify the relative positional deviation between the robot and the train, the module compares the actual positions of the wheel axles with the positions of a preset standard wheel axle template. The standard wheel axle template is a digital model built based on train design parameters, containing data such as the theoretical positions of the wheel axles in the standard carriage coordinate system and the geometric relationships between feature points. During the template construction process, statistical optimization is performed by collecting wheel axle data from multiple sets of standard trains to ensure that it adapts to the common wheel axle characteristics of different train formations.

[0075] The comparison process is achieved through coordinate mapping: the actual position of the wheel axle is transformed from the robot's local coordinate system to the standard carriage coordinate system, and then compared point by point with the theoretical position of the standard wheel axle template to calculate the spatial distance difference of the corresponding feature points. The collection of these differences constitutes the wheel axle position deviation, which directly reflects the robot's pose deviation relative to the current carriage, including translational deviation along the length of the carriage, lateral deviation perpendicular to the direction of the carriage, and angular deviation between the robot's posture and the carriage axis.

[0076] Based on the wheel axle position deviation, the data processing module adjusts the robot's pose relative to the train. The adjustment logic is executed based on the dimension of the deviation: for translational deviation, slight displacement compensation in the forward / backward or left / right directions is achieved by controlling the speed difference of the robot's drive wheels; for angular deviation, the heading angle is corrected by adjusting the robot's steering mechanism to ensure that the robot's travel direction is parallel to the train's axis. During the adjustment process, the module collects robot motion feedback data in real time, performs closed-loop verification of the adjustment effect, and continues until the deviation is reduced to within a preset threshold.

[0077] After adjustment, the module outputs the pose correction result after secondary correction. This result contains three types of core information: six-degree-of-freedom correction values, which quantify the translation correction amount of the robot in the X, Y, and Z axes and the rotation correction amount of the heading angle, pitch angle, and roll angle, used to accurately update the robot pose; the relative coordinate relationship between the robot and the train inspection points, which clarifies the spatial distance and orientation between the robot's current position and the inspection targets such as bolts and pipelines in the carriage, providing a basis for the robotic arm's inspection path planning; and pose calibration markers, which record the correspondence parameters between the global 3D point cloud map coordinate system and the train's local feature coordinate system, serving as the coordinate association benchmark for subsequent map reconstruction, ensuring the spatial consistency between the global map and the train's dynamic features.

[0078] The pose correction results are output to the map reconstruction module via a data interface. The pose calibration markers serve as key correlation elements, integrating the robot's local correction data into the global map framework. This correlation mechanism enables the global 3D point cloud map to not only include static environmental features but also integrate the location information of the train's dynamic features in real time, providing more accurate coordinate references for map retrieval in subsequent inspection tasks.

[0079] By capturing the directional features of the wheel axle, making precise comparisons and quantitative adjustments, the secondary calibration process effectively eliminates the pose deviations caused by individual differences in the train and the cumulative errors in robot navigation. This enables the alignment accuracy between the robot and the train inspection points to reach the millimeter level, providing a core guarantee for the accuracy of the undercarriage equipment inspection.

[0080] The map reconstruction module reconstructs and generates the inspection map based on the 3D point cloud map and pose correction results.

[0081] The map reconstruction module uses the global 3D point cloud map and the pose correction results output by the data processing module to achieve dynamic updating and optimization of the inspection map.

[0082] After receiving the pose correction results, the module integrates the coordinate mapping relationships and pose correction amounts with the global 3D point cloud map, updating key information such as the robot's historical trajectory and the coordinates of train inspection points on the map. By integrating real-time correction data, the module can eliminate the cumulative errors caused by the robot's long-term operation, adapt to scenarios such as changes in train stopping positions and dynamic equipment adjustments, ensuring the map remains consistent with the actual environment. Ultimately, the module generates an inspection map that includes global environmental features, the robot's trajectory, and the train's dynamic characteristics, providing accurate map support for subsequent inspection tasks.

[0083] Specifically, such as Figures 1 to 5 As shown, the map reconstruction module includes: a coordinate mapping submodule, used to establish a dynamic mapping relationship between the global 3D point cloud map coordinate system and the train local feature coordinate system based on the pose calibration marks; a pose correction submodule, used to inversely superimpose the six-degree-of-freedom correction values ​​onto the robot's historical positioning trajectory in the global 3D point cloud map to correct the cumulative error of the robot's pose in the global map; and a feature update submodule, used to update the coordinate information of the train inspection points in the global 3D point cloud map according to the relative coordinate relationship between the robot and the train inspection points combined with the dynamic mapping relationship, and to complete the dynamic optimization and reconstruction of the inspection map.

[0084] The core function of the coordinate mapping submodule is to establish a dynamic mapping relationship between the global 3D point cloud map coordinate system and the train's local feature coordinate system based on pose calibration markers. The global 3D point cloud map coordinate system is a spatial coordinate system established based on the robot's initial positioning value, covering the static features of the entire inspection environment, such as tracks, walls, and fixed equipment; the train's local feature coordinate system is a coordinate system based on the train's own structure, using inherent features of the train such as wheel axles and carriage connections as references to reflect the relative positions of various train components.

[0085] The pose calibration markers contain two types of coordinate information: one is the absolute coordinates of the marker point in the global 3D point cloud map coordinate system, obtained by the 3D LiDAR module through global positioning; the other is the relative coordinates of the same marker point in the train's local feature coordinate system, determined by the upward-looking camera combined with wheel axle feature recognition. The submodule analyzes these dual coordinates of the marker points to calculate the transformation parameters between the two coordinate systems, including translation and rotation, forming a transformation matrix. This transformation matrix can transform the coordinates of any point in the train's local feature coordinate system to the global 3D point cloud map coordinate system, or vice versa.

[0086] The core of dynamic mapping lies in real-time updates. Since the train's stopping position may undergo minor adjustments, the submodule continuously receives new pose calibration marker data. By comparing the coordinate deviations of the new marker points with historical marker points, it dynamically adjusts the transformation matrix parameters. When the train's position changes, the transformation matrix is ​​updated accordingly, ensuring that the association between the global coordinate system and the train's local coordinate system remains consistent with the actual scene, providing an accurate coordinate reference for subsequent data fusion.

[0087] The pose correction submodule is used to inversely superimpose the six-degree-of-freedom correction values ​​onto the robot's historical positioning trajectory in the global 3D point cloud map, correcting the cumulative error of the robot's pose in the global map. During long-term operation, the robot's historical positioning trajectory may gradually deviate from the actual path due to factors such as uneven ground and wheel wear, resulting in cumulative errors. If not corrected, the deviation between the global map and the actual environment will gradually increase.

[0088] Reverse overlay refers to the reverse allocation of the six-DOF correction values ​​obtained from the secondary correction to each segment of the historical trajectory according to the time series. Specifically, the submodule first extracts the complete historical positioning trajectory data of the robot from the starting point to the current position, including the pose parameters at each moment; then, according to the time nodes of the secondary correction, the six-DOF correction values ​​are decomposed into correction components corresponding to the historical stages—for the translation corrections of the X-axis, Y-axis, and Z-axis, linear interpolation is used to allocate them to each historical trajectory segment according to the travel distance ratio, ensuring that the correction amount is evenly distributed with the trajectory length; for the rotation corrections of the heading angle, pitch angle, and roll angle, spherical linear interpolation is used to allocate them according to the time ratio, avoiding abrupt changes in rotational error at the trajectory splicing points.

[0089] By overlaying the images in reverse, the pose parameters at each moment in the historical trajectory are corrected to match the actual driving path, eliminating the long-term accumulated positioning deviation. This ensures that the robot's driving trajectory recorded in the global 3D point cloud map is completely consistent with the actual inspection path, providing a reliable data foundation for subsequent map analysis and backtracking.

[0090] The feature update submodule is used to update the coordinate information of train inspection points in the global 3D point cloud map based on the relative coordinate relationship between the robot and the train inspection points, combined with the dynamic mapping relationship. Train inspection points refer to key parts of the undercarriage that need to be inspected, such as bolts, pipeline interfaces, and shock absorption devices. The accuracy of their coordinates in the global map directly affects the inspection accuracy of the inspection robot.

[0091] The relative coordinates between the robot and the train inspection points are obtained by the data processing module through visual recognition, reflecting the distance and orientation of the inspection points relative to the robot's current position. The submodule first transforms these relative coordinates to the train's local feature coordinate system, and then further transforms them into absolute coordinates in the global 3D point cloud map coordinate system through the transformation matrix established by the coordinate mapping submodule.

[0092] After the conversion, the submodule compares the newly obtained absolute coordinates with the original coordinates of the detected point stored in the global map. If the spatial distance between the two exceeds a preset threshold, it means that the original coordinates can no longer reflect the actual position of the detected point, and the submodule updates the global map data with the new coordinates; if the deviation is within the threshold range, the original coordinates remain unchanged. During the update process, the submodule synchronously records the timestamp of the coordinate update and the corresponding pose calibration mark version to achieve traceability of map data.

[0093] By continuously updating the global coordinates of train inspection points, the global 3D point cloud map can dynamically adapt to scenarios such as changes in train stopping positions and slight equipment displacements, ensuring that the recorded inspection point positions on the map are always consistent with reality, providing reliable map support for the precise alignment and fault detection of inspection robots.

[0094] The three sub-modules work together to achieve global and local spatial correlation through coordinate mapping, ensure the accuracy of historical trajectories through pose correction, and maintain the real-time performance of detection points through feature updates. Ultimately, they complete the dynamic optimization and reconstruction of the inspection map, ensuring that the map always maintains a high degree of consistency with the actual scenario of rail transit vehicle undercarriage inspection.

[0095] Specifically, such as Figures 1 to 5 As shown, the map reconstruction module further includes: calling the pose calibration marker to parse the coordinates of marker points in the global 3D point cloud map coordinate system and the coordinates of marker points in the train's local feature coordinate system; solving the transformation matrix of the two coordinate systems using the least squares method to establish a dynamic mapping relationship, which is updated in real time as the train's position changes; extracting the robot's historical positioning trajectory data from the global 3D point cloud map; superimposing the six-degree-of-freedom correction values ​​in reverse order to the historical pose at the corresponding time; calculating the correction amount for each historical pose; wherein the correction amounts for the X-axis, Y-axis, and Z-axis are allocated to the historical trajectory segments using linear interpolation; and the correction amounts for the heading angle, Y-axis, and Z-axis are calculated. The pitch and roll angle corrections are compensated for rotational errors using spherical linear interpolation. Based on the relative coordinate relationship between the robot and the train inspection points, the local coordinates of the inspection points are converted into absolute coordinates in the global 3D point cloud map coordinate system using a transformation matrix. The converted coordinates are compared with the original coordinates of the inspection points stored in the map. If the deviation exceeds a preset threshold, the map data is updated with the converted coordinates, and the coordinate update timestamp and the corresponding pose calibration mark version are recorded. The corrected historical trajectory data and the updated inspection point coordinate information are integrated to generate an inspection map that includes global environmental features, robot trajectory, and train dynamic features.

[0096] Specifically, such as Figures 1 to 5As shown, the map reconstruction module further includes, after establishing the dynamic mapping relationship, optimizing the transformation matrix through a random sampling consistency algorithm, eliminating abnormal mapping data caused by marker point recognition errors, and retaining transformation matrix parameters with confidence levels higher than a preset threshold; after generating the inspection map, calculating the matching degree between the corrected historical trajectory and the static features in the global 3D point cloud map; if the matching degree is lower than a preset threshold, re-establishing the coordinate mapping relationship and subsequent steps are performed until the matching degree is higher than the preset threshold. The matching degree is calculated by the mean Euclidean distance of the overlapping area of ​​the point cloud.

[0097] The map reconstruction module first calls the pose calibration marker to resolve two types of key coordinates: one is the absolute coordinates of the marker point in the global 3D point cloud map coordinate system. These coordinates are continuously updated by the 3D LiDAR module through global positioning, reflecting the fixed position of the marker point in the environment; the other is the relative coordinates of the same marker point in the train's local feature coordinate system. These coordinates are determined based on wheel axle feature recognition and are dynamically adjusted as the train's position changes.

[0098] To establish a mathematical relationship between the two coordinate systems, the module solves for the transformation matrix using the least squares method. The least squares method calculates the optimal translation and rotation parameters by minimizing the sum of squared deviations of the marker point's coordinates in the two coordinate systems. The transformation matrix, composed of these parameters, enables the mutual transformation of coordinates between any points in the two coordinate systems. For example, when there is a deviation between the marker point's coordinates in the global coordinate system and its coordinates in the train's local coordinate system, the least squares method iteratively optimizes the transformation matrix parameters until the sum of the deviations is minimized, ensuring transformation accuracy. Based on this transformation matrix, the module establishes a dynamic mapping relationship between the global 3D point cloud map coordinate system and the train's local feature coordinate system. This relationship is updated in real time as the train's stopping position is fine-tuned—whenever new pose calibration marker data is input, the module recalculates the transformation matrix parameters to ensure that the coordinate mapping always matches the actual scene.

[0099] The module extracts the robot's historical positioning trajectory data from the global 3D point cloud map. This data records the robot's complete travel path from the starting point to the current position and the corresponding pose parameters. To eliminate pose errors accumulated over long-term travel, the module superimposes the six-degree-of-freedom correction values ​​obtained from the secondary correction onto the historical poses at the corresponding time points in reverse order, accurately calculating the correction amount for each historical pose.

[0100] The correction allocation employs a differentiated strategy: For translation corrections along the X, Y, and Z axes, linear interpolation is used to allocate them proportionally to the travel distance of historical trajectory segments. For example, if a historical trajectory segment accounts for 20% of the total trajectory, then that segment is allocated 20% of the translation correction, ensuring that the correction is evenly distributed along the trajectory and avoiding over- or under-correction in certain areas. For rotation corrections for heading, pitch, and roll angles, spherical linear interpolation is used to allocate them proportionally over time. This method smoothly transitions rotation parameters in three-dimensional spherical space, avoiding trajectory stitching distortion caused by abrupt changes in rotation angles and ensuring the attitude continuity of historical trajectories. Through reverse overlay and differentiated allocation, the pose parameters at each moment in the historical trajectory are corrected to match the actual travel path, eliminating the impact of accumulated errors on map accuracy.

[0101] Based on the relative coordinate relationship between the robot and the inspection points on the train, the module uses a transformation matrix to convert the local coordinates of the inspection points into absolute coordinates in the global 3D point cloud map coordinate system. The relative coordinate relationship reflects the spatial distance and orientation between the inspection point and the robot's current position. The transformation matrix can map the inspection point from the train's local coordinate system to the global coordinate system, enabling precise positioning of the inspection point in the environment.

[0102] After the conversion, the module compares the newly obtained absolute coordinates with the original coordinates of the detected item point stored in the map. If the spatial deviation exceeds a preset threshold, it means that the original coordinates can no longer reflect the actual position of the detected item point, and the module updates the map data with the converted coordinates; if the deviation is within the threshold range, the original coordinates are kept unchanged to reduce redundant calculations. To ensure the traceability of map data, the module synchronously records the timestamp of the coordinate update and the corresponding pose calibration mark version—the timestamp marks the specific time the update occurred, and the version number is associated with the corresponding transformation matrix parameters, facilitating subsequent tracing of the basis for map data updates and ensuring data reliability. Finally, the module integrates the corrected historical trajectory data with the updated detected item point coordinate information to generate an inspection map that includes global environmental static features, the robot's complete driving trajectory, and train dynamic features.

[0103] After establishing the dynamic mapping relationship, the module optimizes the transformation matrix using a random sampling consensus algorithm. This algorithm calculates the corresponding transformation matrix by repeatedly sampling coordinates of a subset of marker points and checking the matching degree between the remaining marker points and the matrix. The matrix parameters that cover the most marker points are selected as the valid solution. This process eliminates abnormal mapping data caused by marker point identification errors, such as coordinate misjudgments due to partial occlusion, and retains transformation matrix parameters with a confidence level higher than a preset threshold, further improving the stability of the coordinate transformation.

[0104] After generating the inspection map, the module calculates the matching degree between the corrected historical trajectory and the static features in the global 3D point cloud map. The matching degree is calculated using the average Euclidean distance of the overlapping point cloud regions: selecting static feature point clouds (such as track edges and fixed equipment) within a certain range around the historical trajectory, the average spatial distance between the trajectory points and the corresponding static feature points is calculated; the smaller the distance, the higher the matching degree. If the matching degree is lower than a preset threshold, it indicates a deviation in the historical trajectory correction or coordinate transformation. The module then backtracks to the coordinate mapping relationship establishment step, recalculates the transformation matrix, and executes subsequent correction processes until the matching degree is higher than the preset threshold. This closed-loop verification mechanism ensures that the final generated inspection map highly matches the actual environment, providing reliable map support for inspection tasks.

[0105] Through a coherent process of solving the coordinate transformation matrix, accurately correcting historical trajectories, removing abnormal data, and verifying matching degree, the map reconstruction module effectively eliminates the influence of coordinate system deviation, cumulative errors, and environmental interference, enabling the inspection map to have both the stability of the global environment and the dynamic adaptability of train characteristics.

[0106] Specifically, such as Figures 1 to 5 As shown, the map construction method of the system includes the following steps: collecting environmental data through closed-loop path scanning, constructing an initial global 3D point cloud map and performing rasterization and noise reduction preprocessing, and setting the initial positioning value of the robot; based on the initial positioning value, fusing inertial measurement data, mileage data and lidar matching results, outputting the optimal pose of the robot through Kalman filtering, controlling the robot to autonomously navigate to the target track, the photoelectric positioning module predicts the parking deviation and adjusts the trajectory in advance during the robot's movement, and after arriving at the target track, the position of the train head is detected by the cross-type photoelectric sensor to complete the correction of the parking deviation; capturing train wheel axle images, calculating the pose deviation by comparing the wheel axle images with a standard template, and outputting the pose correction result including six-degree-of-freedom correction values, relative coordinate relationships and pose calibration marks; establishing coordinate mapping based on the pose correction result, correcting the global pose error and updating the coordinates of the detected items, and completing the dynamic reconstruction and storage of the inspection map.

[0107] The initial steps of map construction are environmental data acquisition and initial map generation. The system controls the robot to travel along a pre-set closed-loop path to complete environmental data acquisition. The closed-loop path is designed to start from the starting point, cover the target inspection area, and return to the starting point, forming a closed trajectory. This path can reduce the cumulative error of long-distance travel through consistency verification of the data at the beginning and end. During travel, the 3D LiDAR module simultaneously collects three types of core data: 3D point cloud data to capture the spatial morphology of static environmental features, inertial measurement data to record the changes in the robot's acceleration and angular velocity, and mileage data to quantify the travel distance and direction.

[0108] Based on the collected multi-frame data, the system constructs an initial global 3D point cloud map. During the construction process, coordinate transformation is used to unify the multi-frame point cloud data into the same coordinate system, and feature matching is used to eliminate inter-frame deviations, forming a complete environmental space model. To improve map usability, the system performs rasterization and denoising preprocessing on the initial map: rasterization divides the 3D space into regular grids, aggregating point cloud data by grid unit to reduce redundancy and accelerate subsequent localization matching; denoising removes isolated anomalies caused by reflections and dust through multi-frame comparison and feature verification, retaining stable environmental features. After preprocessing, the system uses the starting point of the closed-loop path as a reference, combined with surrounding fixed features (such as track edges and right angles of walls) to set the robot's initial positioning value as the origin reference of the map coordinate system.

[0109] Based on the initial positioning values, the system enters the autonomous navigation phase. The data processing module integrates three types of data to optimize positioning accuracy: inertial measurement data reflects the robot's dynamic posture in real time, odometry data records the incremental trajectory, and the LiDAR matching results provide an absolute position reference by aligning the features of the current frame with the preprocessed map frame. These three data are fused using a Kalman filter algorithm: the filtering algorithm uses the LiDAR matching results as high-precision observations to calibrate the cumulative drift of the inertial measurement and odometry data, outputting a continuous and stable optimal robot pose, providing accurate position and attitude references for navigation.

[0110] Based on the optimal pose, the system controls the robot to autonomously navigate along the planned path to the target track. During the journey, the photoelectric positioning module activates a parking deviation prediction mechanism: cross-arranged photoelectric sensors scan the train's front area in real time, combining this with a pre-set front edge feature model to identify the train's position. Simultaneously, it integrates the current speed, remaining distance, and error patterns from similar historical conditions to predict potential deviations upon reaching the theoretical stopping point. Based on the predicted deviation, the system adjusts the robot's trajectory in advance, compensating for potential offsets by fine-tuning the drive wheel speeds. Upon reaching the target track, the photoelectric sensors perform a second detection of the deviation between the actual and theoretical front positions, calculating the final compensation amount based on the predicted deviation. A slight displacement adjustment completes the parking deviation correction, ensuring the robot stops at the pre-set detection starting position.

[0111] After parking correction, the system initiates a second fine-tuning of the posture. The upward-facing camera captures images of the wheel axles of each train car at preset intervals. During the capture process, adaptive light adjustment ensures clear imaging of key features such as wheel axle edges and hubs. The data processing module extracts features from the wheel axle images, identifies the wheel axle contour boundaries and core feature points, and then uses perspective projection transformation to restore the image coordinates to three-dimensional spatial coordinates, obtaining the actual position of the wheel axle.

[0112] The system compares the actual position of the wheel axle with a pre-set standard wheel axle template. The standard wheel axle template is constructed based on train design parameters and includes the theoretical position and characteristic geometric relationships of the wheel axle in the standard carriage coordinate system. The actual position is aligned with the template position through coordinate mapping, and the spatial deviation between the two is calculated. This deviation directly quantifies the robot's pose offset relative to the current carriage. Based on the deviation analysis, the system outputs pose correction results: six-degree-of-freedom correction values ​​clearly indicate the adjustment amount in translation and rotation directions; the relative coordinate relationship between the robot and the detection points marks the spatial orientation of key detection parts; and pose calibration markers record the association parameters between the global map and the train's local coordinate system, providing a coordinate reference for subsequent map updates.

[0113] Based on the pose correction results, the system enters the map dynamic optimization stage. The map reconstruction module first establishes a coordinate mapping through pose calibration markers: it analyzes the dual coordinates of the marker points in the global map coordinate system and the train's local coordinate system, solves the transformation matrix using the least squares method, and achieves a precise association between the two coordinate systems. Moreover, this mapping relationship is updated in real time as the train's position is finely adjusted.

[0114] Subsequently, the system corrects the global pose error: the six-degree-of-freedom correction values ​​are superimposed onto the robot's historical positioning trajectory in reverse order over time; translational corrections are allocated through linear interpolation, and rotational errors are compensated through spherical linear interpolation, eliminating accumulated deviations from long-term travel. Simultaneously, the feature update submodule, in conjunction with a transformation matrix, converts the relative coordinates of the robot and the detected points into absolute coordinates in the global map coordinate system. This is compared to the original stored coordinates; if the deviation exceeds the limit, the data is updated, and a timestamp and version label are recorded. Finally, the system integrates the corrected historical trajectory, the updated coordinates of the detected points, and global environmental features to generate a dynamically optimized inspection map and complete data storage, providing accurate map support for subsequent inspection tasks.

[0115] Through the above steps, the system achieves a closed-loop process from environmental perception to dynamic map optimization, ensuring the global integrity of the initial map while adapting to dynamic changes in the scenario through real-time correction and updates, thus meeting the high-precision positioning requirements of rail transit vehicle undercarriage inspection.

[0116] The foregoing has illustrated and described the basic features, principles, and advantages of the present invention. It should be noted that the present invention is not limited to the above embodiments, but only to some embodiments. Any improvements and additions made without departing from the spirit and scope of the present invention are considered to be within the scope of protection of the present invention.

Claims

1. A patrol map construction system based on radar and photoelectric technology, characterized in that, The application relates to a robot positioning system, which comprises the following parts: a three-dimensional laser radar module for collecting three-dimensional point cloud data, inertial measurement data and mileage data, constructing a global three-dimensional point cloud map according to the collected data and outputting an initial pose of the robot; a photoelectric positioning module, which comprises cross-arranged photoelectric sensors and an upward-looking camera, the photoelectric sensors are used for detecting the position of a train head and correcting the parking deviation of the robot, and the upward-looking camera is used for acquiring train wheel shaft images; a data processing module, which is in communication connection with the photoelectric positioning module and the three-dimensional laser radar module, pre-processes the three-dimensional point cloud map, combines the inertial measurement data, the mileage data and the initial pose, optimizes the robot optimal pose through Kalman filtering and outputs real-time positioning results, calculates the position deviation according to the matching results of the wheel shaft images and a standard wheel shaft template, and carries out secondary correction on the pose of the robot relative to the train according to the position deviation to obtain a pose correction result; a map reconstruction module, which carries out reconstruction of an inspection map according to the three-dimensional point cloud map and the pose correction result.

2. The radar and electro-optical based inspection mapping system of claim 1, wherein, The three-dimensional laser radar module comprises a map collection and construction sub-module, which is used for collecting three-dimensional laser radar data, inertial measurement data and mileage data through a closed loop path scanning of the robot, generating a three-dimensional point cloud map composed of multiple frames of moving data, carrying out rasterization processing and denoising point processing on the three-dimensional point cloud map and setting an initial positioning value of the robot relative to a three-dimensional point cloud map coordinate system.

3. The radar and electro-optical based inspection mapping system of claim 2, wherein, The data processing module comprises a data collection and matching sub-module, which is used for obtaining an initial pose of the robot according to the initial positioning value through the inertial measurement data and the mileage data, matching a current frame of laser radar with a pre-processed frame of three-dimensional point cloud map and outputting a six-degree-of-freedom pose of the robot, the six-degree-of-freedom pose comprises an X-axis position, a Y-axis position, a Z-axis position, a heading angle, a pitch angle and a roll angle, and the six-degree-of-freedom pose is taken as an observation value to carry out data fusion with the inertial measurement data and the mileage data, and the optimal pose of the robot to the inspection target position is output.

4. The radar and electro-optical based inspection mapping system of claim 1, wherein, The secondary correction of the pose in the data processing module comprises the following steps: the upward-looking camera is used for shooting wheel shaft images of each carriage of the train, wheel shaft features are extracted and the actual position of the wheel shaft is analyzed, the actual position of the wheel shaft is compared with a preset standard wheel shaft template position, the wheel shaft position deviation is calculated, the wheel shaft position deviation is the pose deviation of the robot relative to the current carriage, the pose of the robot relative to the train is adjusted according to the wheel shaft position deviation, and a secondary corrected pose correction result is output, the pose correction result comprises a six-degree-of-freedom correction value of the robot relative to the current carriage, a relative coordinate relationship between the robot and a train detection point and a pose calibration mark for map reconstruction, the pose calibration mark is used for associating the coordinate mapping relationship between the global three-dimensional point cloud map and the local features of the train, and is output to the map reconstruction module.

5. The radar and electro-optical based inspection map building system of claim 4, wherein, The map reconstruction module comprises a coordinate mapping sub-module, which is used for establishing a dynamic mapping relationship between a global three-dimensional point cloud map coordinate system and a local feature coordinate system of the train based on the pose calibration mark. The pose correction sub-module is configured to superimpose the six-degree-of-freedom correction value reversely to a historical positioning trajectory of the robot in the global three-dimensional point cloud map, and correct cumulative errors of the robot pose in the global map. The feature updating sub-module is configured to update coordinate information of the train detection points in the global three-dimensional point cloud map according to a relative coordinate relationship between the robot and the train detection points in combination with the dynamic mapping relationship, and complete dynamic optimization and reconstruction of the inspection map.

6. The radar and electro-optical based inspection mapping system of claim 1, wherein, The photoelectric positioning module corrects the parking deviation, including that, during driving of the robot to the preset theoretical parking point based on the optimal pose, the photoelectric sensors arranged in cross are used to scan a train head region in real time, a train head position is identified through a preset train head edge feature model, a parking deviation generated when reaching the theoretical parking point is predicted based on a current driving speed, a distance to the theoretical parking point, and historical driving error data, a driving trajectory of the robot is adjusted according to the predicted parking deviation, and after reaching the theoretical parking point, a deviation between an actual train head position and a theoretical position is detected again through the photoelectric sensors, and the predicted parking deviation is compensated and adjusted to correct the parking deviation.

7. The radar and electro-optical based inspection map building system of claim 5, wherein, The map reconstruction module further includes that, the marker point coordinates in the global three-dimensional point cloud map coordinate system and the marker point coordinates in the train local feature coordinate system are parsed by calling the pose calibration marker, a conversion matrix of the two coordinate systems is solved by using a least square method, and a dynamic mapping relationship is established, which is updated in real time with the change of the train position. The historical positioning trajectory data of the robot in the global three-dimensional point cloud map are extracted, the six-degree-of-freedom correction value is reversely superimposed to the historical pose at the corresponding time, and the correction amount of each historical pose is calculated, wherein the X-axis, Y-axis and Z-axis correction amounts are distributed to the historical trajectory segment by using a linear interpolation method, and the heading angle, pitch angle and roll angle correction amounts are used to compensate the rotation error by using a spherical linear interpolation method; according to the relative coordinate relationship between the robot and the train detection points, the local coordinates of the detection points are converted into absolute coordinates in the global three-dimensional point cloud map coordinate system in combination with the conversion matrix, the converted coordinates are compared with the originally stored detection point coordinates in the map, if the deviation exceeds a preset threshold, the converted coordinates are used to update the map data, and the coordinate update time stamp and the corresponding pose calibration marker version are recorded. The corrected historical trajectory data and the updated detection point coordinate information are integrated to generate an inspection map including global environmental features, a robot driving trajectory and train dynamic features.

8. The radar and electro-optical based patrolling map construction system of claim 7, wherein, The map reconstruction module further includes that, after the dynamic mapping relationship is established, the conversion matrix is optimized by using a random sample consensus algorithm, abnormal mapping data caused by marker point recognition errors are removed, and the conversion matrix parameters with a confidence higher than a preset threshold are retained; after the inspection map is generated, the matching degree of the corrected historical trajectory and the static features in the global three-dimensional point cloud map is calculated, if the matching degree is lower than a preset threshold, the coordinate mapping relationship establishment and the subsequent steps are performed again until the matching degree is higher than the preset threshold, and the matching degree is calculated by using the Euclidean distance mean of the point cloud overlapping area.

9. The radar and electro-optical based inspection mapping system of claim 1, wherein, The steps of the map construction method of the system include collecting environment data through a closed-loop path scan, constructing an initial global three-dimensional point cloud map and performing griding, denoising and pretreatment, and setting an initial positioning value of the robot; based on the initial positioning value, fusing inertial measurement data, mileage data and laser radar matching results, outputting an optimal pose of the robot through Kalman filtering, controlling the robot to autonomously navigate to a target lane, predicting a parking deviation in the process of robot driving and adjusting a trajectory in advance through an optoelectronic positioning module, detecting a vehicle head position through a cross-type optoelectronic sensor after the robot reaches the target lane, and completing correction of the parking deviation; shooting a train axle image, calculating a pose deviation through comparison between the axle image and a standard template, and outputting a pose correction result containing a six-degree-of-freedom correction value, a relative coordinate relationship and a pose calibration mark; based on the pose correction result, establishing coordinate mapping, correcting a global pose error and updating a detection point coordinate, and completing dynamic reconstruction and storage of an inspection map.

Images