Bridge bottom plate detection method and system based on crawling robot

Abstract

The application discloses a bridge bottom plate detection method and system based on a crawling robot, and relates to the technical field of bridge detection. The method comprises the following steps: collecting bridge bottom plate images and robot motion data, and determining a current detection area; a first detection area generates historical disease data through an improved SeaFormer model; disease feature points in an overlapping area are screened, and a theoretical position is generated; an actual position is matched, and a position correction amount is generated; a detection position is corrected according to the position correction amount, and current disease data are generated; the current disease data are mapped to a bridge bottom plate coordinate, and the historical disease data are updated; corresponding diseases of subsequent detection areas are connected, and a bridge bottom plate disease detection result is generated. The application realizes continuous detection of bridge bottom plate diseases, and improves disease positioning accuracy and recognition integrity.

Description

Technical Field

[0001] This invention relates to the field of bridge inspection technology, and in particular to a method and system for inspecting bridge base plates based on a crawling robot. Background Technology

[0002] Bridge foundation slabs are subjected to vehicle loads, environmental erosion, and structural deformation over long periods, making them prone to cracks, spalling, water seepage, and exposed rebar. Current inspection methods primarily involve manual inspections, vehicle-mounted inspection equipment, or crawling robots equipped with cameras to capture images of the bridge foundation slabs. Image recognition methods are then used to extract affected areas, enabling the recording of defect locations and condition assessments.

[0003] When a crawling robot moves continuously along the bridge deck, wheel slippage, posture changes, and cumulative displacement errors can cause image position shifts. Simply relying on robot motion data to determine the location of defects can easily lead to localization errors. Existing image detection methods mostly process individual detection areas independently, lacking a mechanism to use fixed defect features in adjacent detection areas to correct the detection position. The same defect across detection areas is prone to duplicate recording, positional misalignment, and broken extension relationships, affecting the continuous detection and spatial positioning accuracy of bridge deck defects.

[0004] Therefore, how to provide a bridge under-slab inspection method and system based on crawling robots is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a bridge under-slab inspection method and system based on a crawling robot. This invention employs an improved SeaFormer model and combines fixed defect features with reverse correction of the detection position to complete the identification, localization, and continuous connection of bridge under-slab defects. It has the advantages of high positioning accuracy, good defect continuity, and high detection efficiency.

[0006] A bridge under-slab inspection method based on a crawling robot according to an embodiment of the present invention includes the following steps:

[0007] Collect images of the bridge floor and robot motion data to determine the initial detection position, and map the bridge floor image to the bridge floor coordinates to form the current detection area;

[0008] The first detection area is input into the improved SeaFormer model to extract the disease area and skeletonize it. Endpoints, bifurcation points and turning points are selected as disease feature points to determine the location and extension direction of the disease and generate historical disease data.

[0009] Calculate the intersection of the coordinates of the current detection area and the previous detection area, filter the disease feature points within the intersection from historical disease data, and map them to the theoretical position of the current image;

[0010] Extract current image features, match disease feature points in the neighborhood of theoretical position to form actual position, calculate displacement difference, determine common translation and rotation amount based on several sets of displacement differences, and generate position correction amount;

[0011] The initial detection position is corrected based on the position correction amount, the reverse spatial transformation is performed on the current image features, the correction features are arranged according to the longitudinal and transverse bridge directions, the axial attention calculation is performed, and the current disease data is generated.

[0012] Based on the corrected detection location, the current defect data is mapped to the bridge bottom plate coordinates, historical defect data is written, and the current detection area is used as the previous detection area.

[0013] The system performs repeated defect matching, location correction, and defect data updates in subsequent detection areas. It connects corresponding defects based on the distance between defect endpoints and the direction of extension, generating bridge slab defect detection results.

[0014] Optionally, the process of generating the current detection region includes:

[0015] Control the crawling robot to move along the bridge floor, collect images of the bridge floor at each sampling position, and simultaneously record the crawling robot's walking displacement and posture from the previous sampling position to the current sampling position;

[0016] Starting from the position when the crawling robot enters the detection range, the coordinates of the bridge base plate are established along the longitudinal and transverse directions. Based on the walking displacement and posture corresponding to the continuous sampling positions, the walking displacement of each segment is converted into longitudinal and transverse displacements and accumulated to determine the initial detection position corresponding to the current sampling position.

[0017] Read the imaging parameters of the image acquisition device and the installation pose of the image acquisition device relative to the crawling robot, and determine the spatial transformation relationship between the bridge bottom plate image and the bridge bottom plate coordinates based on the pose of the current sampling position;

[0018] The pixel positions of the bridge underside image are transformed to bridge underside coordinates according to the spatial transformation relationship, and the current detection area is determined according to the bridge underside coordinates corresponding to the boundary of the bridge underside image.

[0019] Optionally, the improved SeaFormer model includes an image feature extraction unit, a lesion location matching unit, a common displacement extraction unit, a feature location correction unit, a bridge axial attention unit, and a lesion data generation unit.

[0020] Optionally, generating historical disease data includes the following steps:

[0021] The bridge under-slab image of the first detection area is classified by pixel, and consecutive defect pixels are connected to form defect areas;

[0022] Iterative refinement is performed on the diseased area, preserving the connectivity of the diseased area, and compressing the diseased area into a disease skeleton with a width of one pixel;

[0023] According to the eight-neighbor relationship, the number of adjacent skeleton pixels of each skeleton pixel in the disease skeleton is counted. Skeleton pixels with only one adjacent skeleton pixel are determined as endpoints, and skeleton pixels with no less than three adjacent skeleton pixels are determined as bifurcation points.

[0024] Arrange skeleton pixels along the continuous direction of the disease skeleton, calculate the direction change angle between adjacent skeleton segments, determine the skeleton pixels whose direction change angle reaches the preset turning angle threshold as turning points, and combine the endpoints, bifurcation points and turning points to generate disease feature points.

[0025] Record the pixel coordinates of each defect feature point in the bridge bottom plate image to generate the defect location. Divide the defect skeleton segments according to the skeleton connection relationship between the defect feature points. Determine the corresponding defect extension direction according to the difference between the first and last pixel coordinates of each defect skeleton segment.

[0026] By associating the diseased areas, disease characteristic points, disease locations, and disease extension directions according to their corresponding relationships, historical disease data is generated.

[0027] Optionally, the process of mapping to the theoretical location of the current image includes:

[0028] The coordinate intersection is determined based on the longitudinal and transverse coordinate ranges of the current detection area and the previous detection area, and an overlapping area is generated.

[0029] Read the bridge slab coordinates corresponding to each defect feature point in the historical defect data, and determine the defect feature points whose bridge slab coordinates are located in the overlapping area as the screening results.

[0030] Based on the initial detection location and the spatial transformation relationship between the bridge underslab image and the bridge underslab coordinates, the bridge underslab coordinates corresponding to the screening results are reverse-transformed to the pixel coordinates of the current bridge underslab image to generate the theoretical positions corresponding to each defect feature point.

[0031] Optionally, the process of generating the position correction amount includes:

[0032] Convolutional mapping and feature encoding are performed on the current bridge underside image to generate current image features, and search neighborhoods are established with each theoretical location as the center.

[0033] Pixel classification and skeletonization are performed on the current image features corresponding to each search neighborhood, and candidate disease feature points are extracted according to the determination method of endpoints, bifurcation points and inflection points.

[0034] Read the feature type and disease extension direction of the disease feature points corresponding to each theoretical position, filter candidate disease feature points with the same feature type from the corresponding search neighborhood, compare the direction difference of the disease extension direction, and determine the pixel coordinates of the candidate disease feature point with the smallest direction difference as the actual position.

[0035] Calculate the displacement difference of each disease characteristic point according to the coordinate difference between the theoretical position and the corresponding actual position, and establish several sets of correspondences between theoretical positions and actual positions;

[0036] The correspondence is fitted with a two-dimensional rigid transformation to find the common translation and rotation that minimize the overall coordinate deviation when the theoretical position is transformed to the actual position. The common translation and rotation are then combined to generate the position correction.

[0037] Optionally, the process of generating the current disease data includes:

[0038] Update the spatial transformation relationship corresponding to the current bridge underside image based on the corrected detection location;

[0039] Based on the spatial transformation relationship, the pixel coordinates of the disease feature points in the current disease data are converted into the coordinates of the bridge bottom plate to generate coordinate-based disease data.

[0040] Calculate the location distance and extension direction difference between the disease area in the coordinated disease data and the disease area in the historical disease data, and determine the disease area whose location distance is less than a preset distance threshold and whose extension direction difference is less than a preset direction threshold as the associated disease area.

[0041] The historical disease data corresponding to the associated disease area is merged with the coordinated disease data, and the disease feature points, disease location and disease extension direction are updated.

[0042] Coordinated disease data that did not match associated disease areas were written into the historical disease data as new disease areas, generating updated historical disease data;

[0043] Record the coordinate range of the bridge base plate corresponding to the current detection area, and update the current detection area to the previous detection area.

[0044] Optionally, the process of generating bridge underslab defect detection results includes:

[0045] Extract the disease feature points corresponding to the associated disease area and the current disease area, and determine the duplicate disease feature points and newly added disease feature points according to the bridge bottom plate coordinates;

[0046] For duplicate disease feature points, only one record is retained. Newly added disease feature points are added to the associated disease area, and the set of disease feature points corresponding to the associated disease area is updated.

[0047] The boundaries and locations of disease areas are redefined based on the updated disease feature points and the skeletal connections between them.

[0048] The direction of disease extension is determined based on the updated disease skeleton segment, and the updated disease area, disease feature points, disease location, and disease extension direction are written into the historical disease data.

[0049] A bridge under-slab inspection system based on a crawling robot according to an embodiment of the present invention includes the following modules:

[0050] The acquisition module is used to acquire images of the bridge base plate and robot motion data to determine the initial detection position and the current detection area;

[0051] The disease extraction module is used to extract disease areas, disease feature points, disease locations and extension directions by improving the SeaFormer model, and generate historical disease data.

[0052] The location mapping module is used to calculate the intersection of coordinates between the current detection area and the previous detection area, filter disease feature points, and generate theoretical locations;

[0053] The correction amount generation module is used to match the theoretical position with the actual position, determine the common translation and rotation amounts based on multiple sets of displacement differences, and generate the position correction amount;

[0054] The disease generation module is used to correct the initial detection position and the current image features, perform bridge axial attention calculation, and generate the current disease data;

[0055] The disease update module is used to map the current disease data to the coordinates of the bridge bottom plate and update historical disease data and the previous detection area.

[0056] The results output module is used to connect the corresponding defects in the subsequent detection areas and generate the detection results of bridge bottom slab defects.

[0057] A chip, disposed inside a crawling robot, includes programmable logic circuits and a computer program, which, when running, enables a method for detecting bridge underplates based on a crawling robot.

[0058] The beneficial effects of this invention are:

[0059] (1) The present invention continuously collects images of the bridge base plate and robot motion data by crawling robot, maps the images of the bridge base plate to a unified bridge base plate coordinate system, and uses the defect feature points in adjacent detection areas to reverse correct the detection position, thereby reducing the impact of robot slippage, posture changes and cumulative displacement errors on the spatial positioning of defects and improving the accuracy of determining the location of defects in the bridge base plate.

[0060] (2) This invention extracts disease areas, disease feature points and disease extension directions by improving the SeaFormer model, calculates common translation and rotation based on multiple displacement differences between theoretical and actual positions, and performs reverse spatial transformation and bridge axial attention calculation on the current image features to improve the consistency of disease feature extraction and position correction in continuous detection areas.

[0061] (3) The present invention continuously writes the current defect data into the historical defect data, and connects the corresponding defects in adjacent detection areas according to the bridge bottom plate coordinates, the distance between defect endpoints and the direction of defect extension, thereby reducing the repeated recording and fracture identification of the same defect across detection areas, and realizing the continuous positioning, range updating and overall detection of bridge bottom plate defects. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a flowchart of a bridge under-slab inspection method based on a crawling robot proposed in this invention;

[0064] Figure 2 This is a module connection diagram of a bridge under-slab inspection system based on a crawling robot proposed in this invention;

[0065] Figure 3 This is a structural diagram of a crawling robot used in a bridge under-slab inspection method based on a crawling robot proposed in this invention.

[0066] Figure 4 This is a coordinate projection diagram of the bridge under-slab of a bridge under-slab detection method based on a crawling robot proposed in this invention.

[0067] Figure 5 This is a diagram showing the location of defects in a bridge under-slab detection method based on a crawling robot proposed in this invention.

[0068] Figure 6 This is a two-dimensional rigid fitting diagram of a bridge under-slab detection method based on a crawling robot proposed in this invention;

[0069] Figure 7 This is a diagram showing the position correction results of a bridge under-slab detection method based on a crawling robot proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] Example 1: Reference Figures 1-7 A method for detecting bridge under-slabs based on crawling robots includes the following steps:

[0072] Collect images of the bridge floor and robot motion data, determine the initial detection position based on walking displacement and posture, and map the bridge floor image to the bridge floor coordinates to form the current detection area;

[0073] The first detection area is input into the improved SeaFormer model to extract the disease area and skeletonize it. Endpoints, bifurcation points and turning points are selected as disease feature points to determine the location and extension direction of the disease and generate historical disease data.

[0074] Calculate the intersection of the coordinates of the current detection area and the previous detection area, filter the disease feature points within the intersection from historical disease data, and map them to the theoretical position of the current image;

[0075] Extract current image features, match disease feature points in the neighborhood of theoretical position to form actual position, calculate displacement difference, determine common translation and rotation amount based on several sets of displacement differences, and generate position correction amount;

[0076] The initial detection position is corrected based on the position correction amount, the reverse spatial transformation is performed on the current image features, the correction features are arranged according to the longitudinal and transverse bridge directions, the axial attention calculation is performed, and the current disease data is generated.

[0077] Based on the corrected detection location, the current defect data is mapped to the bridge bottom plate coordinates, historical defect data is written, and the current detection area is used as the previous detection area.

[0078] The system performs repeated defect matching, location correction, and defect data updates in subsequent detection areas. It connects corresponding defects based on the distance between defect endpoints and the direction of extension, generating bridge slab defect detection results.

[0079] In this embodiment, the process of generating the current detection area includes:

[0080] Control the crawling robot to move along the bridge floor, collect images of the bridge floor at each sampling position, and simultaneously record the crawling robot's walking displacement and posture from the previous sampling position to the current sampling position;

[0081] refer to Figure 3In this embodiment, the crawling robot is a detection carrier capable of adhering to the surface of the bridge deck and moving continuously along it. The crawling robot includes a robot body 1, negative pressure adsorption wheel sets 2 and 3, an image acquisition device 4, a drive motor 5, an inertial measurement device 6, a walking displacement detection device 7, and a data processing device 8. The robot body 1 serves as the mounting base for all detection and moving components. The negative pressure adsorption wheel sets 2 and 3 are located at the edges of the robot body 1. After the negative pressure adsorption wheel sets come into contact with the surface of the bridge deck, they form a negative pressure adsorption effect, keeping the robot body 1 attached to the surface of the bridge deck. Under the driving force provided by the drive motor 5, it moves along the bridge deck.

[0082] Negative pressure adsorption wheel sets 2 and 3 can adopt an adsorption wheel structure with a negative pressure chamber and rolling components. The negative pressure chamber forms a negative pressure area lower than the external environmental pressure during operation. The rolling components rotate along the surface of the bridge base plate, enabling the crawling robot to complete longitudinal and transverse bridge movement and adjust its direction of travel while maintaining adsorption. During the movement of the robot body 1 on the bridge base plate surface, the negative pressure adsorption wheel sets continuously maintain adsorption contact with the bridge base plate surface, reducing the possibility of the robot detaching from the bridge base plate due to its own weight.

[0083] Image acquisition device 4 is fixed to robot body 1 by a mounting bracket. Image acquisition device 4 uses an area array camera to acquire images of the bridge underside corresponding to the robot's current position. The installation position and angle of image acquisition device 4 are calibrated before the start of detection, and the relative installation relationship between image acquisition device 4 and robot body 1 remains unchanged during continuous detection. Image acquisition device 4 acquires images at preset sampling intervals, obtaining one image of the bridge underside each time. The corresponding sampling time is recorded by data processing device 8.

[0084] An inertial measurement unit 6 is fixed to the robot body 1. The inertial measurement unit 6 can be an inertial measurement unit including a three-axis accelerometer and a three-axis gyroscope, continuously collecting the three-axis acceleration and three-axis angular velocity of the crawling robot. Through attitude calculation, the heading angle, pitch angle, and roll angle corresponding to the current sampling position are obtained. The heading angle represents the rotational state of the crawling robot's travel direction relative to the longitudinal direction of the bridge; the pitch angle represents the tilting state of the robot body 1 along the forward / backward direction relative to a local surface of the bridge base; and the roll angle represents the tilting state of the robot body 1 along the left / right direction relative to a local surface of the bridge base. The heading angle, pitch angle, and roll angle together constitute the attitude.

[0085] The walking displacement detection device 7 is installed at the corresponding position of the robot's motion transmission components, and can use a rotary encoder to detect the rotation of the drive shaft or the suction wheel. The data processing device 8 calculates the walking distance of the crawling robot from the previous sampling position to the current sampling position based on the rotation recorded by the rotary encoder and the effective rolling circumference of the corresponding wheel. The walking distance calculated between two consecutive bridge underslab image acquisitions constitutes the walking displacement of the corresponding sampling interval.

[0086] The data processing device 8 is connected to the image acquisition device 4, the inertial measurement device 6, and the walking displacement detection device 7 respectively. It sets a unified sampling identifier for the bridge bottom plate image, walking displacement, and attitude, and associates and stores the bridge bottom plate image, walking displacement, heading angle, pitch angle, and roll angle corresponding to the same sampling position to form the robot motion data corresponding to the current sampling position.

[0087] In one specific embodiment, the robot's walking distance corresponding to adjacent bridge deck images can be set to 100mm. The image acquisition device 4 acquires 1920×1080 pixel bridge deck images, the inertial measurement device 6 continuously acquires attitude data at a frequency not lower than the image sampling frequency, and the walking displacement detection device 7 accumulates and records the wheel rotation between two image acquisitions. The data processing device 8 reads the corresponding attitude at the time of image acquisition and reads the accumulated walking displacement obtained from the previous image acquisition time to the current image acquisition time, so that each bridge deck image corresponds to a determined walking displacement and attitude.

[0088] Starting from the position when the crawling robot enters the detection range, the coordinates of the bridge base plate are established along the longitudinal and transverse directions. Based on the walking displacement and posture corresponding to the continuous sampling positions, the walking displacement of each segment is converted into longitudinal and transverse displacements and accumulated to determine the initial detection position corresponding to the current sampling position.

[0089] like Figure 4 As shown, the coordinates of the bridge base plate are based on the projection of the robot's center onto the surface of the bridge base plate when the crawling robot enters the area to be inspected. Using the longitudinal extension direction of the bridge as The positive direction of the axis is defined as the direction perpendicular to the longitudinal direction of the bridge underside plane. The positive direction of the axis is defined by the direction perpendicular to the surface of the bridge's bottom slab. The axial direction is used to establish the spatial coordinate system of the bridge base plate. The bridge base plate coordinate system of this invention mainly adopts... shaft and The axis coordinates represent the longitudinal and transverse positions of the defects on the surface of the bridge deck.

[0090] The longitudinal extension direction of the bridge can be determined based on the bridge design axis, the longitudinal baseline in the bridge structural drawing, or the pre-marked longitudinal bridge reference at the starting position of the inspection. After the crawling robot enters the inspection range, the robot's initial travel direction is aligned with the longitudinal bridge reference, and the coordinate origin corresponding to the robot's center is recorded.

[0091] For the walking displacement between the previous sampling position and the current sampling position, the walking displacement is decomposed into along the heading angle corresponding to the current walking direction. Longitudinal bridge displacement along the axial direction and along The transverse displacement of the bridge in the axial direction. If the first The segment's walking displacement is The corresponding heading angle is The longitudinal bridge displacement is according to Confirmed, the transverse bridge displacement is in accordance with Confirmed. The longitudinal bridge displacements obtained from the starting position to the current sampling position are accumulated separately, and the transverse bridge displacements are accumulated separately to obtain the longitudinal and transverse bridge coordinates of the current sampling position in the bridge bottom plate coordinate system. These longitudinal and transverse bridge coordinates together constitute the initial detection position.

[0092] When the posture data indicates that the crawling robot is pitching or rolling, the measured walking displacement is projected onto a local plane of the bridge base plate before being decomposed into longitudinal and transverse directions. This avoids directly using the spatial movement distance as the plane displacement of the bridge base plate when the robot crosses a local protrusion or experiences a short-term tilt.

[0093] The imaging parameters of the image acquisition device and the installation pose of the image acquisition device relative to the crawling robot are read, and the spatial transformation relationship between the bridge bottom plate image and the bridge bottom plate coordinates is determined based on the pose of the current sampling position.

[0094] Read the imaging parameters of the image acquisition device and the installation pose of the image acquisition device relative to the crawling robot, and determine the spatial transformation relationship between the bridge bottom plate image and the bridge bottom plate coordinates based on the pose of the current sampling position;

[0095] The image acquisition device can be an industrial area scan camera. The imaging parameters of the image acquisition device include the horizontal focal length. Vertical focal length x-coordinate of principal point Principal point ordinate And lens distortion parameters. Lens distortion parameters include radial distortion parameters and tangential distortion parameters. Imaging parameters can be obtained by camera calibration using a checkerboard calibration plate with known geometric dimensions before the crawling robot performs bridge underslab inspection. The calibrated imaging parameters are written into the data processing device and repeatedly retrieved when the image acquisition device's installation status remains unchanged.

[0096] The mounting pose of the image acquisition device relative to the crawling robot represents the fixed spatial position of the optical center of the image acquisition device relative to the robot's main body reference point, and the fixed rotational relationship of the optical axis of the image acquisition device relative to the robot's main body coordinate direction. The mounting pose is represented by translation vectors and rotation matrices. The translation vectors record the forward / backward, left / right, and vertical offsets of the optical center of the image acquisition device relative to the robot's main body reference point, while the rotation matrix records the mounting angles of the optical axis of the image acquisition device relative to the robot's forward direction, lateral direction, and normal direction. After the image acquisition device is fixedly installed, the mounting pose is obtained through installation dimension measurement and external calibration, and the mounting pose remains unchanged during continuous detection.

[0097] Based on the initial detection position and posture of the current sampling position, a first rigid body transformation is established between the coordinates of the robot body and the coordinates of the bridge base plate. Based on the installation posture of the image acquisition device relative to the crawling robot, a second rigid body transformation is established between the coordinates of the image acquisition device and the coordinates of the robot body. The first rigid body transformation and the second rigid body transformation are combined in sequence to obtain the rotational and translational relationships between the coordinates of the image acquisition device and the coordinates of the bridge base plate.

[0098] Specifically, the heading angle, pitch angle, and roll angle of the current sampling position are converted into rotation matrices, and the initial detection position is taken as the translation position of the robot body in the coordinates of the bridge base plate. The rotation matrix and translation position corresponding to the robot body are combined with the installation rotation matrix and installation translation vector of the image acquisition device to obtain the spatial pose of the image acquisition device relative to the coordinates of the bridge base plate at the current sampling position.

[0099] Distortion correction is performed on the pixels in the bridge underside image. The corrected pixel coordinates are then normalized to the focal length and principal point coordinates of the image acquisition device, resulting in a spatial ray passing through the corresponding pixel from the optical center of the image acquisition device. Based on the current spatial pose of the image acquisition device, the spatial ray is transformed to bridge underside coordinates, and the intersection point of the spatial ray with a local surface of the bridge underside is determined. The longitudinal and transverse coordinates of the intersection point are used as the bridge underside coordinates of the corresponding pixel position, thus obtaining the spatial transformation relationship between the bridge underside image and the bridge underside coordinates.

[0100] For a detection state where the image acquisition device is basically perpendicular to the bridge deck and the robot's attachment surface is approximately parallel to a local surface of the bridge deck, the local surface of the bridge deck can be considered as... For a local plane with fixed coordinates, the vertical distance from the optical center of the image acquisition device to the robot's attachment surface is determined by the vertical translation in the installation pose. Based on this, the intersection point of the spatial ray corresponding to each pixel and the local plane is calculated. When the robot has a pitch angle or roll angle, the direction of the spatial ray is adjusted according to the rotation matrix corresponding to the current pose before calculating the intersection point.

[0101] The pixel positions of the bridge underside image are transformed to bridge underside coordinates according to the spatial transformation relationship, and the current detection area is determined according to the bridge underside coordinates corresponding to the boundary of the bridge underside image.

[0102] Specifically, spatial transformations are performed on the pixel coordinates corresponding to the top left, top right, bottom right, and bottom left corners of the bridge underside image to obtain the positions of the four boundary points in the bridge underside coordinates; the area of ​​the bridge underside enclosed by the four boundary points is determined as the current detection region. For bridge underside images that form non-rectangular projection boundaries after pose correction, the current detection region is recorded according to the quadrilateral region formed by the four boundary points.

[0103] The current detection area is stored in correspondence with the current sampling position, recording the longitudinal and transverse coordinate ranges of the current detection area. After the crawling robot enters the next sampling position, it generates the next current detection area according to the same process. The subsequent overlapping area is determined by comparing the coordinate ranges of the bridge base corresponding to two consecutive current detection areas. This establishes a continuous positional correspondence between the robot's motion data, the initial detection position, the bridge base image, and the current detection area.

[0104] In this embodiment, the improved SeaFormer model includes an image feature extraction unit, a lesion location matching unit, a common displacement extraction unit, a feature location correction unit, a bridge axial attention unit, and a lesion data generation unit.

[0105] The image feature extraction unit is used to process bridge under-slab images. The unit performs size unification and convolution mapping on the input bridge under-slab image, converting image pixels into feature representations. Following the feature encoding process of the SeaFormer model, it extracts edge, texture, and region morphology information of the bridge under-slab surface, generating current image features corresponding to the spatial location of the bridge under-slab image. The first detection region directly enters the bridge axial attention calculation using the current image features.

[0106] The defect location matching unit receives the current image features, the initial detection location, and defect feature points and defect locations selected from the overlapping area. Based on the initial detection location and the spatial transformation relationship between the bridge underslab coordinates and the bridge underslab image, it transforms the selected defect locations to the current bridge underslab image to obtain the theoretical location corresponding to each defect feature point. With each theoretical location as the center, it extracts candidate feature points in the corresponding neighborhood of the current image features. It matches the candidate feature points with the selected defect feature points according to local features and spatial morphology to determine the actual location corresponding to each theoretical location, thus establishing a one-to-one correspondence between theoretical and actual locations.

[0107] The common displacement extraction unit receives multiple sets of theoretical and actual positions output by the disease location matching unit. It calculates the two-dimensional displacement difference between the theoretical and actual positions of the same disease feature point, transforms these multiple two-dimensional displacement differences to the same coordinate direction, and compares the displacement direction and amount. Disease feature points with consistent spatial variation are selected, and a two-dimensional rigid transformation fitting is performed based on their corresponding theoretical and actual positions to obtain the translation and rotation amounts that enable overall registration of multiple corresponding positions. These translation and rotation amounts are used as the common displacement deviation. For cases where only a single disease feature point exhibits a significant positional change, and the remaining disease feature points do not have the same spatial variation relationship, the positional change of the single disease feature point is not used to generate the common displacement deviation.

[0108] The feature position correction unit receives the current image features and common displacement deviation, establishes an inverse rigid transformation based on the translation and rotation amounts in the common displacement deviation, remaps each feature position in the current image features according to the inverse rigid transformation, and obtains the corresponding feature value for features that fall into non-integer positions after mapping by interpolation, keeping the number of channels and spatial arrangement size of the current image features unchanged, and generates the current image features after position correction.

[0109] The axial attention unit determines the feature arrangement direction according to the longitudinal and transverse directions in the bridge base plate coordinates. For the first detection area, the current image features are axially compressed along the longitudinal and transverse directions respectively, and the feature association between different positions in the same bridge direction is established according to the axial attention calculation method of the SeaFormer model. For the detection area, the same processing is performed on the current image features after position correction, and the longitudinal and transverse attention calculation results are restored to the original feature space position to generate bridge defect features.

[0110] The defect data generation unit receives bridge-oriented defect features, performs pixel classification on these features, determines the pixel locations belonging to bridge underslab defects, restores defect boundaries, and generates defect regions. It then performs skeletonization processing on the defect regions, compressing continuous defect regions into a centerline that maintains connectivity, and extracts the endpoints, bifurcation points, and turning points where the direction changes as defect feature points. The defect location is determined based on the position of the defect region in the bridge underslab image, and the defect extension direction is determined based on the arrangement direction of adjacent points in the defect skeleton. Historical defect data is generated from the defect region, defect feature points, defect location, and defect extension direction corresponding to the first detected region. Current defect data is generated from the same data. Finally, the defect location corresponding to the current defect data is converted to the bridge underslab location according to the corrected detection location and written into the historical defect data.

[0111] In this embodiment, generating historical disease data includes the following steps:

[0112] By improving the SeaFormer model, pixel classification is performed on the bridge underside image of the first detection area, and consecutive defect pixels are connected to form a defect area.

[0113] The pixel classification process is as follows: The bridge under-slab image of the first detection area is input into the improved SeaFormer model. The bridge-direction defect features are obtained through the image feature extraction unit and the bridge-direction axial attention unit. The bridge-direction defect features are restored to the pixel positions corresponding to the bridge under-slab image. A defect response value is generated for each pixel. The defect response value is normalized using Sigmoid. Pixels whose normalization result is not less than a preset classification threshold are identified as defect pixels, and pixels whose normalization result is less than the preset classification threshold are identified as background pixels. The preset classification threshold can be set to 0.5.

[0114] Continuous defect pixels refer to a group of defect pixels in a bridge underslab image that can be continuously reached from one defect pixel to another according to an eight-neighborhood relationship. The data processing device starts with any unclassified defect pixel and sequentially searches for defect pixels that are horizontally, vertically, and diagonally adjacent to the current defect pixel. The searched adjacent defect pixels are added to the same connected region, and the process continues to search for adjacent defect pixels of newly added defect pixels until there are no more unsearched adjacent defect pixels. The resulting connected region is then defined as a defect region. Thus, defect pixels in different spatial locations that do not have an eight-neighborhood connection form different defect regions.

[0115] Iterative refinement is performed on the diseased area, preserving the connectivity of the diseased area, and compressing the diseased area into a disease skeleton with a width of one pixel;

[0116] Iterative refinement employs the Zhang-Suen refinement algorithm. For each diseased pixel in the diseased region, the eight neighboring pixels surrounding the central diseased pixel are arranged clockwise. The number of diseased pixels among these eight neighboring pixels is counted, as is the number of times a background pixel transforms into a diseased pixel in a clockwise direction. When the number of neighboring diseased pixels is between 2 and 6, and the number of transformations from background to diseased pixels is 1, satisfying the neighborhood retention condition for the current refinement sub-process, the central diseased pixel is marked as a pixel to be deleted. After scanning all pixels in the diseased region, the pixels to be deleted are deleted uniformly, and then the next scan is performed.

[0117] A complete iteration consists of two refinement sub-processes. The first refinement sub-process deletes boundary pixels on one side of the diseased area, and the second refinement sub-process deletes boundary pixels on the opposite side. Each deletion does not sever the pixel connection paths already formed in the original diseased area. The two refinement sub-processes are repeated until there are no more diseased pixels that can be deleted in a consecutive complete iteration. The iteration stops, and the remaining single-pixel width of continuous pixel chains is determined as the disease skeleton. The disease skeleton maintains the main connection relationships, branching relationships, and extension directions of the original diseased area.

[0118] According to the eight-neighbor relationship, the number of adjacent skeleton pixels of each skeleton pixel in the disease skeleton is counted. Skeleton pixels with only one adjacent skeleton pixel are determined as endpoints, and skeleton pixels with no less than three adjacent skeleton pixels are determined as bifurcation points.

[0119] The eight-neighbor relation is established by creating a 3×3 pixel window centered on the skeleton pixel to be determined. Excluding the central skeleton pixel, the eight pixels in the window—top, bottom, left, right, top-left, top-right, bottom-left, and bottom-right—constitute the eight-neighbor region of the central skeleton pixel. Each of these eight positions is checked one by one to determine if it is a skeleton pixel, thus obtaining the number of adjacent skeleton pixels of the central skeleton pixel.

[0120] When the number of adjacent skeleton pixels is 1, it indicates that the disease skeleton terminates at the corresponding position, and the central skeleton pixel is recorded as the endpoint. When the number of adjacent skeleton pixels is 2, the central skeleton pixel is located inside a continuous disease skeleton. When the number of adjacent skeleton pixels is not less than 3, it indicates that at least three skeleton paths are connected at the corresponding position, and the central skeleton pixel is recorded as a bifurcation point. For multiple candidate bifurcation points formed by adjacent skeleton pixels in the same bifurcation region, the candidate bifurcation points connected by their eight neighborhoods are treated as a connected set, and the skeleton pixel closest to the geometric center of the connected set is selected to record the corresponding bifurcation position, thus avoiding the repeated recording of multiple disease feature points in the same actual bifurcation region.

[0121] Arrange skeleton pixels along the continuous direction of the disease skeleton, calculate the direction change angle between adjacent skeleton segments, determine the skeleton pixels whose direction change angle reaches the preset turning angle threshold as turning points, and combine the endpoints, bifurcation points and turning points to generate disease feature points.

[0122] Specifically, arranging skeleton pixels along the continuous direction of the disease skeleton involves: starting from an endpoint or bifurcation point, searching for unvisited adjacent skeleton pixels in the eight-neighborhood, and tracing forward along the adjacent skeleton pixels one by one, using the tracing order as the arrangement order of the skeleton pixels; ending the current path tracing when another endpoint, bifurcation point, or already visited skeleton pixel is reached.

[0123] To determine the inflection point, the skeleton pixels to be judged are... Centered on the skeleton, a preset number of skeleton pixels are selected forward and backward along the skeleton arrangement direction. In one specific embodiment, the following selections are made: One skeleton pixel; the farthest skeleton pixel on the front is denoted as... The furthest skeleton pixel on the rear side is denoted as Establish the front direction vector Rear direction vector The orientation change angle of the corresponding skeleton position is determined based on the angle between two direction vectors; orientation change angle Calculate according to the following formula:

[0124] ;

[0125] When the directions remain basically the same, the angle of change of direction near When the diseased skeleton undergoes significant bending, the angle of change in direction... Increase.

[0126] A preset turning angle threshold is used to distinguish between normal small pixel fluctuations in the skeleton and changes in the actual extension direction of the disease. In one specific implementation, the preset turning angle threshold is set to... When the angle of change of direction is not less than When the corresponding skeleton pixel is determined as the turning point, in order to avoid multiple consecutive adjacent skeleton pixels at the same bending position exceeding the preset turning angle threshold, the skeleton pixel with the largest direction change angle is selected from a group of skeleton pixels that continuously exceed the preset turning angle threshold and recorded as the turning point.

[0127] In the turning point judgment stage, adjacent skeleton segments refer to two local skeleton paths located in front of and behind the skeleton pixel to be judged. They are used to determine the directional change at the skeleton pixel to be judged. Adjacent skeleton segments are different from disease skeleton segments formed later based on disease feature points.

[0128] The obtained endpoints, bifurcation points, and inflection points are summarized according to their respective pixel coordinates within the same disease skeleton. For duplicate points with the same pixel coordinates, only one record is retained. For each retained point, the feature type corresponding to the endpoint, bifurcation point, or inflection point is recorded. The endpoints, bifurcation points, and inflection points after deduplication are collectively used as disease feature points. Specifically, the feature types are divided into three categories:

[0129] Endpoint type: corresponds to the termination position of the disease skeleton, with only 1 adjacent skeleton pixel in the eight-neighbor domain;

[0130] Bifurcation point type: The branch connection position corresponding to the disease skeleton, with no less than 3 adjacent skeleton pixels in the eight-neighbor domain;

[0131] Turning point type: The location where the extension direction of the disease skeleton changes significantly, and the angle of change of direction reaches the preset turning angle threshold.

[0132] Record the pixel coordinates of each defect feature point in the bridge bottom plate image to generate the defect location. Divide the defect skeleton segments according to the skeleton connection relationship between the defect feature points. Determine the corresponding defect extension direction according to the difference between the first and last pixel coordinates of each defect skeleton segment.

[0133] The location of each defect feature point is recorded, including both its horizontal and vertical pixel coordinates in the bridge underside image. Based on the spatial transformation between the bridge underside image and bridge underside coordinates, the pixel coordinates of the defect feature points are converted to their corresponding bridge underside coordinates, and these coordinates are then associated and recorded. Bridge underside coordinates are used to determine if a defect feature point is located in an overlapping area, while the converted pixel position in the current bridge underside image is used when performing defect location matching within the current image features.

[0134] The skeleton connection relationship indicates that there is a connection path between two disease feature points consisting of eight consecutive skeleton pixels in their eight neighborhoods, and the connection path between the two disease feature points does not pass through a third disease feature point. For each disease feature point, starting from the skeleton pixel connected to the eight neighborhood of the disease feature point, the connection is traced along consecutive skeleton pixels until the next disease feature point is reached. The two disease feature points and all the skeleton pixels traversed between the two disease feature points are divided into a disease skeleton segment. When different skeleton paths are connected at the same bifurcation point, the connection is traced along each different skeleton path from the bifurcation point, and each path forms a corresponding disease skeleton segment.

[0135] For a disease skeleton segment, the disease feature point located at the front of the skeleton arrangement is determined as the starting point of the disease skeleton segment, and the disease feature point located at the rear is determined as the ending point of the disease skeleton segment. Let the starting pixel coordinates be... The endpoint pixel coordinates are ; by coordinate difference The direction vector that constitutes the skeleton segment of the defect is used to determine the direction of defect extension. The direction of defect extension can be represented by the angle between the direction vector and the horizontal axis of the bridge bottom plate image. After further converting the start and end points of the defect skeleton segment to the coordinates of the bridge bottom plate, the direction of defect extension in the actual bridge bottom plate can also be represented by the angle between the defect skeleton segment and the longitudinal direction of the bridge.

[0136] In one specific implementation, the coordinates of the starting point of the defective skeleton segment corresponding to the bridge bottom plate are: The coordinates of the bridge bottom slab corresponding to the endpoint are: The corresponding longitudinal bridge coordinate difference is The transverse coordinate difference is The extension direction of the disease skeleton segment is determined based on the two coordinate differences, and the corresponding direction is recorded in the disease extension direction.

[0137] Historical disease data is generated by associating diseased areas, disease feature points, disease locations, and disease extension directions according to their corresponding relationships. Specifically, the association is established based on the connectivity between diseased areas. Each diseased area is assigned a diseased area number, and disease feature points extracted from the skeleton of the same diseased area are recorded under the same diseased area number. For each disease feature point, the corresponding feature type, pixel coordinates, and bridge underslab coordinates are recorded. For each disease skeleton segment, the corresponding disease feature points at both ends of the segment and the disease extension direction are recorded.

[0138] In this embodiment, the process of mapping to the theoretical position of the current image formation includes:

[0139] The coordinate intersection is determined based on the longitudinal and transverse coordinate ranges of the current detection area and the previous detection area, and an overlapping area is generated.

[0140] Read the bridge slab coordinates corresponding to each defect feature point in the historical defect data, and determine the defect feature points whose bridge slab coordinates are located in the overlapping area as the screening results.

[0141] Based on the initial detection location and the spatial transformation relationship between the bridge underslab image and the bridge underslab coordinates, the bridge underslab coordinates corresponding to the screening results are reverse-transformed to the pixel coordinates of the current bridge underslab image to generate the theoretical positions corresponding to each defect feature point.

[0142] In this embodiment, the process of generating the position correction amount includes:

[0143] Convolutional mapping and feature encoding are performed on the current bridge underside image to generate current image features, and search neighborhoods are established with each theoretical location as the center.

[0144] Specifically, the pixel values ​​of the red, green, and blue channels of the current bridge underside image are normalized and then input into the image feature extraction unit of the improved SeaFormer model. A convolution kernel is used to slide on the bridge underside image with a preset stride. The pixel values ​​within the coverage area of ​​the convolution kernel are multiplied and added with the convolution kernel parameters to obtain the channel response corresponding to each pixel position. The channel response is then normalized and nonlinearly activated to form basic image features with spatial position correspondence. A 3×3 convolution kernel with a stride of 1 and an edge padding of 1 can be used to keep the feature positions corresponding before and after the convolution mapping.

[0145] The basic image features obtained from convolution mapping are input into the original feature encoding structure of the SeaFormer model. Local texture, edge morphology, and adjacent region features are encoded according to pixel spatial location. The encoding results of different channels at the same spatial location are combined to form a feature vector. All feature vectors corresponding to all spatial locations are arranged in order of their positions in the bridge underside image to generate the current image features. Each feature position in the current image features can be transformed to a pixel position in the current bridge underside image based on the spatial correspondence during convolution mapping.

[0146] The search neighborhood is established centered on the theoretical location to limit the retrieval range of the corresponding defect feature points in the current bridge underside image. The range of the search neighborhood is determined based on the maximum allowable positional error of the crawler's initial detection position and the actual spatial resolution corresponding to the current bridge underside image. Let the maximum allowable positional error be... The spatial resolution of the current bridge under-slab image is ,in The search radius represents the actual length of the bridge base plate corresponding to one pixel. ; Expand outwards from the theoretical position in both the horizontal and vertical pixel directions. 1 pixel, forming a side length of 1 The square search neighborhood.

[0147] In one specific implementation, the maximum allowable positional error for the initial detection position is set to 10 mm, and the spatial resolution of the current bridge under-slab image is [missing information]. The search radius obtained is 20 pixels, and the corresponding search neighborhood is 41×41 pixels. When the theoretical position is close to the boundary of the bridge bottom plate image, the part that exceeds the image range will be truncated, and the defect feature point matching will only be performed within the effective image range.

[0148] Pixel classification and skeletonization are performed on the current image features corresponding to each search neighborhood, and candidate disease feature points are extracted according to the determination method of endpoints, bifurcation points and inflection points.

[0149] Read the feature type and disease extension direction of the disease feature points corresponding to each theoretical position, filter candidate disease feature points with the same feature type from the corresponding search neighborhood, compare the direction difference of the disease extension direction, and determine the pixel coordinates of the candidate disease feature point with the smallest direction difference as the actual position.

[0150] Consistent feature types mean that endpoints in historical disease data are matched only with endpoints in the current search neighborhood, bifurcation points are matched only with bifurcation points, and inflection points are matched only with inflection points. For cases where multiple candidate disease feature points with consistent feature types exist within the same search neighborhood, the angle difference between the disease extension direction corresponding to each candidate feature point and the disease extension direction corresponding to each historical feature point is calculated, and this angle difference is converted to... to The minimum directional difference range is used to determine the candidate disease feature point with the smallest directional difference as the matching result, and the pixel coordinates corresponding to the matching result are determined as the actual position.

[0151] In one specific implementation, the maximum permissible directional difference can also be set to... When the direction difference is greater than When no corresponding relationship is established, and when there are no candidate disease feature points in the corresponding search neighborhood that satisfy the feature type and direction difference conditions, the current theoretical position is discarded and does not participate in the two-dimensional rigid transformation fitting.

[0152] Calculate the displacement difference of each disease characteristic point according to the coordinate difference between the theoretical position and the corresponding actual position, and establish several sets of correspondences between theoretical positions and actual positions;

[0153] refer to Figure 5 The diagram shows the corresponding locations of defects. The theoretical and actual locations are established in a one-to-one correspondence based on the same defect feature points. Each set of matching points in the diagram represents the theoretical location mapped from historical defect feature points and the actual location detected in the current bridge underside image, respectively. Coordinate centers are calculated for both the theoretical and actual location sets to provide corresponding point data for two-dimensional rigid transformation fitting.

[0154] The correspondence represents a one-to-one correspondence between the theoretical location of the same historical defect feature point in the current bridge underside image and its actual location obtained through re-detection. For the... Each disease characteristic point, its theoretical location is recorded as Record the actual location as Then the first The displacement difference of each disease characteristic point is:

[0155] ;

[0156] Each set of correspondences must record at least the disease feature point number, feature type, theoretical location, and actual location, so that multiple disease feature points can be formed. The corresponding set of points. To avoid the positional error of a single defect feature point directly affecting the robot's position correction, in one specific implementation, only when a set of at least [number missing] points is established... When the groups have valid correspondences, a two-dimensional rigid transformation fitting is performed.

[0157] The correspondence is fitted with a two-dimensional rigid transformation to find the common translation and rotation that minimize the overall coordinate deviation when the theoretical position is transformed to the actual position. The common translation and rotation are then combined to generate the position correction.

[0158] A two-dimensional rigid transformation fitting is performed on the correspondence between multiple sets of theoretical and actual positions. This involves grouping the theoretical and actual positions of the same defect feature point in the current bridge underside image into a set of corresponding points. Based on multiple sets of corresponding points, the in-plane translation and rotation relationships that can jointly describe the set of theoretical positions relative to the set of actual positions are obtained. Both theoretical and actual positions are represented using pixel coordinates in the current bridge underside image, maintaining the same origin, horizontal pixel direction, and vertical pixel direction before performing the fitting.

[0159] When establishing correspondences, a theoretical position corresponds to only one actual position, and an actual position participates in only one set of correspondences. When multiple theoretical positions match the same candidate disease feature point, a unique correspondence is determined based on the disease feature type, disease extension direction, and pixel distance between the theoretical position and the candidate disease feature point. Disease feature points that do not form a unique correspondence do not participate in the two-dimensional rigid transformation fitting.

[0160] Let the first The theoretical location corresponding to each disease characteristic point is: The corresponding actual location is Two-dimensional rigid transformation is expressed as ;

[0161] in, The serial number indicating the correspondence between the characteristic points of the disease; Indicates the first The theoretical location of each disease characteristic point; Indicates the first The actual location of each disease characteristic point; Represents a two-dimensional rotation matrix acting on all theoretical positions; This represents a two-dimensional translation vector acting on all theoretical positions. A two-dimensional rigid transformation only adjusts the overall position and orientation of the set of theoretical positions, without changing the relative distances between different theoretical positions.

[0162] During the fitting process, the coordinate centers of all theoretical positions and all actual positions are calculated separately. The coordinate centers of the theoretical and actual positions are then subtracted from each theoretical position, resulting in two sets of centered positions. This centering process eliminates the overall translational difference between the theoretical and actual position sets, allowing subsequent calculations to derive the common rotational relationships from the relative positional relationships of multiple disease feature points.

[0163] refer to Figure 6The two-dimensional rigid fitting diagram shown uses the theoretical position set and the actual position set as common reference positions after they are centered. The rotation relationship between the theoretical position set and the actual position set is obtained based on the spatial distribution of the corresponding points of each set of centers. Figure 6 The centered point set is used to eliminate the influence of the overall translation, and the two-dimensional rigid transformation fitting is used to determine the common rotation and common translation.

[0164] Each group of theoretically centered positions is paired with its corresponding actual centered positions in a one-to-one correspondence to construct a two-dimensional correlation matrix. Singular value decomposition is then performed on the two-dimensional correlation matrix, and the two-dimensional rotation matrix is ​​obtained from the left and right singular vectors obtained by the singular value decomposition. The obtained two-dimensional rotation matrix is ​​subjected to determinant evaluation to exclude mirror flip results, ensuring that the two-dimensional rotation matrix only represents the rotation within the current bridge underside image plane. The common rotation amount is then determined based on the two-dimensional rotation matrix. The common rotation amount represents the angular change when multiple theoretical disease feature points rotate as a whole to the corresponding actual disease feature point.

[0165] After determining the two-dimensional rotation matrix, the theoretical position coordinate center is rotated according to the two-dimensional rotation matrix. Then, the coordinate difference between the actual position coordinate center and the rotated theoretical position coordinate center is calculated to obtain the common translation amount. ;

[0166] in, The coordinate center representing all theoretical positions; The coordinate center representing all actual locations; This represents a two-dimensional rotation matrix. The common translation is denoted as... , This indicates the amount of horizontal pixel displacement in the current bridge underside image. This indicates the amount of vertical pixel displacement in the current bridge underside image.

[0167] To minimize the overall coordinate deviation, all theoretical positions are transformed using the same two-dimensional rotation matrix and the same common translation, minimizing the sum of the squares of the distances between the transformed theoretical positions and their corresponding actual positions. The overall coordinate deviation is expressed as: ;

[0168] in, This indicates the number of valid correspondences involved in the fitting; This represents the Euclidean distance between two-dimensional pixel coordinates. The obtained two-dimensional rotation matrix and common translation amount make... By obtaining the minimum value, we can obtain the common translation and common rotation that can jointly explain the positional changes of multiple disease feature points.

[0169] After the initial fitting, the fitting residual between each theoretical position and its corresponding actual position after the two-dimensional rigid transformation is calculated. For correspondences where the fitting residual exceeds a preset residual threshold, the correspondence is deleted from the fitting data, and the two-dimensional rigid transformation fitting is re-performed using the retained correspondences. For example, the preset residual threshold is set to... Pixels, with no less than 1 effective correspondence. When grouping, the common translation and common rotation values ​​obtained from refitting are used; the effective correspondence is less than When grouping, the location correction amount is not generated using the disease feature points of the current detection area.

[0170] refer to Figure 7 The position correction diagram shown shows that the theoretical position after two-dimensional rigid transformation and the corresponding actual position tend to coincide in the two-dimensional pixel plane. The corresponding points with large fitting residuals are removed from the fitting data, and the remaining corresponding points are used to recalculate the common translation and common rotation. Figure 7 The position correction results are used to characterize the fitting results of the two-dimensional rigid transformation on the common position deviation of multiple disease feature points. The final common translation and common rotation are used as position correction quantities.

[0171] Two-dimensional rigid transformation fitting addresses the local detection position deviations that still exist after spatial mapping of the bridge underside image. The current detection area overlaps with the previous detection area, and the remaining positional changes between adjacent sampling positions mainly manifest as overall translation and small-angle rotation within the bridge underside image plane. The image projection changes caused by the robot's pitch and roll angles have been processed through attitude and spatial transformation relationships, and significant perspective distortion is not directly used as the object of two-dimensional rigid transformation fitting.

[0172] The final common translation and common rotation amount The system combines and generates position correction values; these values ​​represent the common positional deviation from the theoretical position set to the actual position set. When performing position correction on the current image features, a reverse two-dimensional rigid transformation is established based on the common translation and rotation values, causing the offset current image features to return to their corresponding theoretical structural positions. When correcting the robot's initial detection position, the common translation is converted into longitudinal and transverse bridge displacements based on the spatial transformation relationship between the bridge base image and the bridge base coordinates. The initial detection position and detection direction are compensated in the opposite direction to the detection deviation to generate the corrected detection position.

[0173] In this embodiment, the process of generating the current disease data includes:

[0174] Update the spatial transformation relationship corresponding to the current bridge underside image based on the corrected detection location;

[0175] Specifically, based on the corrected detection location, the position and orientation of the image acquisition device in the bridge floor coordinate system are redefined. The rotation and translation relationships between the image acquisition device coordinate system and the bridge floor coordinate system are updated to obtain the corrected spatial transformation relationship. Using the updated spatial transformation relationship, the pixel coordinates in the current bridge floor image are transformed back to the bridge floor coordinate system, ensuring that the current defect data and historical defect data are in the same bridge floor coordinate system.

[0176] Based on the spatial transformation relationship, the pixel coordinates of the disease feature points in the current disease data are converted into the coordinates of the bridge bottom plate to generate coordinate-based disease data.

[0177] Calculate the location distance and extension direction difference between the disease area in the coordinated disease data and the disease area in the historical disease data, and determine the disease area whose location distance is less than a preset distance threshold and whose extension direction difference is less than a preset direction threshold as the associated disease area.

[0178] A preset distance threshold is used to determine whether two defect areas belong to the same continuous defect within the plane of the bridge deck. The preset distance threshold is determined based on the spatial resolution of the bridge deck image, defect localization error, and robot repetitive detection position error. For example, setting the preset distance threshold to 50mm means that when the minimum distance between corresponding defect feature points of two defect areas is less than 50mm, the two defect areas are considered to meet the spatial proximity condition. A preset direction threshold is used to determine whether the extension directions of two defect areas are consistent. The preset direction threshold is determined based on the defect extension direction detection error. For example, setting the preset direction threshold to 20° means that when the angle difference between corresponding defect extension directions of two defect areas is less than 20°, the two defect areas are considered to meet the directional consistency condition.

[0179] The historical disease data corresponding to the associated disease area is merged with the coordinate-based disease data, and the disease feature points, disease location, and disease extension direction are updated. Specifically, the disease feature points corresponding to the current disease area and the associated disease area are obtained. The disease feature points of the two disease areas are combined according to the coordinates of the bridge bottom plate. Disease feature points with duplicate coordinates are deleted, and newly added disease feature points are retained. The disease area range is redefined based on the combined disease feature points, and the disease location is updated. The disease extension direction is redefined based on the disease extension direction corresponding to the current disease area and the associated disease area. The updated disease area, disease feature points, disease location, and disease extension direction replace the original data corresponding to the associated disease area.

[0180] Coordinated disease data that did not match associated disease areas were written into the historical disease data as new disease areas, generating updated historical disease data;

[0181] Record the coordinate range of the bridge base plate corresponding to the current detection area, and update the current detection area to the previous detection area.

[0182] In this embodiment, the process of generating bridge underslab defect detection results includes:

[0183] Extract the disease feature points corresponding to the associated disease area and the current disease area, and determine the duplicate disease feature points and newly added disease feature points according to the bridge bottom plate coordinates;

[0184] For duplicate disease feature points, only one record is retained. Newly added disease feature points are added to the associated disease area, and the set of disease feature points corresponding to the associated disease area is updated.

[0185] The boundaries and locations of disease areas are redefined based on the updated disease feature points and the skeletal connections between them.

[0186] Specifically, based on the updated skeleton connection relationship between the disease feature points, the disease skeleton segments belonging to the same continuous disease are determined, and the disease pixels corresponding to each disease skeleton segment in the associated disease area and the current disease area are retrieved according to the correspondence of the disease skeleton segments; the disease pixels belonging to the same continuous disease are converted to the unified bridge bottom plate coordinates and then merged, and only one record is retained for disease pixels at overlapping positions. The connected region formed by all retained disease pixels is taken as the updated disease area; the contour is traced along the outermost continuous disease pixels of the updated disease area to obtain the closed outer contour, and the closed outer contour is determined as the boundary of the updated disease area.

[0187] For bifurcation diseases, the bifurcation point is used as the common connection point, and the disease pixels corresponding to each disease skeleton segment connected to the bifurcation point are classified into the same disease area; if there is no skeleton connection relationship between two disease skeleton segments, even if the two disease areas are spatially close, they will not be directly merged into the same disease area.

[0188] The location of the defect is determined based on the updated defect feature points. The pixel coordinates corresponding to the endpoints, bifurcation points, and turning points are converted into bridge base plate coordinates according to the corrected spatial transformation relationship, and the converted bridge base plate coordinates are written into the corresponding defect area. For defect feature points newly added due to the addition of the current defect area, the corresponding bridge base plate coordinates are added. For duplicate defect feature points, only the corrected bridge base plate coordinates are retained, and the updated defect location is formed by the updated bridge base plate coordinates of all defect feature points.

[0189] The direction of disease extension is determined based on the updated disease skeleton segment, and the updated disease area, disease feature points, disease location, and disease extension direction are written into the historical disease data.

[0190] A bridge under-slab inspection system based on a crawling robot according to an embodiment of the present invention includes the following modules:

[0191] The acquisition module is used to acquire images of the bridge base plate and robot motion data to determine the initial detection position and the current detection area;

[0192] The disease extraction module is used to extract disease areas, disease feature points, disease locations and extension directions by improving the SeaFormer model, and generate historical disease data.

[0193] The location mapping module is used to calculate the intersection of coordinates between the current detection area and the previous detection area, filter disease feature points, and generate theoretical locations;

[0194] The correction amount generation module is used to match the theoretical position with the actual position, determine the common translation and rotation amounts based on multiple sets of displacement differences, and generate the position correction amount;

[0195] The disease generation module is used to correct the initial detection position and the current image features, perform bridge axial attention calculation, and generate the current disease data;

[0196] The disease update module is used to map the current disease data to the coordinates of the bridge bottom plate and update historical disease data and the previous detection area.

[0197] The results output module is used to connect the corresponding defects in the subsequent detection areas and generate the detection results of bridge bottom slab defects.

[0198] A chip, disposed inside a crawling robot, includes programmable logic circuits and a computer program, which, when running, enables a method for detecting bridge underplates based on a crawling robot.

[0199] Example 2: To verify the feasibility of this invention in practice, it was applied to the continuous defect detection scenario of a concrete bridge slab. The bridge slab to be inspected is a continuous concrete structure. A section of the slab, approximately 24m long and 4.8m wide laterally, was selected as the verification range, with a detection area of ​​approximately 115.2m². The surface of the bridge slab exhibited longitudinal cracks, transverse cracks, diagonal cracks, localized spalling, and water seepage marks. Some cracks exceeded the coverage of a single image, and the same crack appeared consecutively in multiple detection areas. The surface of the bridge slab also showed localized roughness, joints, and attached contaminants. During the continuous movement of the crawling robot, there was slight wheel slippage and deviation in the direction of travel. When the detection position was determined solely based on the cumulative walking displacement and posture, position accumulation errors would occur as the detection distance increased, causing coordinate misalignment of the same defect in adjacent detection areas. Long cracks were repeatedly recorded or recorded separately in multiple detection areas.

[0200] This implementation uses a negative pressure adsorption crawling robot attached to the surface of the bridge deck. The robot body is equipped with an image acquisition device, an inertial measurement unit, a walking displacement detection device, and a data processing device. The image acquisition device acquires bridge deck images at a resolution of 1920×1080 pixels, with a calibrated average spatial resolution of approximately 0.50 mm / pixel. The robot triggers an image acquisition at 100 mm walking intervals, maintaining a large spatial overlap between adjacent bridge deck images. The inertial measurement unit outputs the heading angle, pitch angle, and roll angle, while the walking displacement detection device records the incremental movement distance between adjacent sampling positions. After completing the entire detection range, a total of 1168 valid bridge deck images were obtained. Of these, 17 images with significant out-of-focus issues were removed, leaving 1151 bridge deck images for subsequent processing.

[0201] When the robot enters the detection range, the coordinate origin is established using the projection of the robot's center onto the bridge deck surface, with the longitudinal direction as the X-axis and the transverse direction as the Y-axis. Based on the walking displacement and heading angle corresponding to each sampling interval, the walking displacement is decomposed into the longitudinal and transverse directions and accumulated to obtain the initial detection position for each sampling location. A spatial transformation relationship is established using the focal length, principal point coordinates, lens distortion parameters, and the installation pose of the image acquisition device relative to the robot body. This transforms the four boundary points of the bridge deck image to the bridge deck coordinates, forming the current detection area. During the detection process, the cumulative deviation of the robot's original position was recorded to increase with the moving distance. When manually measured at the end of the detection path for verification, the maximum planar deviation of the original position reached 31.4 mm, and the maximum directional deviation reached 2.1°, indicating that it is difficult to maintain the consistency of the defect location throughout the continuous detection process using only robot motion data.

[0202] The first detection region does not contain the previous detection region, and the first bridge underside image is directly input into the improved SeaFormer model. The improved SeaFormer model performs convolutional mapping and feature encoding on the bridge underside image to generate current image features. After bridge axial attention processing, the defect pixels are classified. The defect classification threshold is set to 0.5, and consecutive defect pixels in the classification results are connected to form defect regions. The Zhang-Suen thinning method is used to skeletonize the defect regions, converting defect regions with area into single-pixel defect skeletons that maintain connectivity. The endpoints and bifurcation points of the defect skeleton are identified according to the eight-neighbor relationship. The local directional change angle is calculated along the skeleton direction, and skeleton pixels with a directional change angle of 30° are identified as inflection points. The endpoints, bifurcation points, and inflection points are used together as defect feature points. Each defect feature point records the pixel coordinates and bridge underside coordinates, and each defect skeleton segment records the corresponding extension direction, thereby establishing the historical defect data for the first detection region.

[0203] When the robot enters a new detection area, it calculates the intersection of the current and previous detection areas in the bridge slab coordinates. From historical defect data, it filters defect feature points whose bridge slab coordinates fall within this intersection. These selected defect feature points are mapped to the current bridge slab image based on their initial detection positions and spatial transformation relationships, forming theoretical positions. During verification, an average of 7.6 historical defect feature points were selected from adjacent detection areas, with crack endpoints accounting for approximately 31.8%, bifurcation points approximately 18.5%, and turning points approximately 49.7%. Overlapping areas with fewer than three defect feature points do not undergo defect reverse position correction; instead, the initial detection positions determined by the robot's motion data are used.

[0204] In the current bridge underslab image, a search neighborhood is established centered on each theoretical location. The maximum allowable error for the initial detection location is set to 10 mm, the spatial resolution is 0.50 mm / pixel, and the corresponding search radius is set to 20 pixels, forming a 41×41 pixel search neighborhood. In the current image features corresponding to the search neighborhood, defect pixel classification and skeletonization are performed again to extract candidate endpoints, candidate bifurcation points, and candidate turning points. Historical endpoints are matched only with candidate endpoints, historical bifurcation points only with candidate bifurcation points, and historical turning points only with candidate turning points. For candidate defect feature points of the same type, the defect extension direction is compared, with a maximum allowable direction difference set to 20°. The candidate defect feature point with the smallest direction difference that meets the direction condition is selected, and its corresponding pixel coordinates are determined as the actual location.

[0205] This implementation method generated 6842 preliminary correspondences between theoretical and actual positions. After filtering by feature type and extension direction, 6417 valid correspondences were retained, with an initial effective matching rate of 93.8%. Two-dimensional rigid transformation fitting was performed on at least three valid correspondences within each detection area. The fitting process utilized multiple sets of disease feature points to jointly calculate the common translation and rotation amounts of the current detection area relative to historical disease positions. After each round of fitting, the residual between the theoretical position and the actual position after rigid transformation was calculated. Correspondences with residuals exceeding 3 pixels were deleted, and the remaining correspondences were refitted. After residual filtering, 6226 stable correspondences were retained, accounting for 91.0% of the preliminary correspondences. The average coordinate residual between the theoretical and actual positions before fitting was 9.6 pixels, and the average residual after fitting and outlier removal was reduced to 1.8 pixels.

[0206] After obtaining the common translation and common rotation, these are combined to form the position correction value. The position correction value describes the common positional deviation between the theoretical set of defect locations and the current actual set of defect locations. The improved SeaFormer model establishes an inverse two-dimensional rigid transformation based on the position correction value, remapping the positions of the current image features to return them to their corresponding bridge structure positions. The robot's initial detection position is corrected according to the opposite deviation direction, generating the corrected detection position. The corrected current image features are rearranged along the longitudinal and transverse bridge directions, and axial attention calculations are performed to obtain the current defect data. Continuous detection verification shows that the maximum planar deviation of the robot's detection position decreased from 31.4 mm to 8.6 mm, the average positional deviation decreased from 18.7 mm to 6.1 mm, and the maximum directional deviation decreased from 2.1° to 0.7°.

[0207] After the current defect data is generated, the spatial transformation relationship between the current bridge slab image and the bridge slab coordinates is re-established based on the corrected detection location, and defect feature points and defect areas are remapped to a unified bridge slab coordinate system. The current defect area is associated with historical defect areas using spatial location and defect extension direction. The association distance threshold between defect feature points is set to 50mm, and the defect extension direction difference threshold is set to 20°. Defect areas that meet the spatial distance and direction conditions are identified as associated defect areas. Only one record is retained for duplicate defect feature points in associated defect areas, and newly detected endpoints, bifurcation points, and turning points are written to the corresponding historical defect areas. The defect skeleton is reconnected according to the updated connection relationship, defect pixels are merged under the unified bridge slab coordinate system, and overlapping pixels are recorded only once, completing the historical defect data update. The bridge slab coordinate range corresponding to the current detection area is used as the previous detection area for the next sampling position, and the same processing is performed continuously.

[0208] After the inspection, a reference result was established through manual verification. Manual verification used a steel ruler, crack width measuring tools, and image annotation records to confirm the location, length, and connection relationships of the main cracks. A total of 63 cracks with a clear continuous direction were identified within the verification area, of which 27 cracks spanned at least two robot inspection areas, with the longest crack spanning nine inspection areas. Without the location correction and historical defect update processing of this invention, the 63 cracks were identified as 78 independent defect records, including 11 cross-regional cracks with breakage records and 7 defects recorded repeatedly in overlapping areas. After adopting this invention, the 63 cracks formed 66 defect records, with only 2 cross-regional cracks showing partial separation, and the number of repeated records reduced to 2. The continuous identification rate of cross-regional defects increased from 84.1% to 96.8%, and the defect duplication rate decreased from 11.1% to 3.2%.

[0209] In terms of defect image recognition, using the SeaFormer processing results without position correction as a control, the pixel cross-union ratio (CUN) of the defect area was 0.843. After using the improved SeaFormer model of this invention and completing feature position inverse correction, the CUN of the defect area increased to 0.879. The average positioning error of crack endpoints decreased from 6.8 pixels to 2.3 pixels, the average positioning error of bifurcation points decreased from 7.5 pixels to 2.6 pixels, and the average positioning error of turning points decreased from 8.1 pixels to 2.9 pixels. The average structural coordinate error of defect locations within the detection range decreased from 18.7 mm to 6.1 mm. The above data show that fixing bridge defect feature points can provide a stable reference for the positional deviation caused by continuous robot movement. The translation and rotation relationships formed by multiple sets of defect feature points can be used to correct the robot's detection position and improve the correspondence of defect features in different detection areas in a unified bridge base plate coordinate system.

[0210] In terms of detection efficiency, all 1151 bridge under-slab images were processed continuously by the data processing device. The average time for feature extraction per image was approximately 86 ms, for defect location matching and two-dimensional rigid transformation it was approximately 17 ms, for feature inverse spatial transformation and bridge axial attention it was approximately 24 ms, and the average total processing time per image was approximately 127 ms. The average time for updating historical defect data was approximately 8 ms. The entire detection process was comparable to the slow, continuous crawling acquisition process of a robot. The verification results show that this invention can not only identify local defects in bridge under-slabs but also use identified fixed defects to correct subsequent detection positions, continuously recovering the same defect across multiple detection areas. Specific data are shown in Table 1 below.

[0211] Table 1: Verification data on continuous defects in bridge underslabs

[0212] As shown in Table 1, the position correction effect of this invention in continuous inspection of bridge underslabs is quite significant. Without position inverse correction, the average defect position error was 18.7 mm, and the maximum detection position deviation was 31.4 mm. After adopting the method of this invention, the average defect position error decreased to 6.1 mm, and the maximum detection position deviation decreased to 8.6 mm. The maximum directional deviation decreased from 2.1° to 0.7°, and the average residual of the two-dimensional rigid fitting decreased from 9.6 pixels to 1.8 pixels. This indicates that obtaining the common translation and rotation amounts based on the correspondence between theoretical and actual positions can effectively correct the cumulative position deviation formed during the continuous movement of the crawling robot, making the spatial correspondence of defect features in the bridge underslab coordinates more accurate.

[0213] In terms of disease identification and continuous correlation, the crossover ratio of disease areas increased from 0.843 to 0.879, the continuous identification rate of cross-regional diseases increased from 84.1% to 96.8%, the duplicate recording rate of diseases decreased from 11.1% to 3.2%, and the average positioning error of crack endpoints decreased from 6.8 pixels to 2.3 pixels. These data indicate that the present invention not only improves the positioning accuracy of disease feature points within a single detection area but also enhances the continuous connection capability of the same disease between adjacent detection areas. By utilizing historical disease feature points to participate in the position matching of the current detection area and using position correction to modify the current image features and detection position, the identification and duplicate recording of cross-regional crack fractures can be reduced, improving the completeness and spatial consistency of bridge under-slab disease detection results.

[0214] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting bridge under-slabs based on a crawling robot, characterized in that, Includes the following steps: Collect images of the bridge floor and robot motion data to determine the initial detection position, and map the bridge floor image to the bridge floor coordinates to form the current detection area; The first detection area is input into the improved SeaFormer model to extract the disease area and skeletonize it. Endpoints, bifurcation points and turning points are selected as disease feature points to determine the location and extension direction of the disease and generate historical disease data. Calculate the intersection of the coordinates of the current detection area and the previous detection area, filter the disease feature points within the intersection from historical disease data, and map them to the theoretical position of the current image; Extract current image features, match disease feature points in the neighborhood of theoretical position to form actual position, calculate displacement difference, determine common translation and rotation amount based on several sets of displacement differences, and generate position correction amount; The initial detection position is corrected based on the position correction amount, the reverse spatial transformation is performed on the current image features, the correction features are arranged according to the longitudinal and transverse bridge directions, the axial attention calculation is performed, and the current disease data is generated. Based on the corrected detection location, the current defect data is mapped to the bridge bottom plate coordinates, historical defect data is written, and the current detection area is used as the previous detection area. The system performs repeated defect matching, location correction, and defect data updates in subsequent detection areas. It connects corresponding defects based on the distance between defect endpoints and the direction of extension, generating bridge slab defect detection results.

2. The bridge under-slab inspection method based on a crawling robot according to claim 1, characterized in that, The process of generating the current detection region includes: Control the crawling robot to move along the bridge floor, collect images of the bridge floor at each sampling position, and simultaneously record the crawling robot's walking displacement and posture from the previous sampling position to the current sampling position; Starting from the position when the crawling robot enters the detection range, the coordinates of the bridge base plate are established along the longitudinal and transverse directions. Based on the walking displacement and posture corresponding to the continuous sampling positions, the walking displacement of each segment is converted into longitudinal and transverse displacements and accumulated to determine the initial detection position corresponding to the current sampling position. Read the imaging parameters of the image acquisition device and the installation pose of the image acquisition device relative to the crawling robot, and determine the spatial transformation relationship between the bridge bottom plate image and the bridge bottom plate coordinates based on the pose of the current sampling position; The pixel positions of the bridge underside image are transformed to bridge underside coordinates according to the spatial transformation relationship, and the current detection area is determined according to the bridge underside coordinates corresponding to the boundary of the bridge underside image.

3. The bridge under-slab inspection method based on a crawling robot according to claim 2, characterized in that, The improved SeaFormer model includes an image feature extraction unit, a lesion location matching unit, a common displacement extraction unit, a feature location correction unit, a bridge axial attention unit, and a lesion data generation unit.

4. The bridge under-slab inspection method based on a crawling robot according to claim 3, characterized in that, The generation of historical disease data includes the following steps: The bridge under-slab image of the first detection area is classified by pixel, and consecutive defect pixels are connected to form defect areas; Iterative refinement is performed on the diseased area, preserving the connectivity of the diseased area, and compressing the diseased area into a disease skeleton with a width of one pixel; According to the eight-neighbor relationship, the number of adjacent skeleton pixels of each skeleton pixel in the disease skeleton is counted. Skeleton pixels with only one adjacent skeleton pixel are determined as endpoints, and skeleton pixels with no less than three adjacent skeleton pixels are determined as bifurcation points. Arrange skeleton pixels along the continuous direction of the disease skeleton, calculate the direction change angle between adjacent skeleton segments, determine the skeleton pixels whose direction change angle reaches the preset turning angle threshold as turning points, and combine the endpoints, bifurcation points and turning points to generate disease feature points. Record the pixel coordinates of each defect feature point in the bridge bottom plate image to generate the defect location. Divide the defect skeleton segments according to the skeleton connection relationship between the defect feature points. Determine the corresponding defect extension direction according to the difference between the first and last pixel coordinates of each defect skeleton segment. By associating the diseased areas, disease characteristic points, disease locations, and disease extension directions according to their corresponding relationships, historical disease data is generated.

5. The bridge under-slab inspection method based on a crawling robot according to claim 4, characterized in that, The process of mapping to the current image to form a theoretical position includes: The coordinate intersection is determined based on the longitudinal and transverse coordinate ranges of the current detection area and the previous detection area, and an overlapping area is generated. Read the bridge slab coordinates corresponding to each defect feature point in the historical defect data, and determine the defect feature points whose bridge slab coordinates are located in the overlapping area as the screening results. Based on the initial detection location and the spatial transformation relationship between the bridge underslab image and the bridge underslab coordinates, the bridge underslab coordinates corresponding to the screening results are reverse-transformed to the pixel coordinates of the current bridge underslab image to generate the theoretical positions corresponding to each defect feature point.

6. The bridge under-slab inspection method based on a crawling robot according to claim 5, characterized in that, The process of generating the position correction amount includes: Convolutional mapping and feature encoding are performed on the current bridge underside image to generate current image features, and search neighborhoods are established with each theoretical location as the center. Pixel classification and skeletonization are performed on the current image features corresponding to each search neighborhood, and candidate disease feature points are extracted according to the determination method of endpoints, bifurcation points and inflection points. Read the feature type and disease extension direction of the disease feature points corresponding to each theoretical position, filter candidate disease feature points with the same feature type from the corresponding search neighborhood, compare the direction difference of the disease extension direction, and determine the pixel coordinates of the candidate disease feature point with the smallest direction difference as the actual position. Calculate the displacement difference of each disease characteristic point according to the coordinate difference between the theoretical position and the corresponding actual position, and establish several sets of correspondences between theoretical positions and actual positions; The correspondence is fitted with a two-dimensional rigid transformation to find the common translation and rotation that minimize the overall coordinate deviation when the theoretical position is transformed to the actual position. The common translation and rotation are then combined to generate the position correction.

7. The bridge under-slab inspection method based on a crawling robot according to claim 6, characterized in that, The process of generating the current disease data includes: Update the spatial transformation relationship corresponding to the current bridge underside image based on the corrected detection location; Based on the spatial transformation relationship, the pixel coordinates of the disease feature points in the current disease data are converted into the coordinates of the bridge bottom plate to generate coordinate-based disease data. Calculate the location distance and extension direction difference between the disease area in the coordinated disease data and the disease area in the historical disease data, and determine the disease area whose location distance is less than a preset distance threshold and whose extension direction difference is less than a preset direction threshold as the associated disease area. The historical disease data corresponding to the associated disease area is merged with the coordinated disease data, and the disease feature points, disease location and disease extension direction are updated. Coordinated disease data that did not match associated disease areas were written into the historical disease data as new disease areas, generating updated historical disease data; Record the coordinate range of the bridge base plate corresponding to the current detection area, and update the current detection area to the previous detection area.

8. The bridge under-slab inspection method based on a crawling robot according to claim 7, characterized in that, The process of generating bridge underslab defect detection results includes: Extract the disease feature points corresponding to the associated disease area and the current disease area, and determine the duplicate disease feature points and newly added disease feature points according to the bridge bottom plate coordinates; For duplicate disease feature points, only one record is retained. Newly added disease feature points are added to the associated disease area, and the set of disease feature points corresponding to the associated disease area is updated. The boundaries and locations of disease areas are redefined based on the updated disease feature points and the skeletal connections between them. The direction of disease extension is determined based on the updated disease skeleton segment, and the updated disease area, disease feature points, disease location, and disease extension direction are written into the historical disease data.

9. A bridge under-slab inspection system based on a crawling robot, applied to the bridge under-slab inspection method based on a crawling robot as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The acquisition module is used to acquire images of the bridge base plate and robot motion data to determine the initial detection position and the current detection area; The disease extraction module is used to extract disease areas, disease feature points, disease locations and extension directions by improving the SeaFormer model, and generate historical disease data. The location mapping module is used to calculate the intersection of coordinates between the current detection area and the previous detection area, filter disease feature points, and generate theoretical locations; The correction amount generation module is used to match the theoretical position with the actual position, determine the common translation and rotation amounts based on multiple sets of displacement differences, and generate the position correction amount; The disease generation module is used to correct the initial detection position and the current image features, perform bridge axial attention calculation, and generate the current disease data; The disease update module is used to map the current disease data to the coordinates of the bridge bottom plate and update historical disease data and the previous detection area. The results output module is used to connect the corresponding defects in the subsequent detection areas and generate the detection results of bridge bottom slab defects.

10. A chip, characterized in that, The chip is disposed inside the crawling robot. The chip includes a programmable logic circuit and a computer program. When the chip is running, it is used to implement the bridge underplate detection method based on the crawling robot as described in any one of claims 1 to 8.

Images