Dynamic panoramic intelligent detection system and method for variable-height beam bottom diseases of bridge

Abstract

The invention discloses a dynamic panoramic intelligent detection system and method for variable-height beam bottom diseases of a bridge. Variable-cross-section and large-gradient beam bottom global cross-pier walking is realized through a split type electromagnetic adsorption hanging rail detection platform based on a rail; an array camera acquisition module is constructed, laser ranging and steering engine control are combined to dynamically adjust the distance and the inclination angle of a camera, an intelligent light supplementing module is matched to guarantee image acquisition standardization, a high-precision panoramic image is generated through camera joint calibration, distortion correction and SIFT + RANSAC feature matching and self-correction splicing algorithms, multi-thread synchronous processing is carried out by means of an NUC edge calculation module, and a high-precision panoramic image is obtained. Deep learning recognition, stereoscopic vision and a point cloud clustering technology are integrated to complete two-dimensional / three-dimensional feature extraction of diseases, precise positioning and visualization of the diseases are achieved through a GNSS-odometer-camera coordinate fusion positioning technology and butt joint with a BIM system, remote control deployment, automatic cross-pier transposition and remote transmission are supported, and the system is suitable for large-scale popularization and application. The intelligent and visual level of complex bridge detection can be comprehensively improved, and the digital maintenance requirement of the bridge is met.

Description

Technical Field

[0001] This invention belongs to the field of bridge defect detection technology, and more specifically, relates to a dynamic panoramic intelligent detection system and method for bridge bottom defects with varying height. Background Technology

[0002] With the rapid advancement of transportation infrastructure construction in my country, bridges, as key hubs in the transportation network, are receiving increasing attention for their structural safety and long-term service performance. Under the influence of natural environmental erosion, vehicle loads, and material aging, bridge beam bottoms are prone to defects such as cracks, spalling, exposed reinforcement, and honeycomb-like pitting. If these defects are not detected accurately and promptly, they will gradually deteriorate and threaten the structural safety of the bridge. Therefore, automated, efficient, and accurate detection of beam bottom defects has become a core requirement for intelligent bridge operation and maintenance and digital management.

[0003] Traditional bridge girder bottom defect detection mainly relies on manual inspection. Inspectors need to use lifting platforms or bridge inspection vehicles to approach the bottom of the girder, observing the defects using flashlights, binoculars, and other tools, recording data and labeling defect information on-site. This method has significant drawbacks: it relies excessively on the inspector's experience, is highly subjective, and has low inspection efficiency; for girder bottom structures with high piers, large spans, and enclosed spaces, inspection is extremely difficult and carries very high safety risks, with high-altitude operations easily leading to accidents; and the inspection data lacks quantitative and visual support, making it difficult to conduct long-term defect trend analysis.

[0004] In actual engineering, many bridges adopt beam bottom structures with variable cross-sections, large longitudinal slopes, or complex spatial curvatures. Such structures bring more severe challenges to defect detection: First, the large variation in beam bottom height and irregular geometric shape make it difficult for traditional detection platforms (such as tracked robots and ground-based remote control equipment) to move stably along the beam bottom. UAVs are prone to drifting, collisions, or crashes when flying over the corners and narrow gaps of the beam bottom, making it impossible to achieve full-area coverage detection. Second, the tilt of the beam bottom causes image acquisition to be prone to distortion, and the image distribution is discontinuous, making data fusion difficult and making it difficult to construct a complete defect heat map or achieve 3D reconstruction. Third, defect coordinate positioning needs to be combined with the three-dimensional configuration information of the bridge. Traditional photogrammetric projection models are no longer applicable, and point cloud registration and stereo matching technologies are required. Fourth, the joint calibration of sensors is difficult, and standardized image acquisition cannot be achieved, resulting in insufficient accuracy in defect detection and feature extraction.

[0005] Among existing related technologies, some solutions focus on the mechanical structure design of the beam bottom inspection platform, such as using a guide rail walking mechanism or a hydraulically driven rotating structure to achieve cross-pier inspection. However, such platforms are mostly designed for human transport, with low levels of automation and intelligence, and cannot dynamically adjust their movement posture according to the beam bottom structure with variable cross-section and large slope, resulting in insufficient adaptability to complex structures. Another set of solutions uses image processing algorithms as the core for defect identification, but lacks organic integration with the inspection platform, failing to form a complete systematic solution from hardware deployment, data acquisition, signal processing to defect identification, location, and visualization. In addition, existing technologies generally suffer from problems such as lack of multi-dimensional perception, low spatial positioning accuracy of defects, and large image stitching errors, making it difficult to meet the accurate inspection requirements of beam bottom structures with variable height.

[0006] In summary, existing bridge beam bottom defect detection technologies have shortcomings in terms of adaptability to complex structures, degree of automation in detection, standardization of data acquisition, accuracy of defect identification, and reliability of spatial positioning. There is an urgent need for a full-process intelligent detection system that can adapt to beam bottom structures with variable cross sections and large slopes to solve the above-mentioned technical problems. Summary of the Invention

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a dynamic panoramic intelligent detection system and method for bridge beam bottom defects of varying height. Through an innovative design of the detection platform, a split electromagnetic adsorption connection structure between the crossbeam and the ball joint is used. Combined with PID closed-loop control, adaptive attitude adjustment of the beam bottom and blind-spot-free detection across piers are achieved. High-precision, distortion-free panoramic images of the beam bottom are generated using multi-camera joint calibration, Brown-Conrady distortion correction, SIFT+RANSAC feature matching, and multi-view joint optimization techniques. A lightweight deep convolutional neural network is used to achieve two-dimensional defect recognition and three-dimensional measurement. By integrating NUC edge computing units to build a multi-threaded parallel processing architecture, and combining GNSS and wheeled odometer EKF positioning technology to form a dual-coordinate indexing system, the system simultaneously achieves real-time detection, accurate defect identification (0.1mm-level crack identification, 3D quantization accuracy down to the millimeter level), reliable positioning (longitudinal error ≤5cm), and cloud-edge collaborative data management. This effectively solves the problems of traditional beam bottom detection, such as structural limitations, poor imaging quality, insufficient defect quantification, and ambiguous positioning. It can provide full-process, refined data support for bridge maintenance, significantly improving detection efficiency and operation and maintenance economy.

[0008] To achieve the above objectives, one aspect of the present invention provides a dynamic panoramic intelligent detection system for bridge beam bottom defects of varying height, including a detection platform module, a camera array acquisition module, an attitude control module, an intelligent supplementary lighting module, an image processing module, a defect identification and feature extraction module, an edge computing and control module, and a positioning fusion module; The detection platform module includes a main body of the beam bottom detection platform, guide rails respectively arranged on the outside of the bridge piers on both sides of the beam bottom, a steel truss support structure for supporting the main body of the beam bottom detection platform, a transmission device for driving the main body of the beam bottom detection platform to move longitudinally, and a split electromagnetic adsorption connecting beam connected to the main body of the beam bottom detection platform. The camera array acquisition module includes multiple moving cameras arranged laterally at preset intervals; the lateral overlap rate of images acquired by adjacent moving cameras is 20%, and all moving cameras maintain a constant detection distance from the bottom of the beam; the attitude control module includes linear laser rangefinders and single-axis servos mounted on the moving cameras of the camera array acquisition module; the intelligent supplementary lighting module is a ring LED supplementary lighting system with ambient light perception; the image processing module is used to realize joint camera calibration, image distortion correction, and panoramic image stitching of the bottom of the beam; the defect identification and feature extraction module is used to realize defect type identification, two-dimensional morphological parameter calculation, and three-dimensional volume quantization; the edge computing and control module is used to realize multi-threaded synchronous control of camera acquisition, image processing, defect identification, and remote data transmission based on NUC industrial computing units; the positioning fusion module is used to realize precise defect positioning based on GNSS and wheeled odometer fusion positioning technology, combined with the lateral numbering of the camera array to form a dual-coordinate index system; through the efficient collaboration of various modules, the level of intelligence and visualization of complex bridge inspection is comprehensively improved.

[0009] Furthermore, the main body of the beam bottom detection platform is equipped with rollers adapted to the guide rail; the steel truss support structure includes longitudinal support columns on both sides of the bridge beam bottom and a steel truss located on top of the longitudinal support columns on both sides of the bridge beam bottom; the crossbeam is connected to the longitudinal support columns through a ball joint structure. When facing the need for pier inspection, the electromagnetic switch is controlled to disconnect the connection in the middle of the crossbeam, rotate the crossbeam to a collinear state with the columns, and after crossing the pier, restore the original state to continue the inspection, realizing full accessibility of the bridge.

[0010] Furthermore, the transmission device is connected to the edge computing and control module via control signals to achieve precise control of the detection platform's travel speed; the transmission device integrates a positioning sensor, which is a wheel-type odometer, and is connected to the positioning fusion module via communication; the detection platform module's split electromagnetic adsorption connecting beam integrates a GNSS module. The LED supplemental lighting system includes an ambient light sensor and an LED driving circuit; the ambient light sensor is connected to the edge computing and control module via an analog signal to provide real-time brightness data; the LED driving circuit is connected to the edge computing and control module via a PWM signal to receive brightness adjustment commands. The linear laser rangefinder of the attitude control module is connected to the edge computing and control module via analog / digital signals to transmit "camera-beam bottom distance" data in real time; the single-axis servo of the attitude control module is connected to the edge computing and control module via PWM control signals to receive attitude adjustment commands and drive camera angle changes. The image processing module is connected to the camera array acquisition module via a data interface to receive raw image data; the image processing module is connected to the edge computing and control module via a data bus to transmit corrected and stitched panoramic image data. The edge computing and control module serves as the control and data processing hub of the system. It is bidirectionally connected to the detection platform module, camera array acquisition module, attitude control module, intelligent supplementary lighting module, image processing module, disease identification and feature extraction module, and positioning fusion module through control signals or data links, enabling synchronous scheduling, data interaction, and command issuance among multiple modules. The positioning fusion module is connected to the detection platform module via a serial port / network port to receive multi-source positioning raw data; the positioning fusion module is connected to the edge computing and control module via a data bus to output the precise coordinates of the disease in the form of "vertical distance + horizontal number".

[0011] The second aspect of this invention provides a dynamic panoramic intelligent detection method for bridge beam bottom defects of varying height, implemented using the aforementioned dynamic panoramic intelligent detection system for bridge beam bottom defects of varying height, comprising the following steps: S1. Assemble and deploy the guide rail, detection platform module, camera array acquisition module, intelligent supplementary lighting module, and attitude control module; clarify the core system parameters; and calculate the constant detection distance of the camera. and platform speed The intelligent supplementary lighting module was debugged and the multi-camera joint calibration was carried out to establish a three-dimensional geometric model of the beam bottom and a unified coordinate system. S2. The edge computing and control module outputs control commands to the transmission device, driving the detection platform to operate at a preset speed. Moving forward at a constant speed, the attitude control module uses a PID algorithm to adjust the camera attitude to ensure that the imaging surface is parallel to the bottom of the beam. Simultaneously, the camera array is activated to acquire images, and the intelligent supplementary lighting module dynamically supplements the light. At the same time, the odometer and GNSS auxiliary data are acquired and transmitted to the edge computing and control module. S3, the image processing module corrects distortion in the acquired raw images and completes the panoramic image stitching of the beam bottom through feature matching and multi-view joint optimization; the disease identification and feature extraction module uses a pre-trained convolutional neural network to identify diseases and extract two-dimensional morphological parameters, and combines stereo view pairs and point cloud clustering to quantify the three-dimensional volume of the disease. S4, the positioning fusion module integrates GNSS and odometer data, and constructs a dual coordinate system by combining the horizontal numbering of the cameras, to achieve precise positioning of each defect in the coordinate system of the bottom of the structural beam, and maps the defect information to the BIM and GIS systems for visualization; when the detection platform encounters a pier, the edge computing and control module controls the split electromagnetic adsorption beam to complete the crossing of the pier, continuously promotes the detection of the entire bridge, and uploads all data to the cloud to form a closed loop, completing the dynamic panoramic intelligent detection of defects at the bottom of the beam with varying height.

[0012] Furthermore, in step S1, a constant detection distance is maintained between the camera and the bottom of the beam. The calculation formula is: in: The camera is positioned at a constant detection distance from the bottom of the beam. For pixel physical resolution, For camera focal length, This refers to the camera pixel size; Detection platform forward speed The calculation formula is: in, The vertical length of the field of view of adjacent frames; Capture frame rate for camera; The desired overlap rate for the vertical image. .

[0013] Furthermore, the multi-camera joint calibration in step S1 includes: A checkerboard calibration board is placed at a fixed geometric position, and images of the calibration board are simultaneously acquired by all cameras in the array. Using the "Zhang Zhengyou calibration method" combined with the "global bundled adjustment algorithm", the intrinsic parameter matrix and extrinsic parameter matrix of each camera are solved; the intrinsic parameter matrix includes focal length, pixel size, and principal point coordinates; the extrinsic parameter matrix includes rotation matrix and translation vector. Relative calibration relationship: for adjacent cameras and Its extrinsic parameter matrix satisfies: in, It's a camera. The rotation matrix, It's a camera. Translation vector; It's a camera. The rotation matrix, It's a camera. The translation vector; It's a camera. To the camera The relative rotation matrix, It's a camera. To the camera The relative translation vector; Introducing first-order rigid body constraint optimization, the expression for first-order rigid body constraint optimization is: ; in, It refers to the overall rotational attitude of the array. It refers to the overall translational orientation of the array; For the fixed rotational offset of each camera during the calibration phase, This refers to the fixed translation offset of each camera during the calibration phase.

[0014] Further, step S3 includes: S31. Image preprocessing: Radial and tangential distortion correction is performed on the raw images acquired by the camera array based on the Brown-Conrady model; S32. Panoramic image stitching: Using the laser line output by the linear laser rangefinder as the external constraint, the SIFT+RANSAC algorithm is used to extract feature matching point pairs of the overlapping areas of adjacent camera images and adjacent frame images. The coordinate system is unified through homography matrix mapping, and then combined with multi-view joint optimization to eliminate the cumulative stitching error and the brightness jump of the seam, generating a full-coverage, distortion-free panoramic image of the beam bottom. S33. Two-dimensional identification and feature extraction of diseases: The stitched panoramic image is input into a pre-trained convolutional neural network model. The model infers and outputs a mask image of the disease. Then, the two-dimensional contour of the disease is extracted based on the connected component analysis of the mask area. Combined with the "pixel-object size mapping relationship" of the camera calibration in the early stage, the two-dimensional morphological parameters of the disease are calculated. The two-dimensional morphological parameters of the disease include the length, width and area of ​​the disease. S34. Three-dimensional quantitative analysis of defects: Using the overlapping areas of adjacent camera images or adjacent frame images as stereo pairs, stereo reconstruction is performed using the camera intrinsic and extrinsic parameter matrices to generate three-dimensional point cloud data of the beam bottom; the DBSCAN clustering algorithm is used to segment the three-dimensional point cloud, removing background and noise points, and the reference plane of the beam bottom and the abnormal change area of ​​the defect are separated by plane fitting; the three-dimensional quantitative parameters of the defect are calculated by using voxel reconstruction or triangular mesh construction algorithms for the abnormal point cloud clusters corresponding to the defect, realizing the complete analysis of the defect from two-dimensional identification to three-dimensional quantification; the three-dimensional quantitative parameters of the defect include the depth and volume of the defect.

[0015] Further, step S32 includes: Using the markings formed on the bottom of the beam by the linear laser rangefinder as a reference, a feature matching algorithm is used to extract matching point pairs within the overlapping area. and ; The homography matrix model based on perspective projection is used to unify the images from multiple cameras to the same plane coordinate system, thus completing the initial stitching. When the bottom of the beam is approximately a plane, the correspondence between the two camera images can be expressed as: ; in, It is a proportionality coefficient. It's a camera. To the camera The homography matrix; For camera A feature point in the image; For camera A feature point in the image; camera To the camera The formula for calculating the homography matrix is: ; in, It's a camera. The intrinsic parameter matrix; It's a camera. The intrinsic parameter matrix; It's a camera. The extrinsic parameter rotation matrix; It's a camera. The extrinsic parameter rotation matrix; It's a camera. The extrinsic parameter translation vector; It is the normal vector of the local plane at the bottom of the beam; The distance from the bottom plane of the beam to the origin of the coordinate system.

[0016] Furthermore, step S32 introduces multi-view joint optimization; the multi-view joint optimization includes geometric error optimization and illumination consistency optimization; The objective function for geometric error optimization is: ; in, Geometric errors in image stitching; The homography matrix between all cameras or views that need to be optimized simultaneously; It's a camera. To the camera The homography matrix; The illumination consistency optimization is as follows: while ensuring geometric alignment, the influence of uneven illumination from multiple cameras at the bottom of the beam is considered, and an optical flow consistency constraint is introduced to ensure continuous brightness changes between adjacent images. The objective function for the joint optimization of geometry and illumination is: ; in, The homography matrix between all cameras or views that need to be optimized simultaneously; These are the weighting coefficients; For feature points In the camera Pixel brightness values ​​in the image; For feature points In the camera Pixel brightness values ​​in the image; The squared error of brightness of the same physical point in different views.

[0017] Further, step S4 includes: S41. Multi-source positioning data fusion and defect coordinate calculation: The extended Kalman filter algorithm is used to fuse GNSS positioning data with displacement data from the wheel odometer integrated with the transmission device to obtain high-precision longitudinal positioning results of the detection platform; combined with the lateral numbering of the camera array, a dual coordinate indexing system of "longitudinal distance + lateral numbering" is constructed to assign a unique spatial coordinate to each identified defect and clarify the specific location of the defect at the bottom of the beam. S42. Integration and visualization of disease information with BIM / GIS system: The spatial coordinates, two-dimensional morphological parameters, three-dimensional quantitative data, original images, and panoramic image information of the disease are mapped to the bridge BIM three-dimensional model through the API interface to achieve accurate labeling of the disease in the model; at the same time, it is integrated with the GIS geographic information system to associate the geographical location and station information of the bridge, forming an integrated display of "spatial location-disease characteristics-visualized model". S43. Cross-pier inspection connection and full-area coverage: When the inspection platform moves to the pier position, the NUC edge computing unit identifies the pier position through positioning data and triggers the cross-pier control command: controls the electromagnetic switch of the split electromagnetic adsorption beam to open, drives the beam to rotate around the ball joint structure to a collinear state with the longitudinal support column, avoiding interference with the pier; after the inspection platform crosses the pier along the guide rail through the transmission device, controls the electromagnetic switch to close, restores the original connection state of the beam, and continues to advance at the preset speed and inspection process to ensure that there are no blind spots at the bottom of the beam and achieve full bridge coverage; S44. Data Upload and Full-Process Closed Loop Formation: During the inspection process, the collected raw data, processing results, positioning information, and BIM / GIS fused data are transmitted to the cloud database in real time; after the inspection is completed, a full bridge bottom defect inspection report is generated, forming a full-process closed loop of "deployment-collection-processing-identification-positioning-storage-reporting".

[0018] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) The dynamic panoramic intelligent detection system and method for bridge beam bottom defects with variable height of the present invention has strong adaptability to the whole area and can break through the limitations of detection scenarios. It adopts a split electromagnetic adsorption connection between the crossbeam and the ball hinge structure, and with the double-sided guide rail constraint, it can realize the crossbeam disconnection and rotation to cross the pier when crossing the pier, solve the problem of fracture in the detection area of ​​traditional equipment, and achieve full-area detection of the beam bottom without blind spots. At the same time, it dynamically adjusts the camera posture and distance through PID closed-loop control, accurately adapts to the beam bottom structure with variable cross section and large slope, ensures stable attachment and safe operation of the equipment, and can adapt to the complex detection scenarios of different types of bridges.

[0019] (2) The dynamic panoramic intelligent detection system and method for bridge beam bottom defects of the present invention has excellent imaging quality and can ensure the basis for defect identification; the Zhang Zhengyou method combined with global binding adjustment is used to realize the joint calibration of multiple cameras, establish a unified coordinate system, and significantly reduce the camera installation offset error; the Brown-Conrady model is combined to correct radial and tangential distortion, making the imaging geometric relationship linear and ensuring the true geometric proportion of features such as cracks and detachments; combined with homography matrix calculation, SIFT+RANSAC feature matching and multi-view joint optimization (geometric error + photometric consistency constraint), the cumulative splicing error and brightness jump of splicing seam are effectively eliminated, generating a distortion-free and uniformly bright panoramic image of the beam bottom; the brightness is automatically adjusted by controlling the current through the ring LED dynamic supplementary lighting system, which can realize the brightness consistency of multi-camera images, eliminate light interference, and the image splicing error is ≤1 pixel, providing high-quality image data support for the identification of 0.1mm-level micro cracks.

[0020] (3) The dynamic panoramic intelligent detection system and method for bridge beam bottom defects of the present invention is accurate and efficient, and realizes full-dimensional quantification of defects; relying on a lightweight deep convolutional neural network, it automatically identifies various defects such as cracks and spalling, with an accuracy rate of ≥95%; through camera calibration parameters, it completes pixel-to-physical size conversion, accurately calculates indicators such as defect area, length, and shape factor, and provides basic data for quantitative analysis of defects; using the overlapping area of ​​adjacent cameras (or adjacent frames) as a binocular pair, it obtains three-dimensional structural data of defects through disparity calculation and point cloud reconstruction, and realizes mixed The depth and volume of defects such as concrete spalling and honeycomb pitting are quantified; based on DBSCAN clustering and depth anomaly threshold, planar points and anomaly points are separated, and the volume of defects is calculated with the voxel reconstruction algorithm with millimeter-level accuracy; by fusing GNSS and wheel odometer data through EKF, smooth and continuous dynamic positioning can still be achieved in GNSS obstruction scenarios, realizing high-precision longitudinal displacement measurement of the detection platform; a dual coordinate system of "longitudinal distance + lateral number" is constructed to achieve accurate positioning of each defect in the coordinate system at the bottom of the structural beam, with a defect positioning error of ≤5cm.

[0021] (4) The dynamic panoramic intelligent detection system and method for bridge beam bottom defects of the present invention is intelligent and collaborative, which can improve the value of operation and maintenance management; with NUC edge computing unit as the core, it realizes real-time processing of image acquisition, stitching, recognition and transmission through multi-threaded parallel architecture, with end-to-end latency ≤300ms; adopts cloud-edge collaborative mode of "local caching + core upload" to reduce bandwidth consumption and cloud dependence, and supports offline autonomy and fault self-healing; the defect data can be integrated with BIM / GIS system to realize visual management, forming a closed loop of "detection-analysis-storage-management" process, which significantly improves the scientific nature of bridge maintenance decision-making and operation and maintenance economy. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a dynamic panoramic intelligent detection method for bridge beam bottom defects of varying height according to an embodiment of the present invention. Figure 2 This is a structural schematic diagram of a dynamic panoramic intelligent detection system for bridge beam bottom defects of varying height according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] like Figure 1 As shown, one aspect of the present invention provides a dynamic panoramic intelligent detection system for bridge beam bottom defects of varying height, including a detection platform module, a camera array acquisition module, an attitude control module, an intelligent supplementary lighting module, an image processing module, a defect identification and feature extraction module, an edge computing and control module, and a positioning fusion module; the detection platform module includes a beam bottom detection platform body, guide rails respectively arranged on the outside of the bridge piers on both sides of the bridge beam bottom, a steel truss support structure for supporting the beam bottom detection platform body, a transmission device for driving the longitudinal movement of the beam bottom detection platform body, and a system connected to the beam bottom detection platform body. The system comprises a split-type electromagnetic adsorption connecting crossbeam; the main body of the beam bottom detection platform is equipped with rollers adapted to the guide rail; the steel truss support structure includes longitudinal support columns on both sides of the bridge beam bottom and a steel truss located on top of the longitudinal support columns on both sides of the bridge beam bottom; the crossbeam is connected to the longitudinal support columns through a ball joint structure; when facing the need for pier inspection, the electromagnetic switch is controlled to disconnect the connection in the middle of the crossbeam, rotate the crossbeam to a collinear state with the columns, and after crossing the pier, return to the original state to continue the inspection, achieving full accessibility of the bridge; the split-type electromagnetic adsorption connecting crossbeam of the detection platform module integrates a GNSS module. The camera array acquisition module includes multiple moving cameras arranged laterally at preset intervals; the lateral overlap rate of images acquired by adjacent moving cameras is 20%, and all moving cameras maintain a constant detection distance from the bottom of the beam; the attitude control module includes a linear laser rangefinder and a single-axis servo motor mounted on the moving cameras of the camera array acquisition module. The linear laser rangefinder is used to sense the real-time distance between the camera and the bottom of the beam, and the single-axis servo motor is used to adjust the camera angle so that the camera's optical imaging surface is parallel to the bottom surface of the beam; the intelligent supplementary lighting module is a ring LED supplementary lighting system with ambient light sensing, which maintains the brightness consistency of the images acquired by multiple cameras by dynamically adjusting the LED array current; the LED supplementary lighting system includes an ambient light sensor and an LED driving circuit; the intelligent supplementary lighting module has waterproof, dustproof, and ambient brightness self-compensation characteristics, and can operate stably at night, in backlight, and in shadow conditions; the image processing module includes a camera joint calibration submodule, an image distortion correction submodule, and a beam bottom panoramic image stitching submodule, used for practical... The system includes camera joint calibration, image distortion correction, and panoramic image stitching of the beam bottom. The defect identification and feature extraction module comprises a two-dimensional defect identification and segmentation submodule, a two-dimensional defect morphological parameter calculation submodule, and a three-dimensional defect quantification analysis submodule, used to achieve defect type identification, two-dimensional morphological parameter calculation, and three-dimensional volume quantization. The edge computing and control module includes an equipment linkage control submodule, a multi-source data preprocessing submodule, a localized data processing submodule, a multi-source data fusion and positioning calibration submodule, and an edge-cloud collaborative interaction submodule, used to achieve multi-threaded synchronous control of camera acquisition, image processing, defect identification, and remote data transmission based on the NUC industrial computing unit. The positioning fusion module includes a multi-source positioning data receiving and preprocessing submodule, a multi-source data fusion calculation submodule, and a defect coordinate construction and cross-system mapping submodule, used to achieve accurate defect positioning based on GNSS and wheeled odometer fusion positioning technology, combined with the horizontal numbering of the camera array to form a dual-coordinate index system. The modules of this invention are interconnected, with the edge computing and control module as the central hub and physical load, data transmission, and control commands as the links, forming a closed-loop system encompassing acquisition, control, processing, analysis, positioning, and visualization. Through the efficient collaboration of these modules, a comprehensive improvement in the intelligence and visualization capabilities of complex bridge inspection is achieved.

[0025] Furthermore, the camera joint calibration submodule is used to solve for the intrinsic parameters (focal length, pixel size, principal point coordinates, distortion coefficients, etc.) and extrinsic parameters (rotation matrix, translation vector) of the camera array using Zhang Zhengyou's method combined with a global binding adjustment algorithm; it establishes a unified coordinate system for all cameras in the array, optimizes and solidifies the relative attitude between cameras through rigid body constraints, and avoids repeated calibration during the detection process; it outputs a standardized calibration parameter set (including intrinsic parameter matrix, extrinsic parameter matrix, and distortion coefficients) to provide data input for distortion correction and homography matrix calculation; the calibration error is controlled within ≤0.1 pixels. The image distortion correction submodule is used to eliminate hardware imaging defects, restore the true geometric shape of the image, and provide qualified input for feature matching and stitching; its core functions include: Based on the distortion parameters output by the camera joint calibration submodule, and using the Brown–Conrady classical distortion model, the original image is subjected to bidirectional error correction of radial distortion and tangential distortion. Combined with the beam bottom plane fitting data fed back by the attitude control module, the viewing angle deviation caused by the variable height beam bottom is finely adjusted to ensure that the image proportions of different slope areas are consistent and to avoid the measurement error of "nearer is larger and farther is smaller" in the appearance of defects. The beam bottom panoramic image stitching submodule is used to generate a blind-spot-free, high-precision panoramic image of the beam bottom through multi-view optimization technology; its core functions include: Feature matching and coordinate unification: Using the laser line of the linear laser rangefinder as the external constraint, the SIFT+RANSAC algorithm is used to extract the matching point pairs of overlapping areas of adjacent cameras (horizontal) and adjacent frames (vertical), and coordinate mapping is completed through homography matrix; SIFT algorithm: Scale-Invariant Feature Transform; RANSAC algorithm: Random Sample Consensus. Multi-view joint optimization: Construct a dual objective function of "geometric error + illumination consistency" to eliminate cumulative stitching error (≤1 pixel) and seam brightness abrupt changes (brightness fluctuation ≤±5%). Full-domain panoramic generation: Integrates the connecting images from the cross-pier detection to generate a distortion-free panoramic image covering the entire bottom of the beam, supporting coordinate indexing of "horizontal number + vertical distance"; Furthermore, the two-dimensional defect recognition and segmentation submodule is used to receive the distortion-free panoramic image of the bridge bottom output by the image processing module, input a pre-trained deep convolutional neural network model (supporting multiple model adaptations such as Mask-R-CNN, U-Net, and YOLOv8); automatically identify various common bridge bottom defects such as cracks, concrete spalling, honeycomb pitting, and corrosion, and output a binary mask map corresponding to each defect (the defect area is marked as the foreground and the background area is marked as the background); based on the connected component analysis of the mask map, noise areas with an area smaller than a preset threshold (such as 0.1 mm², adapting to the 0.1 mm crack recognition requirement) are removed to optimize the integrity of the defect outline; The disease 2D morphology parameter calculation submodule is used to call the "pixel-physical size mapping relationship" output by the camera joint calibration submodule to convert the pixel coordinates of the disease mask into real physical coordinates; extract the geometric features of the disease 2D contour and calculate the core parameters; generate a list of disease 2D parameters, associate the corresponding panoramic image area with coordinate information, and provide a position index for 3D quantization; The core parameters include: Crack-related defects: Length (cumulative distance along the crack centerline), maximum width / average width (statistical distance perpendicular to the crack centerline), and orientation angle (angle with the longitudinal direction of the beam bottom). Peeling / honeycomb-like surface defects: area, shape factor, boundary perimeter; The 3D quantitative analysis submodule for disease is used to generate 3D point cloud data of the beam bottom using the overlapping areas of adjacent cameras (horizontal) or adjacent frames (vertical) as stereo pairs, combined with the camera intrinsic and extrinsic parameter matrices, through a stereo reconstruction algorithm. Based on the positional information in the 2D disease identification results, it accurately extracts a subset of the point cloud corresponding to the disease area, and uses the DBSCAN clustering algorithm to segment the point cloud data, removing background and noise points. It constructs a reference plane for the beam bottom through a plane fitting algorithm, calculates the vertical distance (i.e., disease depth) between the point cloud of the disease area and the reference plane, and calculates the 3D volume of the disease by combining voxel reconstruction or triangular mesh construction algorithms. It outputs the 3D parameters of the disease (maximum depth, average depth, volume), and associates them with the 2D parameters and spatial coordinates to form a complete disease feature dataset. Furthermore, the edge computing and control module uses a NUC edge computing unit as its core hardware carrier, integrating multi-protocol communication interfaces (such as Ethernet, 5G, and CAN bus), an AI acceleration chip (supporting lightweight model inference such as TensorFlow Lite), and a real-time control system, deployed locally on the beam bottom detection platform. Parallel processing of image acquisition, stitching, recognition, and transmission is achieved through multi-threaded pipeline synchronous control, ensuring real-time and stable data flow. A GPU-accelerated inference framework is used for defect identification at the edge, resulting in low processing latency, fast response speed, and improved real-time detection performance. The equipment linkage control submodule is used for: Motion control of the beam bottom inspection platform: Control commands are output to the transmission device via CAN bus to drive the beam bottom inspection platform to move at a preset speed at a constant speed; speed feedback data from the wheel odometer integrated on the transmission device is collected in real time, and the transmission power is dynamically adjusted based on the incremental PID algorithm to control the speed error within ±0.05m / s and ensure that the longitudinal overlap rate of the image is stable at 20%; Camera attitude and distance closed-loop control: Real-time reception of distance data from linear laser rangefinder and attitude parameters (pitch angle, roll angle) from attitude sensor; calculation of distance control deviation and tilt angle error through PID fusion algorithm; output of target control angle to single-axis servo motor; dynamic adjustment of camera position and attitude; maintenance of constant distance between camera and beam bottom and parallel relationship of imaging surface; response delay ≤50ms. Intelligent control for pier connection: By fusing GNSS positioning data and odometer displacement data, the pier position is accurately identified; the pier control logic is triggered, and the electromagnetic adsorption device is sent with the command "electromagnetic disconnection - crossbeam rotation - electromagnetic closure", which controls the split crossbeam to complete the attitude conversion and reset, realizing pier detection without blind spots. The entire connection process takes ≤30 seconds. Dynamic adjustment of the supplementary lighting system: It receives brightness data from the ambient light sensor and dynamically adjusts the array current of the ring LED intelligent supplementary lighting module through PWM signal based on the preset brightness reference value to ensure the brightness consistency of images acquired by multiple cameras, with a brightness fluctuation range of ≤±10%; The multi-source data preprocessing submodule is used to synchronously receive raw images, laser ranging data, GNSS positioning data, attitude sensor data, and other multi-source information acquired by the camera array acquisition module. It removes sensor outliers based on the 3σ criterion and corrects the timestamp deviation of the multi-source data based on the camera trigger signal. After removing sensor outliers and correcting data timestamp deviations, the data is uniformly packaged with a processing efficiency of ≥1000 data entries / second. Distributed caching technology is used to temporarily store the preprocessed data and dynamically schedule it to subsequent modules according to the "acquisition-processing-analysis" process to avoid data loss and blockage. The localized data processing submodule is used for: Image preprocessing and panoramic stitching: The Brown-Conrady distortion correction algorithm is used to correct radial and tangential distortion in the original images from a single camera; the matching point pairs between adjacent cameras / adjacent frames are extracted based on the SIFT+RANSAC feature matching algorithm, coordinate unification is achieved through homography matrix mapping, and then distortion-free panoramic images are generated through multi-view joint optimization, with stitching error ≤ 1 pixel; Lightweight Intelligent Disease Recognition: Deploys a pre-trained lightweight convolutional neural network model, uses an NPU acceleration chip to perform edge-side inference on panoramic images, and outputs mask images and 2D contours of diseases such as cracks, spalling, and honeycomb surface in real time; combines camera calibration parameters to complete pixel-to-object size conversion, calculates parameters such as disease length and area, with recognition latency ≤200ms / frame and accuracy ≥95%; Rapid 3D data processing: Local stereo reconstruction is performed on stereo view images to generate beam bottom point cloud data; noise points and background points are segmented by DBSCAN clustering algorithm, disease areas are separated based on plane fitting, and the volume and depth of disease are calculated by voxel reconstruction algorithm to achieve 3D quantization of disease, with a single frame point cloud processing efficiency of ≥500,000 points / second. The multi-source data fusion and positioning calibration submodule is used for: The extended Kalman filter (EKF) algorithm is used to fuse GNSS positioning data with wheeled odometer displacement data (high-precision longitudinal displacement) to output longitudinal positioning results of the detection platform with an error ≤5cm. Combined with the lateral numbering of the camera array, a dual coordinate system of "longitudinal distance + lateral numbering" is constructed to assign a unique spatial coordinate to each defect, thereby completing the accurate location of the defect. The edge-cloud collaborative interaction submodule is used for: Data hierarchical transmission: Large-capacity data such as original images and point cloud data are stored locally via SSD, while only core data such as disease parameters (two-dimensional / three-dimensional features), positioning information, panoramic images, and equipment operating status are transmitted to the cloud database via 5G / Ethernet encryption, reducing bandwidth usage by more than 60%. Model and parameter collaboration: Receives optimized model and detection parameter adjustment instructions from the cloud, and updates local algorithms and control parameters in real time; feeds back information such as device status and data processing success rate during the detection process to the cloud, supporting global scheduling and model iteration; Offline autonomy and fault self-healing: It has the ability to operate independently offline and can continue to complete detection tasks after disconnecting from the cloud; it monitors the hardware status in real time through fault monitoring algorithms, and automatically restarts key modules within 3 minutes when the device malfunctions, ensuring the continuity of detection.

[0026] Furthermore, the multi-source positioning data receiving and preprocessing submodule is used to receive and standardize various types of positioning-related raw data; the multi-source data fusion calculation submodule is used for extended Kalman filter (EKF) fusion calculation and reliability verification of the fusion results; specifically including: An EKF state equation and observation equation are constructed, using global coordinates provided by GNSS as observed values ​​and displacement increments provided by odometers as predicted values. The filter gain is calculated in real time to dynamically correct positioning errors. When GNSS signals are obstructed by bridge piers or shadowed by the bottom of beams, smoothing compensation using odometer data ensures continuous and stable longitudinal positioning, ultimately outputting longitudinal positioning results with an error ≤5cm for the detection platform. The reliability of the fused data is judged by calculating the sum of squared residuals. When the residuals exceed a preset threshold, the weight ratio of the two types of sensor data is automatically adjusted (e.g., 70% weight when GNSS signals are stable, and 90% weight when obstructed), further improving positioning robustness. The disease coordinate construction and cross-system mapping submodule is used to construct a disease-specific coordinate system and achieve precise adaptation with BIM / GIS systems; specifically, it includes: Extract the fused longitudinal positioning data (corresponding to the longitudinal distance to the bottom of the beam), and combine it with the fixed lateral number of the camera array (corresponding to the lateral position to the bottom of the beam) to construct a dual coordinate index system of "longitudinal distance + lateral number". Based on the pixel position of the defect in the panoramic image output by the defect identification and feature extraction module, associate the lateral number of the corresponding camera with the longitudinal distance of the detection platform to assign a unique spatial coordinate to each defect and clarify its precise physical position at the bottom of the beam. By using a preset coordinate transformation matrix, the local "vertical-horizontal" coordinates of the defects are converted into three-dimensional coordinates in the global coordinate system of the bridge. A standardized API interface is provided to integrate and map information such as defect coordinates and positioning accuracy with the three-dimensional spatial structure of the BIM model and the geographic latitude and longitude of the GIS system, so as to achieve accurate labeling of defects in the BIM model and geographic display on the electronic map.

[0027] Furthermore, the transmission device is connected to the edge computing and control module via control signals; the transmission device integrates a positioning sensor, which is a wheel-type odometer, and is connected to the positioning fusion module via communication to provide longitudinal positioning raw data; The camera array acquisition module is mechanically connected to the linear laser rangefinder and single-axis servo motor of the attitude control module, and is connected to the intelligent supplementary lighting module through a circuit. It provides consistent illumination to the camera through a ring LED supplementary lighting system. The camera array acquisition module is connected to the image processing module and the edge computing and control module through a data bus for real-time transmission of the acquired image data. The linear laser rangefinder of the attitude control module is connected to the edge computing and control module via analog / digital signals to transmit "camera-beam bottom distance" data in real time; the single-axis servo of the attitude control module is connected to the edge computing and control module via PWM control signals to receive attitude adjustment commands and drive camera angle changes. The ambient light sensor of the intelligent fill light module is connected to the edge computing and control module via analog signals to provide real-time brightness data; the LED driver circuit of the intelligent fill light module is connected to the edge computing and control module via PWM signals to receive brightness adjustment commands. The image processing module is connected to the camera array acquisition module through a data interface to receive raw image data; the image processing module is connected to the edge computing and control module through a data bus to transmit the corrected and stitched panoramic image data. The disease identification and feature extraction module is connected to the image processing module through a data interface to receive panoramic image data; the disease identification and feature extraction module is connected to the edge computing and control module through a data link to output analysis results such as disease type and morphological parameters. The edge computing and control module serves as the control and data processing hub of the system. It is bidirectionally connected to the detection platform module, camera array acquisition module, attitude control module, intelligent supplementary lighting module, image processing module, disease identification and feature extraction module, and positioning fusion module through control signals or data links, enabling synchronous scheduling, data interaction, and command issuance among multiple modules. The positioning fusion module is connected to the detection platform module via a serial port / network port to receive multi-source positioning raw data; the positioning fusion module is connected to the edge computing and control module via a data bus to output the precise coordinates of the defects in the form of "longitudinal distance + horizontal number"; if the system contains a BIM fusion module, the positioning fusion module is connected to the BIM module via a data interface to map the defect coordinates to the bridge 3D model.

[0028] like Figure 2 As shown, a second aspect of the present invention provides a dynamic panoramic intelligent detection method for bridge beam bottom defects of varying height, implemented using the aforementioned dynamic panoramic intelligent detection system for bridge beam bottom defects of varying height, comprising the following steps: S1. System Deployment and Parameter Determination: Assemble and deploy the guide rail, detection platform module, camera array acquisition module, intelligent supplementary lighting module, and attitude control module; clarify the core system parameters; and calculate the constant detection distance of the camera. and platform speed The intelligent supplementary lighting module was debugged and the multi-camera joint calibration was carried out to establish a three-dimensional geometric model of the beam bottom and a unified coordinate system. S2. Standardized Data Acquisition: The edge computing and control module outputs control commands to the transmission device, driving the detection platform to operate at a preset speed. Proceed at a constant speed, while maintaining a constant distance between the camera and the bottom of the beam. The attitude control module uses a PID algorithm to adjust the camera attitude to ensure that the imaging surface is parallel to the bottom of the beam. Simultaneously, the camera array is activated to acquire images, the intelligent supplementary lighting module dynamically supplements light, and the odometer and GNSS auxiliary data are acquired and transmitted to the edge computing and control module. S3. Data Processing and Disease Analysis: The image processing module corrects distortion in the acquired raw images and completes the panoramic image stitching of the beam bottom through feature matching and multi-view joint optimization; the disease identification and feature extraction module uses a pre-trained convolutional neural network to identify diseases and extract two-dimensional morphological parameters, and combines stereo pairs and point cloud clustering to quantify the three-dimensional volume of the disease. S4. Defect Location and Closed-Loop Inspection of the Entire Bridge: The positioning fusion module integrates GNSS and odometer data, and constructs a dual coordinate system by combining the horizontal numbering of the cameras. This enables precise location of each defect in the coordinate system at the bottom of the structural beam, and maps the defect information to the BIM and GIS systems for visualization. When the inspection platform encounters a pier, it uses the edge computing and control module to control the split electromagnetic adsorption beam to cross the pier, continuously advancing the inspection of the entire bridge and uploading all data to the cloud to form a closed loop, completing the dynamic panoramic intelligent inspection of defects at the bottom of the bridge beams at varying heights. Further, step S1 includes Hardware Assembly and Deployment: Two parallel guide rails are laid on the outside of the piers on both sides of the bridge beam bottom. The main body of the inspection platform (including steel truss support structure + split electromagnetic adsorption crossbeam + longitudinal support column + fixed rollers) is assembled. The frame is built through the steel truss support structure, and the split electromagnetic adsorption crossbeam is connected to the longitudinal support column on both sides through a ball joint structure. Fixed rollers adapted to the guide rail are installed on the top of the support column to complete the assembly of the inspection platform and the guide rail. The inspection platform is adapted to the guide rail. According to the transverse width of the bridge and the inspection requirements, a camera array is arranged transversely on the crossbeam of the inspection platform. A linear laser rangefinder (one unit for each camera), a single-axis servo motor (one unit for each camera) and a ring LED intelligent supplementary lighting module are installed to ensure that the transverse overlap rate of adjacent cameras is 20%. Define the core system parameters: record camera pixel size Camera focal length Camera frame rate; setting pixel physical resolution. (Take 0.1mm according to bridge inspection specifications), Longitudinal expected overlap rate (Default is 20%) Baseline parameter calculation: based on the formula Calculate the constant detection distance between the camera and the bottom of the beam. According to the formula Calculate the forward speed of the detection platform ; Equipment pre-testing: Start the intelligent supplementary lighting module, collect the initial ambient brightness through the light sensor, set the reference current of the LED array, simulate changes in ambient brightness to verify the dynamic adjustment function of the supplementary lighting system, and ensure the brightness consistency of the images captured by multiple cameras; Multi-camera joint calibration is completed: The image processing module uses Zhang Zhengyou's method combined with a global bundled adjustment algorithm to acquire images from all cameras through a checkerboard calibration board, solves the intrinsic parameter matrix and extrinsic parameter matrix (rotation matrix, translation vector) of each camera, and establishes a unified coordinate system benchmark. Preliminary data preparation: 3D point cloud data of the beam bottom is obtained through preprocessing, and a 3D geometric model of the beam bottom is constructed to provide a foundation for subsequent camera attitude adjustment and plane fitting; Furthermore, the multi-camera joint calibration in step S1 specifically includes: Calibration data acquisition: Place a calibration board, such as a checkerboard, at a fixed geometric position, and simultaneously acquire images of the calibration board using all cameras in the array; Parameter Solving: Using the "Zhang Zhengyou Calibration Method" combined with the "Global Bundling Adjustment Algorithm", the intrinsic parameter matrix (including focal length, pixel size, principal point coordinates, etc.) and extrinsic parameter matrix (including rotation matrix) of each camera are solved. Translation vector ); Relative calibration relationship: for adjacent cameras and Its extrinsic parameter matrix satisfies: in, It's a camera. The rotation matrix, It's a camera. The translation vector; It's a camera. The rotation matrix, It's a camera. The translation vector; It's a camera. To the camera The relative rotation matrix, It's a camera. To the camera The relative translation vector; Since the physical orientation of the array camera is fixed, a first-order rigid body constraint is introduced for optimization: in, It refers to the overall rotational attitude of the array. It refers to the overall translational orientation of the array; For the fixed rotational offset of each camera during the calibration phase, This involves fixed translational offsets for each camera during the calibration phase; during system operation (when the detection platform moves), only the overall array attitude needs to be updated. , This allows for the rapid acquisition of the extrinsic parameters of all cameras; it eliminates the need to repeatedly calculate the extrinsic parameters of each camera, significantly reducing the amount of real-time computation and improving system operating efficiency. Furthermore, bridge inspection standards require the effective identification of cracks as small as 0.1 mm. Therefore, in step S1 of this invention, the camera array layout needs to be combined with the camera pixel size and camera focal length. fTo meet the requirement of a pixel physical resolution of 0.1mm, the constant detection distance of the camera is calculated; the field of view of the camera is determined based on the set constant detection distance between the camera and the bottom of the beam; the effective sensing area of ​​a single camera is calculated according to a 20% overlap rate between adjacent images, and then the number of cameras required for the camera array is determined in combination with the transverse width of the bridge; at the same time, the detection speed of the beam bottom detection platform is calculated according to the expected overlap rate of 20% for the longitudinal images and the camera's acquisition frame rate, so as to realize motion control of the detection platform and ensure the standardization of the images acquired by the camera; The formula for calculating the constant detection distance between the motion camera and the bottom of the beam is as follows: in: The camera is positioned at a constant detection distance from the bottom of the beam. The pixel physical resolution is 0.1mm. For camera focal length, This refers to the camera pixel size; The forward speed of the detection platform module The expression is: in, The vertical length of the field of view of adjacent frames; Capture frame rate for camera; The desired overlap rate for the vertical image. ; Furthermore, steps S2-S4 utilize NUC (Next Unit of Computing) industrial computing units for edge computing and multi-threaded synchronous control, enabling high-speed parallel processing of multi-camera data acquisition, image stitching, defect identification, and remote data transmission. The NUC (Next Unit of Computing) industrial computing unit, with its core micro high-performance computing module, employs a multi-threaded architecture and distributed data caching technology to construct an intelligent control system for the testing platform, ensuring real-time, stable, and efficient data flow during the testing process. The calculation results and stored data from the NUC industrial computing unit are uploaded in real-time via 5G / Ethernet. Furthermore, in step S2, through ranging feedback and PID control, the optical axis of the motion camera is always perpendicular to the bottom surface of the bridge beam, ensuring that the camera's optical imaging surface is always parallel to the bottom surface of the beam, thus guaranteeing the standardization and accuracy of image acquisition; step S2 includes: S21, Platform Start-up and Speed ​​Control: Calculate the forward speed... The transmission control system of the input detection platform drives the detection platform longitudinally along the guide rail at a speed Proceed at a constant speed; the actual speed is fed back in real time by a wheel-mounted odometer, and the output power of the transmission device is dynamically adjusted to ensure that the platform speed remains stable. This ensures that the vertical overlap rate of the image remains constant at 20%. S22. Constant Camera Distance Control: Activate the linear laser rangefinder to collect real-time data on the actual distance between each camera and the bottom of the beam. The data is transmitted in real time to the edge computing and control module; the edge computing and control module then calculates the actual distance between the camera and the bottom of the beam. A constant detection distance between the camera and the bottom of the beam Comparison calculation of distance control error The camera position is dynamically adjusted via a transmission device. When the value is greater than 0, it is close to the bottom of the beam. (When the camera is at 0, move away from the bottom of the beam), ensuring that all cameras maintain a constant distance from the bottom of the beam. ; S23. Precise Camera Attitude Calibration: The edge computing and control module combines the 3D geometric model of the beam bottom to fit the equation of the local plane at the beam bottom and derive the normal vector of the local plane at the beam bottom. The camera's pitch angle is acquired via an attitude sensor. With roll angle Calculate the current normal vector of the camera's optical imaging surface. The target control angle of the servo motor is calculated based on the PID feedback control algorithm, which drives the camera to rotate so that the camera imaging plane is always parallel to the bottom plane of the beam, thus eliminating perspective distortion. S24. Multi-source data synchronous acquisition: The camera array acquisition module and the intelligent supplementary lighting module are started synchronously. The camera continuously acquires images of the beam bottom at a preset frame rate. The intelligent supplementary lighting module dynamically adjusts the LED array current according to the ambient brightness fed back by the light sensor in real time to maintain the brightness consistency of the multi-camera images. At the same time, the longitudinal displacement data of the wheel odometer, GNSS spatial position data, and camera attitude parameters are acquired. All data are transmitted to the NUC edge computing unit in real time. Furthermore, step S21 also includes: collecting the platform's actual moving speed in real time and comparing it with the theoretical forward speed using a wheeled odometer mounted on the platform. A comparison is made, and if any deviation exists, the output power of the transmission device is dynamically adjusted until the actual forward speed of the testing platform stabilizes at a certain level. To avoid affecting the image overlap rate due to speed fluctuations; Further, step S23 includes: S231: The edge computing and control module calls the beam bottom point cloud data previously acquired through binocular stereo imaging, performs local plane fitting on the beam bottom region at the current detection position, and calculates the local plane equation of the beam bottom using the least squares method; then, it derives the normal vector of the local plane of the beam bottom based on the local plane equation. ; S232: Acquires the current camera pitch angle using attitude sensors (such as gyroscopes and tilt sensors) mounted on the camera. With roll angle The data is then transmitted to the edge computing and control module; the edge computing and control module adjusts the current camera tilt angle accordingly. With roll angle The current normal vector of the camera's optical imaging plane is calculated. ; S233: Calculate the current normal vector of the camera's optical imaging surface. Normal vector to the local plane at the bottom of the beam The angle between the two sides is used to obtain the tilt angle error. Simultaneously adjust the camera attitude and distance control error, combining preset weighted coefficients for distance control error and tilt error. , The overall system error was calculated. ; S234: The edge computing and control module integrates system errors. Input to the PID controller, and calculate the target control angle of the servo motor according to the PID control algorithm. The generated servo control commands are transmitted to the single-axis servo in real time. S235: After receiving the control command, the single-axis servo drives its own axis to rotate to the target control angle, causing the camera to adjust its angle synchronously. After the angle adjustment is completed, distance data and camera attitude data are collected again through the linear laser rangefinder and attitude sensor. Steps S231-S234 are repeated to form a closed-loop control until the distance control error is reached. with tilt angle error All are within the preset threshold range, ultimately ensuring that the camera's optical imaging surface is always parallel to the bottom surface of the beam, eliminating image perspective distortion caused by changes in the slope of the beam bottom. Furthermore, under conditions of varying height or curved beam bottom, the curved surface along the detection direction can be locally fitted to the beam bottom plane; the plane fitting equation in step S231 is:

[0029] in, , This represents the x and y coordinates of a point on the curved surface at the bottom of the beam in a plane rectangular coordinate system. This indicates that a point on the curved surface at the bottom of the beam is perpendicular to... - Elevation coordinates in the plane direction; , The slope parameter of the bottom plane of the beam. Reflecting the bottom plane of the beam along The degree of tilt in the direction, Reflecting the bottom plane of the beam along The degree of inclination in the direction; For the bottom plane of the beam Intercept on the axis; This yields the normal vector of the local plane at the bottom of the beam in step S231. for: ; In step S232, the normal vector of the camera's optical imaging surface The expression is: in: The camera's tilt angle; This refers to the camera's roll angle; In step S233, it is assumed that the laser emission angle is... The measured distance from the camera to the bottom of the beam is The theoretical fixed distance from the camera to the bottom of the beam is The distance control error from the camera to the bottom of the beam is then... for: ; The system synthesis error expression in step S233 is: In the formula: This refers to the overall system error. , These are the weighting coefficients for distance control error and tilt angle error, respectively. For time, This refers to the tilt angle error; In step S234, the servo motor target control angle The PID control formula is: In the formula: For the target control angle of the servo motor, The proportional gain determines the sensitivity adjustment. To achieve integral gain and eliminate steady-state error; The differential gain improves dynamic response and reduces overshoot; through the combined effect of proportional gain, integral gain and differential gain, the target control angle of the servo motor is precisely adjusted, driving the camera to rotate, and ultimately ensuring that the camera optical axis is always perpendicular to the bottom surface of the beam, and that the camera optical imaging surface is always parallel to the bottom plane of the beam. Further, step S3 includes: S31. Image preprocessing: Radial and tangential distortion correction is performed on the original images acquired by the camera array based on the Brown-Conrady model to eliminate hardware imaging errors. S32. Panoramic image stitching: Using the laser line output by the linear laser rangefinder as the external constraint, the SIFT+RANSAC algorithm is used to extract feature matching point pairs of the overlapping areas of adjacent camera images and adjacent frame images. The coordinate system is unified through homography matrix mapping, and then combined with multi-view joint optimization to eliminate the cumulative stitching error and the brightness jump of the seam, generating a full-coverage, distortion-free panoramic image of the beam bottom. S33. Two-dimensional identification and feature extraction of diseases: The stitched panoramic image is input into a pre-trained convolutional neural network model (such as Mask-R-CNN, U-Net, YOLOv8). The model infers and outputs a mask map of the diseases (cracks, peeling, honeycomb pitting, etc.). Then, the two-dimensional contour of the disease is extracted based on the connected component analysis of the mask area. Combined with the "pixel-object size mapping relationship" of the camera calibration in the early stage, the two-dimensional morphological parameters such as the length, width and area of ​​the disease are calculated. S34. Three-dimensional quantitative analysis of defects: Using the overlapping areas of adjacent camera images or adjacent frame images as stereo pairs, stereo reconstruction is performed using the camera intrinsic and extrinsic parameter matrices to generate three-dimensional point cloud data of the beam bottom; for areas with defects such as concrete spalling and honeycomb pitting detected by two-dimensional vision, the area is quickly located, and the DBSCAN clustering algorithm is used to segment the three-dimensional point cloud, removing background and noise points, and separating the beam bottom reference plane and the abnormal change area of ​​the defect through plane fitting; for the abnormal point cloud clusters corresponding to the defects, voxel reconstruction or triangular mesh construction algorithms are used to calculate the three-dimensional parameters such as the depth and volume of the defects, realizing a complete analysis of defects from two-dimensional identification to three-dimensional quantitative analysis; Furthermore, camera imaging exhibits radial and tangential distortion. The intrinsic parameter matrix obtained from multi-camera joint calibration includes distortion coefficients (such as radial and tangential distortion coefficients). These distortion coefficients can be used to correct distortion in each acquired original image, eliminating hardware imaging errors and restoring the image to its true geometric shape. In step S31, the Brown-Conrady model is used to convert the original distorted pixel coordinates... Convert to corrected accurate pixel coordinates The relationship between the original pixels (distorted state) and the corrected pixels is described by the following set of equations: in:( , () represents the normalized distortion coordinates; , () represents the corrected, normalized coordinates; The distance from an image point to the center of the optical axis. ; All are radial distortion coefficients. Corresponding to low-order radial distortion, Corresponding to higher-order radial distortion; 、 , where is the tangential distortion coefficient, used to correct tangential distortion caused by the lens not being parallel to the imaging plane; Furthermore, in step S32, there is approximately a 20% overlap in both the horizontal and vertical directions between adjacent camera images. Using the markings formed by the linear laser rangefinder on the bottom of the beam as a reference, a feature matching algorithm is used to extract matching point pairs within the overlap area. and ;camera A feature point in the image is denoted as: , ( ) is the feature point in the camera Pixel coordinates in the image ( Corresponding image horizontal, (Corresponding to the vertical direction of the image); [·] T This represents the transpose of a vector, where "1" is the normalization term for homogeneous coordinates; camera A feature point in the image is denoted as: , ( ) is the feature point in the camera Pixel coordinates in the image, and Corresponding to the same physical location (such as the same marking point at the bottom of a beam); A homography matrix model based on perspective projection is used to unify multi-camera images to the same plane coordinate system; in the case of an approximate plane at the bottom of the beam, the correspondence between the images from two cameras can be expressed as: in, It is a proportionality coefficient. It's a camera. To the camera The homography matrix; camera To the camera The formula for calculating the homography matrix is: in, It's a camera. The intrinsic parameter matrix; It's a camera. The intrinsic parameter matrix; It's a camera. The extrinsic parameter rotation matrix; It's a camera. The extrinsic parameter rotation matrix; It's a camera. The extrinsic parameter translation vector; It is the normal vector of the local plane at the bottom of the beam; The distance from the bottom plane of the beam to the origin of the coordinate system; using this homography matrix, images from all cameras can be mapped to the same plane coordinate system, achieving initial stitching; Furthermore, to address the issues of "cumulative splicing error" and "joint problems caused by uneven lighting," step S32 introduces multi-view joint optimization to obtain a geometrically distortion-free and uniformly lit panoramic image of the beam bottom; the multi-view joint optimization includes geometric error optimization and lighting consistency optimization. Geometric error optimization aims to minimize the coordinate deviation of the same physical point in different views, ensuring that the coordinates of the same physical point are aligned after mapping in different views. The objective function for geometric error optimization is: in, This represents the geometric error of image stitching, indicating the degree of coordinate deviation of the same physical point in different views. The smaller the value, the higher the geometric accuracy of the stitching. The homography matrix between all cameras or views that need to be optimized simultaneously; It's a camera. To the camera The homography matrix; by adjusting the homography matrix This allows the projected coordinates of the same physical point to be aligned in different camera images. The illumination consistency optimization is as follows: while ensuring geometric alignment, the influence of uneven illumination from multiple cameras at the bottom of the beam is considered, and an optical flow consistency constraint is introduced to ensure continuous brightness changes between adjacent images. The objective function for the joint optimization of geometry and illumination is: in, The homography matrix between all cameras or views that need to be optimized simultaneously; These are weighting coefficients used to balance the optimization priorities of "geometric error" and "illumination error"; For feature points In the camera Pixel brightness values ​​(grayscale / color) in an image; For feature points In the camera Pixel brightness values ​​in the image; The objective function is to calculate the squared error of the brightness of the same physical point in different views. The grayscale value or color of the same physical point in the overlapping area should be similar in different images. By constraining the brightness consistency of the overlapping area, the "brightness jump" of the joint can be reduced. The objective function optimizes "coordinate alignment" and "brightness consistency" at the same time, and finally obtains a panoramic image of the beam bottom with geometric distortion-free and uniform lighting. The panoramic result is stored in a unified pixel coordinate system for subsequent defect identification and BIM mapping. Further, step S4 includes: S41. Multi-source positioning data fusion and defect coordinate calculation: The extended Kalman filter (EKF) algorithm is used to fuse GNSS positioning data (providing global spatial position) and wheel odometer displacement data (providing high-precision longitudinal displacement) to eliminate single sensor errors and obtain high-precision longitudinal positioning results of the detection platform; combined with the lateral numbering of the camera array (corresponding to the lateral position at the bottom of the beam), a dual coordinate indexing system of "longitudinal distance + lateral number" is constructed to assign a unique spatial coordinate to each identified defect, clarifying the specific location of the defect at the bottom of the beam; S42. Integration and Visualization of Defect Information with BIM / GIS Systems: Spatial coordinates, two-dimensional morphological parameters (length, area, etc.), three-dimensional quantitative data (volume, depth, etc.), original images, panoramic images, and other information of defects are mapped to the bridge BIM 3D model through API interfaces to achieve accurate labeling of defects in the model; at the same time, it is integrated with the GIS geographic information system to associate information such as the bridge's geographical location and station number, forming an integrated display of "spatial location - defect characteristics - visualization model" for easy and intuitive viewing and management; S43. Cross-pier inspection connection and full-area coverage: When the inspection platform moves to the pier position, the NUC edge computing unit identifies the pier position through positioning data and triggers the cross-pier control command: controls the electromagnetic switch of the split electromagnetic adsorption beam to open, drives the beam to rotate around the ball joint structure to a collinear state with the longitudinal support column, avoiding interference with the pier; after the inspection platform crosses the pier along the guide rail through the transmission device, controls the electromagnetic switch to close, restores the original connection state of the beam, and continues to advance at the preset speed v and inspection process to ensure that there are no blind spots at the bottom of the beam and achieve full bridge coverage; S44. Data Upload and Full-Process Closed-Loop Formation: During the inspection process, edge computing and multi-threaded synchronous control are performed through the NUC industrial computing unit. The raw data (images, point clouds, sensor data), processing results (stitched images, defect parameters), positioning information, BIM / GIS fusion data, etc., collected by the camera array acquisition module are transmitted to the cloud database in real time via 5G / Ethernet for classified storage and backup. After the inspection is completed, a full bridge bottom defect inspection report (including defect list, location, characteristic parameters, visualization model links, etc.) is generated, forming a full-process closed loop of "deployment-collection-processing-identification-positioning-storage-reporting", providing complete data support for bridge maintenance and repair. Furthermore, step S41 also includes: when the detection platform moves in the longitudinal direction under the bridge beam, it acquires the relative displacement between the GNSS output measurement and the wheel odometer output in real time; in the transverse direction under the beam, array cameras are fixedly arranged to form multiple camera channels, and each camera is assigned a unique transverse number, thereby forming a unified "longitudinal distance + transverse number" dual coordinate positioning system for the unique spatial number of each defect. Furthermore, step S42 also includes: obtaining the image number of the defect through the defect identification module, calling the EKF fusion positioning module to output the vertical position corresponding to the same timestamp, locating the position of the defect on the bottom plane of the beam according to the dual coordinate system rules, combining the bottom height information of the beam to map to the BIM model to achieve three-dimensional display, and transmitting the defect location information in the BIM model to the GIS system through the API interface, overlaying it on the electronic map for display, and realizing visual management.

[0030] This invention proposes a split-type electromagnetic adsorption rail-mounted inspection platform to achieve full-area cross-pier movement of beams with variable cross-sections and steep slopes; it constructs a camera array to achieve full-bridge visual perception, and dynamically controls the platform's movement speed by adjusting the camera's frame rate and overlap rate to achieve stable and standardized data acquisition; it utilizes linear laser distance sensing to perceive the distance between the camera and the beam bottom, dynamically adjusting the distance and tilt angle between the camera and the beam bottom to achieve image acquisition at constant distance and posture; it proposes a method for beam bottom image distortion correction and full-frame stitching, and uses image recognition and point cloud segmentation technology to identify beam bottom defects and extract two-dimensional and three-dimensional features; it proposes a NUC joint control scheme to achieve synchronous control of the camera array; it establishes a GNSS-odometer-camera coordinate fusion defect localization technology, and integrates it with a BIM system to achieve bridge information visualization and digital twin; thus, it achieves fully automated inspection of beam bottoms with variable cross-sections and steep slopes.

[0031] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic panoramic intelligent detection system for bridge beam bottom defects of varying height, characterized in that, It includes a detection platform module, a camera array acquisition module, an attitude control module, an intelligent supplementary lighting module, an image processing module, a disease identification and feature extraction module, an edge computing and control module, and a positioning fusion module; The detection platform module includes a main body of the beam bottom detection platform, guide rails respectively arranged on the outside of the bridge piers on both sides of the beam bottom, a steel truss support structure for supporting the main body of the beam bottom detection platform, a transmission device for driving the main body of the beam bottom detection platform to move longitudinally, and a split electromagnetic adsorption connecting beam connected to the main body of the beam bottom detection platform. The camera array acquisition module includes multiple moving cameras arranged laterally at preset intervals; the lateral overlap rate of images acquired by adjacent moving cameras is 20%, and all moving cameras maintain a constant detection distance from the bottom of the beam; the attitude control module includes linear laser rangefinders and single-axis servos mounted on the moving cameras of the camera array acquisition module; the intelligent supplementary lighting module is a ring LED supplementary lighting system with ambient light perception; the image processing module is used to realize joint camera calibration, image distortion correction, and panoramic image stitching of the bottom of the beam; the defect identification and feature extraction module is used to realize defect type identification, two-dimensional morphological parameter calculation, and three-dimensional volume quantization; the edge computing and control module is used to realize multi-threaded synchronous control of camera acquisition, image processing, defect identification, and remote data transmission based on NUC industrial computing units; the positioning fusion module is used to realize precise defect positioning based on GNSS and wheeled odometer fusion positioning technology, combined with the lateral numbering of the camera array to form a dual-coordinate index system; through the efficient collaboration of various modules, the level of intelligence and visualization of complex bridge inspection is comprehensively improved.

2. The dynamic panoramic intelligent detection system for bridge beam bottom defects of varying heights according to claim 1, characterized in that, The main body of the beam bottom detection platform is equipped with rollers that are compatible with the guide rail; The steel truss support structure includes longitudinal support columns on both sides of the bottom of the bridge beam and a steel truss on top of the longitudinal support columns on both sides of the bottom of the bridge beam. The crossbeam is connected to the longitudinal support columns through a ball joint structure. When facing the need for pier inspection, the electromagnetic switch is controlled to disconnect the connection in the middle of the crossbeam, rotate the crossbeam to the collinear state of the columns, and after crossing the pier, restore the original state to continue the inspection, so as to achieve full accessibility of the bridge.

3. The dynamic panoramic intelligent detection system for bridge beam bottom defects of varying heights according to claim 2, characterized in that, The transmission device integrates a positioning sensor, which is a wheel-type odometer; the detection platform module has a GNSS module integrated on its split electromagnetic adsorption connecting beam. The LED supplemental lighting system includes an ambient light sensor and an LED driving circuit; the ambient light sensor is connected to the edge computing and control module via an analog signal to provide real-time brightness data; the LED driving circuit is connected to the edge computing and control module via a PWM signal to receive brightness adjustment commands. The linear laser rangefinder of the attitude control module is connected to the edge computing and control module via analog / digital signals to transmit "camera-beam bottom distance" data in real time; the single-axis servo of the attitude control module is connected to the edge computing and control module via PWM control signals to receive attitude adjustment commands and drive camera angle changes. The image processing module is connected to the camera array acquisition module via a data interface to receive raw image data; the image processing module is connected to the edge computing and control module via a data bus to transmit corrected and stitched panoramic image data. The edge computing and control module serves as the control and data processing hub of the system. It is bidirectionally connected to the detection platform module, camera array acquisition module, attitude control module, intelligent supplementary lighting module, image processing module, disease identification and feature extraction module, and positioning fusion module through control signals or data links. This enables synchronous scheduling, data interaction, and command issuance among multiple modules. The positioning fusion module is connected to the detection platform module and is used to receive multi-source positioning raw data; the positioning fusion module is connected to the edge computing and control module through a data bus and is used to output the precise coordinates of the disease in the form of "vertical distance + horizontal number".

4. A dynamic panoramic intelligent detection method for bridge beam bottom defects of varying height, characterized in that, The application of the dynamic panoramic intelligent detection system for bridge beam bottom defects of varying heights as described in any one of claims 1-3 includes the following steps: S1. Assemble and deploy the guide rail, detection platform module, camera array acquisition module, intelligent supplementary lighting module, and attitude control module; clarify the core system parameters; and calculate the constant detection distance of the camera. and platform speed The intelligent supplementary lighting module was debugged and the multi-camera joint calibration was carried out to establish a three-dimensional geometric model of the beam bottom and a unified coordinate system. S2. The edge computing and control module outputs control commands to the transmission device, driving the detection platform to operate at a preset speed. Moving forward at a constant speed, the attitude control module uses a PID algorithm to adjust the camera attitude to ensure that the imaging surface is parallel to the bottom of the beam. Simultaneously, the camera array is activated to acquire images, and the intelligent supplementary lighting module dynamically supplements the light. At the same time, the odometer and GNSS auxiliary data are acquired and transmitted to the edge computing and control module. S3, the image processing module corrects distortion in the acquired raw images and completes the panoramic image stitching of the beam bottom through feature matching and multi-view joint optimization; the disease identification and feature extraction module uses a pre-trained convolutional neural network to identify diseases and extract two-dimensional morphological parameters, and combines stereo view pairs and point cloud clustering to quantify the three-dimensional volume of the disease. S4, the positioning fusion module integrates GNSS and odometer data, and constructs a dual coordinate system by combining the horizontal numbering of the cameras, to achieve precise positioning of each defect in the coordinate system of the bottom of the structural beam, and maps the defect information to the BIM and GIS systems for visualization; when the detection platform encounters a pier, the edge computing and control module controls the split electromagnetic adsorption beam to complete the crossing of the pier, continuously promotes the detection of the entire bridge, and uploads all data to the cloud to form a closed loop, completing the dynamic panoramic intelligent detection of defects at the bottom of the beam with varying height.

5. The method for dynamic panoramic intelligent detection of bridge beam bottom defects with varying height according to claim 4, characterized in that, In step S1, the constant detection distance between the camera and the bottom of the beam The calculation formula is: in: The camera is positioned at a constant detection distance from the bottom of the beam. For pixel physical resolution, For camera focal length, Camera pixel size; Detection platform forward speed The calculation formula is: in, The vertical length of the field of view of adjacent frames; Capture frame rate for camera; The desired overlap rate for the vertical image. .

6. The method for dynamic panoramic intelligent detection of bridge beam bottom defects with varying height according to claim 5, characterized in that, Step S1, multi-camera joint calibration, includes: A checkerboard calibration board is placed at a fixed geometric position, and images of the calibration board are simultaneously acquired by all cameras in the array. Using the "Zhang Zhengyou calibration method" combined with the "global bundled adjustment algorithm", the intrinsic parameter matrix and extrinsic parameter matrix of each camera are solved. The intrinsic parameter matrix includes focal length, pixel size, and principal point coordinates; the extrinsic parameter matrix includes rotation matrix and translation vector. Relative calibration relationship: for adjacent cameras and Its extrinsic parameter matrix satisfies: in, It's a camera. The rotation matrix, It's a camera. Translation vector; It's a camera. The rotation matrix, It's a camera. Translation vector; It's a camera. To the camera The relative rotation matrix, It's a camera. To the camera The relative translation vector; Introducing first-order rigid body constraint optimization, the expression for first-order rigid body constraint optimization is: ； in, It refers to the overall rotational attitude of the array. It refers to the overall translational orientation of the array; For the fixed rotational offset of each camera during the calibration phase, This refers to the fixed translation offset of each camera during the calibration phase.

7. The method for dynamic panoramic intelligent detection of bridge beam bottom defects with varying height according to claim 5, characterized in that, Step S3 includes: S31. Image preprocessing: Radial and tangential distortion correction is performed on the raw images acquired by the camera array based on the Brown-Conrady model; S32. Panoramic image stitching: Using the laser line output by the linear laser rangefinder as the external constraint, the SIFT+RANSAC algorithm is used to extract feature matching point pairs of the overlapping areas of adjacent camera images and adjacent frame images. The coordinate system is unified through homography matrix mapping, and then combined with multi-view joint optimization to eliminate the cumulative stitching error and the brightness jump of the seam, generating a full-coverage, distortion-free panoramic image of the beam bottom. S33. Two-dimensional identification and feature extraction of diseases: The stitched panoramic image is input into a pre-trained convolutional neural network model. The model infers and outputs a mask image of the disease. Then, the two-dimensional contour of the disease is extracted based on the connected component analysis of the mask area. Combined with the "pixel-object size mapping relationship" of the camera calibration in the early stage, the two-dimensional morphological parameters of the disease are calculated. The two-dimensional morphological parameters of the disease include the length, width and area of ​​the disease. S34. Three-dimensional quantitative analysis of defects: Using the overlapping areas of adjacent camera images or adjacent frame images as stereo pairs, stereo reconstruction is performed using the camera intrinsic and extrinsic parameter matrices to generate three-dimensional point cloud data of the beam bottom; the DBSCAN clustering algorithm is used to segment the three-dimensional point cloud, removing background and noise points, and the reference plane of the beam bottom and the abnormal change area of ​​the defect are separated by plane fitting; the three-dimensional quantitative parameters of the defect are calculated by using voxel reconstruction or triangular mesh construction algorithms for the abnormal point cloud clusters corresponding to the defect, realizing the complete analysis of the defect from two-dimensional identification to three-dimensional quantification; the three-dimensional quantitative parameters of the defect include the depth and volume of the defect.

8. The method for dynamic panoramic intelligent detection of bridge beam bottom defects with varying height according to claim 7, characterized in that, Step S32 includes: Using the markings formed on the bottom of the beam by the linear laser rangefinder as a reference, a feature matching algorithm is used to extract matching point pairs within the overlapping area. and ; The homography matrix model based on perspective projection is used to unify the images from multiple cameras to the same plane coordinate system, thus completing the initial stitching. When the bottom of the beam is approximately a plane, the correspondence between the two camera images can be expressed as: ； in, It is a proportionality coefficient. It's a camera. To the camera The homography matrix; For camera A feature point in the image; For camera A feature point in the image; camera To the camera The formula for calculating the homography matrix is: ； in, It's a camera. The intrinsic parameter matrix; It's a camera. The intrinsic parameter matrix; It's a camera. The extrinsic parameter rotation matrix; It's a camera. The extrinsic parameter rotation matrix; It's a camera. The extrinsic parameter translation vector; It is the normal vector of the local plane at the bottom of the beam; The distance from the bottom plane of the beam to the origin of the coordinate system.

9. The method for dynamic panoramic intelligent detection of bridge beam bottom defects with varying height according to claim 6, characterized in that, Step S32 introduces multi-view joint optimization; the multi-view joint optimization includes geometric error optimization and illumination consistency optimization; The objective function for geometric error optimization is: ； in, Geometric errors in image stitching; The homography matrix between all cameras or views that need to be optimized simultaneously; It's a camera. To the camera The homography matrix; The illumination consistency optimization is as follows: while ensuring geometric alignment, the influence of uneven illumination from multiple cameras at the bottom of the beam is considered, and an optical flow consistency constraint is introduced to ensure continuous brightness changes between adjacent images. The objective function for the joint optimization of geometry and illumination is: ； in, The homography matrix between all cameras or views that need to be optimized simultaneously; These are the weighting coefficients; For feature points In the camera Pixel brightness values ​​in the image; For feature points In the camera Pixel brightness values ​​in the image; The squared error of brightness of the same physical point in different views.

10. A dynamic panoramic intelligent detection method for bridge beam bottom defects of varying height according to any one of claims 5-9, characterized in that, Step S4 includes: S41. Multi-source positioning data fusion and defect coordinate calculation: The extended Kalman filter algorithm is used to fuse GNSS positioning data with displacement data from the wheel odometer integrated with the transmission device to obtain high-precision longitudinal positioning results of the detection platform; combined with the lateral numbering of the camera array, a dual coordinate indexing system of "longitudinal distance + lateral numbering" is constructed to assign a unique spatial coordinate to each identified defect and clarify the specific location of the defect at the bottom of the beam. S42. Integration and visualization of disease information with BIM / GIS system: The spatial coordinates, two-dimensional morphological parameters, three-dimensional quantitative data, original images, and panoramic image information of the disease are mapped to the bridge BIM three-dimensional model through the API interface to achieve accurate labeling of the disease in the model; at the same time, it is integrated with the GIS geographic information system to associate the geographical location and stationing information of the bridge, forming an integrated display of "spatial location-disease characteristics-visualized model". S43. Cross-pier inspection connection and full-area coverage: When the inspection platform moves to the pier position, the NUC edge computing unit identifies the pier position through positioning data and triggers the cross-pier control command: controls the electromagnetic switch of the split electromagnetic adsorption beam to open, drives the beam to rotate around the ball joint structure to a collinear state with the longitudinal support column, avoiding interference with the pier; after the inspection platform crosses the pier along the guide rail through the transmission device, controls the electromagnetic switch to close, restores the original connection state of the beam, and continues to advance at the preset speed and inspection process to ensure that there are no blind spots at the bottom of the beam and achieve full bridge coverage; S44. Data Upload and Full-Process Closed Loop Formation: During the inspection process, the NUC edge computing unit transmits the collected raw data, processing results, positioning information, and BIM / GIS fusion data to the cloud database in real time; after the inspection is completed, a full bridge bottom defect inspection report is generated, forming a full-process closed loop of "deployment-collection-processing-identification-positioning-storage-reporting".

Images