High-speed rail cable trench polyurethane spraying partition method based on visible light and laser radar

Abstract

The application discloses a high-speed rail cable trench polyurethane spraying partition method based on visible light and laser radar, and relates to the field of automatic spraying construction. The application significantly improves the identification accuracy and robustness of special-shaped structures such as a trench bottom, a trench wall, a dark corner, a bright corner, a drain hole and a junction box by fusing visible light deep learning identification and laser radar point cloud geometry analysis to construct a cable trench semantic point cloud model. On this basis, the identification result is directly converted into a partition spraying control strategy of a standard spraying area, a strengthened spraying area, a forbidden spraying area, an edge compensation area, a rechecking area and a manual review area, and a forbidden spraying area safety expansion and early gun control are introduced, which effectively reduces the risk of mis-spraying of the forbidden spraying object, and improves the coating quality of weak areas through the strengthened spraying area and the edge compensation area control, thereby providing an overall solution of automation, intelligentization and quality traceability for high-speed rail cable trench polyurethane waterproof and insulating material spraying construction.

Description

Technical Field

[0001] This invention relates to the field of automated spraying construction, specifically to a method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar. Background Technology

[0002] Currently, the application of polyurethane waterproof insulation materials for high-speed railway cable trenches is mostly done manually by brushing, rolling, or using handheld spray guns. However, manual methods make it difficult to reliably identify complex structures such as the trench bottom, trench walls, internal and external corners, cover plate supports, drainage holes, junction boxes, embedded parts, and obstacles. Furthermore, irregularly shaped areas such as internal and external corners and step edges are prone to missed areas, thin coatings, or discontinuous coatings. Areas that should not be sprayed, such as drainage holes and junction boxes, are also frequently accidentally sprayed or blocked due to improper operation. At the same time, relying solely on visible light vision is easily affected by factors such as lighting, shadows, stains, and water accumulation, making it difficult to accurately assess the coating. While stable acquisition of three-dimensional geometric boundaries is essential, relying solely on LiDAR point clouds, though capable of reflecting spatial morphology, struggles to accurately determine the semantic categories of objects like drainage holes and junction boxes. Existing recognition technologies largely focus on target detection or defect identification, failing to directly translate perception results into zoning control strategies applicable to polyurethane spraying construction, such as standard spraying areas, reinforced spraying areas, no-spraying areas, edge compensation areas, re-inspection areas, and manual verification areas. Furthermore, traditional construction processes lack quality traceability archives linked to railway mileage, images, point clouds, spraying parameters, and no-spraying protection records. Therefore, a method integrating visible light recognition and LiDAR point cloud analysis is urgently needed to achieve accurate semantic modeling, zoning spraying control, and mileage-based quality recording. Summary of the Invention

[0003] To address the technical problems mentioned above, this invention provides a method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar, comprising: S1. Acquire visible light images and lidar 3D point clouds of the cable trench; S2. Perform geometric analysis based on the 3D point cloud and deep learning recognition based on the visible light image to identify the geometric structure and object category of the cable trench; S3. Based on the spatial calibration relationship between the visible light camera and the lidar, the visual recognition results are mapped to the three-dimensional point cloud to generate a semantic point cloud model of the cable trench with semantic and geometric categories. S4. Identify irregular structures based on the semantic point cloud model of the cable trench, and divide the cable trench into multiple spraying zones based on the identification results of the irregular structures and the constraints of the spraying process. S5. Generate a spraying control strategy based on the spraying zoning results and execute the spraying, while simultaneously associating and storing data to form a milestone-based construction quality record.

[0004] Preferably, in S1, the three-dimensional point cloud is represented as follows: , in, This represents a set of lidar point clouds; Represents the i-th point in the point cloud; ) represents the three-dimensional coordinates of a point in the point cloud.

[0005] Preferably, in step S2, when performing geometric analysis on the 3D point cloud, the normal vector of a local region of the point cloud is denoted as... n i The normal vector of the adjacent region is denoted as n j The angle between the two is: , in, This represents the normal vector of a local region between two adjacent point clouds. Indicates the geometric angle between adjacent regions; Indicates the threshold for determining geometric mutations; when If so, then a significant geometric abrupt change is considered to exist.

[0006] Preferably, in step S2, the deep learning recognition includes: setting... t Visible light image at any time The visual recognition model output is: , in, , To identify the object category, To identify confidence levels, This refers to the target boundary, detection box, or segmentation region in the image.

[0007] Preferably, in step S3, the method for mapping the visual recognition result to a 3D point cloud includes: Point cloud points in the lidar coordinate system Through the external parameter matrix T CL Transform to camera coordinate system Then, it is projected onto the image plane through the camera intrinsic parameter matrix K: ,in, Represents the point cloud points in the lidar coordinate system; Represents the point cloud points in the camera coordinate system; This represents the extrinsic parameter matrix of the lidar to the camera; Represents the camera intrinsic parameter matrix; This represents the pixel coordinates of a point cloud point projected onto an image. This represents the projection scale factor.

[0008] Preferably, in step S3, the semantic point cloud model includes fusion confidence. The calculation method is as follows: , in: , The consistency function is: , in, Indicates the fusion confidence level; Indicates the confidence level of visible light recognition; This represents the confidence level for point cloud geometric recognition; This indicates consistency between visual and geometric categories; This represents the weighting coefficient.

[0009] Preferably, in S4, for the first k Each area is a designated spraying zone. According to semantic category Geometric categories Fusion confidence and grassroots status Sure: ,and Selected from standard spraying area, enhanced spraying area, no-spray area, edge compensation area, re-inspection area and manual verification area.

[0010] Preferably, in step S5, for the original area of ​​the prohibited spraying target... Create a no-spraying-protected zone ,in , These are respectively: recognition error compensation, robot positioning error compensation, spray diffusion compensation, and wind disturbance or scattering compensation; Let the robot's walking speed be... v The spray gun shut-off response time is t q The control system delay is t c The safety margin is The distance for prematurely stopping the gun is: .

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates visible light deep learning recognition with lidar point cloud geometric analysis to construct a semantic point cloud model for cable trenches, significantly improving the accuracy and robustness of identifying irregular structures such as trench bottoms, trench walls, internal and external corners, drainage holes, and junction boxes. Based on this, the recognition results are directly converted into a zoned spraying control strategy, including standard spraying areas, enhanced spraying areas, no-spray areas, edge compensation areas, re-inspection areas, and manual verification areas. Safety expansion and early gun shut-off control in no-spray areas are introduced to effectively reduce the risk of accidental spraying of prohibited objects. Simultaneously, the enhanced spraying areas and edge compensation areas improve the coating quality in weak areas. Furthermore, the system automatically associates and stores images, point clouds, zoning results, spraying parameters, and railway mileage to form a traceable, mileage-based construction quality record, providing an automated, intelligent, and traceable overall solution for the spraying of polyurethane waterproof insulation materials in high-speed railway cable trenches. Attached Figure Description

[0012] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the specific implementation steps of an embodiment of the present invention; Figure 2 This is a schematic diagram of the system composition according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the fusion of visible light recognition and lidar point cloud analysis in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the partitioning of the spraying area according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the semantic point cloud model structure according to an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the no-spray zone protection and early gun shut-off control in an embodiment of the present invention; Figure 7 This is a flowchart illustrating the milestone-based construction quality recording process according to an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] Example 1: like Figure 1 The diagram shown is a flowchart of the method in this embodiment, and its steps include: S1. Acquire visible light images and lidar 3D point clouds of the cable trench.

[0017] The spraying robot is controlled to move along the direction of the high-speed rail cable trench, with the visible light camera, lidar, and spray gun facing inwards. Images of the cable trench are acquired using the visible light camera, and finally, a 3D point cloud of the cable trench is acquired using the lidar.

[0018] Point clouds are used to describe the three-dimensional spatial morphology inside cable trenches, providing geometric basis for the identification of trench bottoms, trench walls, internal and external corners, cover plate supports, holes, and obstacles. The three-dimensional point cloud acquired by lidar is represented as follows: , in, This represents a set of lidar point clouds; Represents the i-th point in the point cloud; ) represents the three-dimensional coordinates of a point in the point cloud.

[0019] S2. Perform geometric analysis based on the 3D point cloud and deep learning recognition based on the visible light image to identify the geometric structure and object category of the cable trench.

[0020] Geometric analysis is performed on the point cloud to identify the trench bottom plane, trench wall plane, internal corners, external corners, cover plate support platform, holes, protrusions, and obstacles. The normal vector of a local region of the point cloud is denoted as... n i The normal vector of the adjacent region is denoted as n j The angle between the two is: , in, This represents the normal vector of a local region between two adjacent point clouds. Indicates the geometric angle between adjacent regions; Indicates the threshold for determining geometric mutations; when If so, then it is considered that there is a significant geometric abrupt change in the region.

[0021] This formula is used to identify inside corners, outside corners, step edges, and local protrusions in cable trenches. The trench bottom is a horizontal plane, and the trench wall is a vertical plane. The normal vector changes significantly at the junction of the two, so it can be identified as an inside corner. The cover plate support platform and the top edge of the trench wall can be identified as outside corners or step edges due to abrupt changes in height and direction.

[0022] Simultaneously, a deep learning visual recognition model is used to identify the trench bottom, trench walls, internal and external corners, cover plate support platforms, drainage holes, junction boxes, embedded parts, and obstacles in the cable trench. This deep learning visual recognition model represents the category, location, and confidence level of each object within the cable trench after the image has passed through the deep learning model. The model can be implemented using object detection, semantic segmentation, or instance segmentation methods, without limiting the specific network structure, as detailed below: set up t Images captured by the visible light camera at all times are The visual recognition model output is: , in, The visible light image was acquired at time t. A deep learning recognition model for visible light; The visible light recognition result is represented as follows: , in, The object category identified by visible light; Confidence level for visible light identification; This refers to the target boundary, detection box, or segmentation region in the image.

[0023] The visual semantic category set is defined as: , The meanings of the symbols are shown in Table 1: Table 1: .

[0024] S3. Based on the spatial calibration relationship between the visible light camera and the lidar, the visual recognition results are mapped to a 3D point cloud to generate a semantic point cloud model of the cable trench with semantic and geometric categories.

[0025] Based on the spatial calibration relationship between the camera and the lidar, the image visual recognition results are mapped to a 3D point cloud to generate a cable trench semantic point cloud model with visual semantic category, point cloud geometric category, fusion confidence score and railway mileage information.

[0026] The 3D point cloud is then projected onto the 2D image. The steps include: (1) Image semantics and point cloud spatial mapping model.

[0027] The point cloud representation in the lidar coordinate system is as follows: , Transform it to the camera coordinate system using the extrinsic parameter matrix TCL between the camera and the LiDAR: , Then project it onto the image plane using the camera intrinsic parameter matrix K: , in, Represents the point cloud points in the lidar coordinate system; Represents the point cloud points in the camera coordinate system; This represents the extrinsic parameter matrix of the lidar to the camera; Represents the camera intrinsic parameter matrix; This represents the pixel coordinates of a point cloud point projected onto an image. This represents the projection scale factor.

[0028] After projection, if a point cloud point falls into the "drainage hole" region of the image, it receives the semantic label "drainage hole"; if it falls into the "ditch wall" region, it receives the semantic label "ditch wall". This achieves the fusion of 2D visual results and 3D point cloud results. The fusion process of visual and 3D point cloud results is as follows: Figure 3 As shown.

[0029] (2) Semantic point cloud model of cable trench.

[0030] The fused semantic point cloud definition of the cable trench is as follows: The semantic point cloud set is as follows: in, This represents the semantic category obtained from visible light recognition. This represents the geometric category obtained from point cloud analysis. Indicates the fusion confidence level. This is represented as railway mileage information.

[0031] (3) Fusion confidence model.

[0032] Visible light recognition confidence level is P v The confidence level for point cloud geometric recognition is P g The consistency between the visual category and the point cloud geometric category is δ( C v , C g If ), then the fusion confidence level is: , in: , The consistency function is: , in, Indicates the fusion confidence level; Indicates the confidence level of visible light recognition; This represents the confidence level for point cloud geometric recognition; This indicates consistency between visual and geometric categories; This represents the weighting coefficient.

[0033] This model is used to determine whether visual recognition and point cloud geometric analysis support each other. For example, if visible light recognition identifies it as a "drain hole" and point cloud analysis identifies it as a "hole or depression," then the consistency is high; if visible light recognition identifies it as a "ditch wall," but the point cloud shows it as a "protruding obstacle," then the consistency is low. When: P f ≥ P If the value is 0, the recognition result is valid. When: P f < P If the value is 0, then the area is designated as the re-inspection area. See the illustration below. Figure 5 As shown.

[0034] S4. Identify irregular structures based on the semantic point cloud model of the cable trench, and divide the cable trench into multiple spraying zones based on the identification results of the irregular structures and the constraints of the spraying process.

[0035] Based on the semantic point cloud model of cable trenches, irregular structures in high-speed railway cable trenches were identified, and the 3D boundaries, identification confidence scores, and railway mileage ranges of each irregular structure were calculated. Simultaneously, based on the irregular structure identification results and the constraints of the polyurethane waterproof insulation material spraying process, the cable trenches were divided into standard spraying areas, reinforced spraying areas, no-spraying areas, edge compensation areas, re-inspection areas, and manual verification areas. The spraying zoning diagram is shown below. Figure 4 As shown.

[0036] For the k Each identification region is used to determine the spraying zone based on semantic category, geometric category, fusion confidence, and substrate condition. , in, Indicates the first k The spraying zone classification for each area; Indicates the semantic category of the region; Indicates the geometric category of the region; Indicates the confidence level of regional integration; This indicates the state of affairs at the grassroots level.

[0037] The spraying zone set is as follows: , The meanings of the characters are shown in Table 2: Table 2: .

[0038] As can be seen from the above steps, the system does not only identify the cable trench structure, but also transforms the identification results into functional zones that can be directly used in the spraying process. Different zones correspond to different spraying control strategies.

[0039] S5. Generate a spraying control strategy based on the spraying zoning results and execute the spraying, while simultaneously associating and storing data to form a milestone-based construction quality record.

[0040] Based on the zoning results of the spraying area, a corresponding spraying control strategy is generated, and the spraying of polyurethane waterproof insulation material is carried out according to the spraying control strategy. Finally, the cable trench image, 3D point cloud, irregular structure recognition results, spraying area zoning results, spraying control strategy, and railway mileage information are linked and stored to form a mileage-based construction quality record.

[0041] No-spray targets should not be avoided solely based on the identified physical boundary; instead, a safe distance should be extended outwards to form a no-spray protection zone. This prevents accidental spraying due to spray diffusion or positioning errors. The original area of ​​the identified no-spray target is recorded as follows: Taking into account identification errors, positioning errors, spray diffusion, and wind disturbance, a no-spraying zone is generated: , in: , The meanings of the characters are shown in Table 3: Table 3: .

[0042] The robot should not wait until the spray gun reaches the boundary of the no-spray zone before shutting off; instead, it should shut off in advance based on speed, spray gun response time, and control delay. This prevents polyurethane material from entering no-spray areas such as drain holes and junction boxes due to spray gun lag or spray diffusion. The no-spray zone protection and early shut-off control process is as follows: Figure 6 As shown.

[0043] Let the robot's walking speed be... v The spray gun shut-off response time is t q The control system delay is t c The safety margin is The distance for prematurely stopping the gun is: .

[0044] With a fixed spray width, a higher spray flow rate results in a thicker coating; a faster robot movement speed results in less material per unit area and a thinner coating. This model is used to guide the adjustment of spraying parameters in standard spraying areas, enhanced spraying areas, and edge compensation areas.

[0045] Let the spraying flow rate be Q, and the material utilization rate be... η The robot's walking speed is v Effective spray width is W The theoretical wet film thickness is h ,but: , The meanings of the characters are shown in Table 4: Table 4: .

[0046] When the target wet film thickness is At that time, the spraying flow rate can be calculated in reverse: , The robot's walking speed can also be calculated in reverse: .

[0047] Example 2: This embodiment applies to cable trenches with relatively regular structures in high-speed railway bridge sections. Such cable trenches typically include a trench bottom, left and right trench walls, the internal corner where the trench bottom and trench walls meet, the external corner at the top of the trench walls, and a cover plate support platform, among other structures.

[0048] In this embodiment, the spraying robot is positioned above or near the centerline of the cable trench and moves along the centerline of the high-speed rail cable trench. Visible light cameras and lidar are both positioned facing inwards from the cable trench, ensuring their sensing range covers the trench bottom, left and right walls, internal and external corners, and the cover plate support platform.

[0049] During the robot's movement, a visible light camera captures images of the cable trench, a lidar simultaneously acquires three-dimensional point clouds of the cable trench, and a pose and mileage positioning module records the robot's current position and railway mileage information.

[0050] The visible light recognition module processes the cable trench image, identifying target categories such as the trench bottom, trench walls, internal corners, external corners, and cover plate support platforms, and outputs the visual semantic category, image boundary, and recognition confidence score. The lidar point cloud analysis module performs geometric analysis on the 3D point cloud, extracting the trench bottom plane, the left and right trench wall planes, the boundary line between the trench bottom and the trench wall, the edge of the cover plate support platform, and the top edge of the trench wall.

[0051] For local areas of the point cloud, the system determines geometric abrupt changes based on the angle between the normal vectors of adjacent areas. When there is a significant abrupt change in normal between the bottom plane and the wall plane, the corresponding boundary area is identified as a concave angle; when there is abrupt change in height and direction at the top of the wall or the edge of the cover plate support, it is identified as a convex angle or edge structure.

[0052] Based on the extrinsic parameter calibration relationship between the camera and the lidar, the system maps the semantic categories of the trench bottom, trench walls, internal corners, external corners, and cover plate support platform identified from the visible light images to corresponding point cloud points, generating a semantic point cloud model of the cable trench. The point cloud points in this semantic point cloud model include three-dimensional coordinates, visual semantic categories, point cloud geometric categories, fusion confidence scores, and railway mileage information.

[0053] In this embodiment, the system divides the cable trench into the following spraying areas based on the semantic point cloud model: The areas with intact structures and normal base layers at the bottom and left and right sides of the trench are designated as standard spraying areas. The area at the inner corner where the trench bottom meets the trench wall is designated as the reinforced spraying zone; The top corner of the trench wall and the edge of the cover plate support platform are designated as edge compensation zones. Local areas where the visual recognition results and point cloud geometric results are inconsistent are designated as re-inspection areas.

[0054] For standard spraying areas, the spraying robot sprays according to preset standard spraying speed, flow rate, and spray width. For reinforced spraying areas, the system reduces the robot's walking speed or increases the spray overlap rate to improve the coverage quality of polyurethane material at inside corners. For edge compensation areas, the system performs local compensation spraying to reduce the risk of missed or thin spraying at outside corners and step edges.

[0055] During the spraying process, the system associates and stores images, point clouds, recognition results, spraying zoning results, spray gun start / stop status, spraying flow rate, robot walking speed, and railway mileage information to form a construction quality record for the corresponding mileage section.

[0056] This embodiment is applicable to construction scenarios where there are drainage holes, junction boxes, embedded parts, grounding terminals, exposed cables, or other objects unsuitable for spraying in high-speed railway cable trenches.

[0057] In this embodiment, the spraying robot moves along the centerline of the cable trench. A visible light camera continuously captures images of the trench sidewalls and bottom, while a lidar simultaneously acquires 3D point clouds of the corresponding areas. The visible light recognition module identifies targets such as drain holes, junction boxes, embedded parts, grounding terminals, and obstacles based on a deep learning detection model, and outputs the target category, image boundaries, and visual confidence level.

[0058] The lidar point cloud analysis module performs geometric analysis on the areas containing the aforementioned objects. For drainage holes, the point cloud shows local depressions, holes, or missing point cloud areas; for junction boxes, embedded parts, and grounding terminals, the point cloud shows local protrusions or protruding structures. The system matches the visual recognition results with the point cloud geometric results. If the visible light recognition identifies it as a drainage hole and the point cloud geometry shows a hole or depression, the fusion confidence of that area is increased; if the visible light recognition identifies it as a junction box and the point cloud shows a local protrusion, the fusion confidence of that target is increased.

[0059] The system divides objects such as drain holes, junction boxes, embedded parts, grounding terminals, exposed cables, nameplates, and equipment interfaces into no-spray zones. For no-spray objects, the system first obtains their three-dimensional boundaries and then performs safety expansion on the boundaries of the no-spray objects based on recognition error, robot positioning error, spray diffusion compensation, and wind disturbance compensation to form a no-spray protection zone.

[0060] No-spray protection zones can be represented as: , in, This is the original no-spray zone; This is the expanded no-spraying protection zone; For safety buffer; For a safe buffer distance; Region expansion operation.

[0061] The safe buffer distance can be determined by the following formula: , in, Indicates the total safe buffer distance of the no-spray zone; Indicates the recognition error compensation distance; This indicates the robot's positioning error compensation distance; Indicates the spray diffusion compensation distance; Indicates the compensation distance for wind disturbance or material scattering. The design of no-spray zones should not only consider the size of the target itself, but also take into account identification bias, robot motion errors, and the effects of spray material diffusion.

[0062] For example, the actual identification area of ​​a certain drain hole is If visual recognition error compensation Robot positioning error compensation Spray diffusion compensation Wind disturbance compensation The safe buffer distance is: d s =10+15+20+5=50mm, In other words, the system does not actually avoid the original boundary of the drainage hole, but rather the no-spray protection zone formed after expanding outward by 50 mm.

[0063] When the spray gun is less than the pre-shutdown distance from the front boundary of the no-spray zone, the system controls the spray gun to shut off prematurely. For objects such as drain holes and junction boxes that are easily contaminated or clogged, the system can also control the shielding device to position a shield or flexible shield between the spray area and the no-spray object. After the robot passes through the no-spray zone, the system controls the spray gun to resume spraying.

[0064] Because the spray gun has a shut-off response time, the control system also has a signal delay. If the spray gun is shut off only when the robot has moved to the boundary of the no-spray zone, the polyurethane material may have already been sprayed into the no-spray zone. Therefore, this embodiment uses early spray gun shut-off control.

[0065] When the painting robot approaches the no-spray zone, the system calculates the advance gun-closing distance based on the robot's walking speed, the spray gun's shut-off response time, the control system delay, and the safety margin. Let the robot's walking speed be... The spray gun shut-off response time is The control system delay is The safety margin is The distance for prematurely stopping the gun is: , in, This indicates the distance at which the gun is turned off in advance; This is expressed as the robot's walking speed; This represents the spray gun shut-off response time; This is represented as a control system delay; This is expressed as a safety margin.

[0066] The robot should be within a certain distance of the front boundary of the no-spray protection zone. Turn off the spray gun in advance, instead of waiting until the spray gun reaches the no-spray zone.

[0067] For example, the robot's walking speed is: , The spray gun shutdown response time is: , The control system delay is: , The safety margin is: , The distance for prematurely stopping the gun is: , The system should shut off the spray gun when it is about 130 mm away from the front boundary of the no-spraying zone.

[0068] In this embodiment, the no-spray zone protection record includes the category of the no-spray object, the three-dimensional boundary, the range of the no-spray protection zone, the distance for early gun shut-off, the gun shut-off time, the status of the masking protection action, and the time for resuming spraying. After the above data is linked to railway mileage information, a mileage-based construction quality record is formed.

[0069] In practice, the system performs spray-prevention protection according to the following procedure: Step 1: Identify prohibited spraying targets such as drain holes, junction boxes, embedded parts, and grounding terminals; Step 2: Obtain the image boundary and 3D point cloud boundary of the no-spray object to obtain the original no-spray object region. ; Step 3: Calculate the safe buffer distance based on the identification error, positioning error, spray diffusion, and wind disturbance. ; Step 4: Based on Create a no-spraying zone; Step 5: Calculate the advance gun shut-off distance based on robot speed, spray gun shut-off response time, and control delay. ; Step 6: When the spray gun is less than or equal to the front boundary of the no-spray protection zone At that time, control the spray gun to turn off in advance; Step 7: For easily contaminated objects such as drain holes and junction boxes, control the action of the shielding plate or flexible shielding component to position the shielding device between the spray area and the prohibited spraying object; Step 8: After the robot passes through the no-spray zone, the spray gun resumes spraying.

[0070] The above control methods can reduce the risk of polyurethane material being accidentally sprayed into drain holes, junction boxes, embedded parts, and grounding terminals.

[0071] Example 3: This embodiment is applicable to construction scenarios where the base layer of the cable trench is in poor condition or the identification conditions are complex, such as situations where there are cracks, falling pieces, water accumulation, oil stains, debris obstruction, severe damage, or missing point clouds in the cable trench.

[0072] In this embodiment, the spraying robot moves along the cable trench and acquires visible light images and lidar point clouds. The visible light recognition module identifies targets such as the trench bottom, trench walls, internal corners, external corners, cracks, loose pieces, water accumulation, and oil stains. The lidar point cloud analysis module identifies the trench bottom plane, trench wall plane, steps, protrusions, depressions, and locally missing areas.

[0073] For cracks, beam end joints, expansion joints, areas with historical leaks, or areas requiring patch repair, the system classifies them as reinforced spraying zones. For reinforced spraying zones, the system can reduce robot walking speed, increase spray overlap, or increase the number of spray coats to improve the coating coverage quality of weak waterproof areas.

[0074] For areas with water accumulation, oil stains, severe damage, obstruction by debris, severe image occlusion, or severe point cloud loss, the system classifies them as manual review areas. For these areas, the system does not directly perform automatic spraying; instead, it outputs a manual confirmation prompt and marks the anomaly type, location, and suggested handling method in the mileage quality record.

[0075] For regions where the visible light recognition results and point cloud geometric analysis results are inconsistent, the system calculates the fusion confidence score. The fusion confidence score can be expressed as: , in, Indicates the fusion confidence level; Represented as visual recognition confidence level; Represented as point cloud geometric recognition confidence; This represents the semantic category obtained from visible light recognition; This represents the geometric category obtained from point cloud analysis; This indicates the consistency between the visual category and the point cloud geometric category. Represents the weighting coefficients; the weighting coefficients satisfy: .

[0076] The class consistency function is defined as follows: , The value is 1 when the visible light recognition result and the point cloud geometry result are completely consistent; the value is 0.5 when there is some correlation between the two but it cannot be completely confirmed; and the value is 0 when the two are obviously conflicting.

[0077] The identification results are shown in Table 5.

[0078] Table 5: .

[0079] When the fusion confidence level is lower than a preset threshold, the system classifies the corresponding area as a re-inspection area. For the re-inspection area, the robot can reduce its walking speed and re-acquire images and point clouds; if the confidence level meets the requirements after the second recognition, it is reclassified as the corresponding spraying area; if it still cannot be confirmed after the second recognition, it is classified as a manual verification area.

[0080] For the re-examination area determination, the fusion confidence threshold is set as follows: P 0. When the following conditions are met: If the identification reliability of that area is insufficient, it will be designated as a re-inspection area. The current confidence level of regional integration; The threshold for re-inspection is set.

[0081] For example, let The confidence level of visual recognition in a certain area is The confidence level for point cloud geometric recognition is Furthermore, the two are only partially related. .like , , ,but: , because: , Therefore, this area was designated as a re-inspection area.

[0082] If the following conditions are detected during the manual review process, the area should be classified as a manually reviewed area even if the fusion confidence level is not low: ① There is obvious water accumulation on the surface of the cable trench; ② There is oil on the surface of the cable trench; ③ The base layer is severely damaged or has broken pieces; ④ There are obstructions; ⑤ The image is severely overexposed, underexposed, or obscured; ⑥ Point cloud is severely missing; ⑦ The robot is unable to confirm the condition of the painted surface.

[0083] The manual review area does not directly enter the automatic spraying process; instead, it prompts on-site personnel for processing or confirmation. After processing, the robot can re-capture images and point clouds for re-identification and zoning. In this embodiment, the system avoids blind spraying when the base layer does not meet the spraying requirements or the identification results are unreliable by using a re-inspection area and a manual verification area mechanism, thereby improving construction safety and quality reliability.

[0084] Example 4: This embodiment is applicable to construction scenarios where the spraying flow rate, robot walking speed, and spraying overlap rate need to be automatically adjusted according to different spraying areas.

[0085] In this embodiment, the system generates a spraying control strategy based on the spraying area zoning results. The spraying control strategy includes standard spraying control, enhanced spraying control, no-spray zone protection control, edge compensation control, re-inspection control, and manual review prompts.

[0086] Let the spraying flow rate be Material utilization rate The robot's walking speed is Effective spray width is The theoretical wet film thickness is ,but: in, Indicates the spraying flow rate; Indicates the theoretical wet film thickness; Indicates material utilization rate; Indicates the robot's walking speed; Indicates the effective spray width.

[0087] With a fixed effective spray width, the more material sprayed per unit time, the thicker the coating; the faster the robot walks, the less material is obtained per unit area, and the thinner the coating.

[0088] For areas requiring enhanced spraying, the system can reduce the robot's walking speed, increase the spraying overlap rate, or increase the number of spraying coats, thereby providing higher coverage for vulnerable waterproof areas such as corners, cracks, beam end seams, and expansion joints.

[0089] For edge compensation areas, the system can reduce the robot's walking speed, adjust the spray gun angle, or perform local compensation spraying to improve the coating continuity at external corners, cover plate support edges, and around obstacles.

[0090] For no-spray zones, the system sets the spray flow rate to zero and implements early gun shut-off and masking protection controls.

[0091] During the spraying process, the quality recording module binds the spraying area partitioning, spraying flow rate, robot walking speed, spray gun start / stop status, spraying time, no-spray protection actions, and railway mileage information to form a construction quality record. After construction is completed, images, point clouds, spraying partitions, spraying parameters, and anomaly markers corresponding to any mileage segment can be queried based on railway mileage, providing a basis for subsequent project acceptance and quality traceability.

[0092] Example 3: This embodiment illustrates the collaborative working process among the mobile painting robot platform, visible light camera, lidar, spray gun and painting actuator, pose and odometer positioning module, control unit, image recognition module, point cloud analysis module, odometer mapping module, fusion modeling module, semantic point cloud model, irregular structure recognition module, painting area partitioning module, painting control module, and quality recording module.

[0093] The system composition in this embodiment is similar to that of... Figure 2 The corresponding system composition diagram is shown in Table 6, and the relationship between each module and its functions is shown in Table 6.

[0094] Table 6: .

[0095] Step 1: Initialize the mobile spraying robot platform.

[0096] After the mobile spraying robot platform A enters the construction section of the high-speed railway cable trench, the control unit F initializes each module of the system.

[0097] The initialization content should include at least: (1) Check the communication status of visible light camera B, lidar C, and pose and odometer positioning module E; (2) Check the spray gun start / stop, flow rate adjustment, attitude adjustment and shielding protection functions of the spray gun and spraying actuator D; (3) Read the camera intrinsic parameters, the extrinsic parameters between the camera and the lidar, and the sensor installation parameters relative to the robot platform; (4) Read the spray flow rate, spray width, target wet film thickness, standard walking speed and spray restriction protection parameters of polyurethane material; (5) Establish the starting point of the railway mileage for the current construction section; (6) Synchronize the image timestamp, point cloud timestamp, pose timestamp and spraying control timestamp.

[0098] After the system self-test passes, the control unit F controls the mobile spraying robot platform A to move along the centerline of the cable trench. The visible light camera B, lidar C, and spray gun and spraying actuator D are preferably arranged in the transverse central area of ​​the cable trench so that the sensing range and spraying range cover the bottom of the trench, the left and right trench walls, the inside corners, the outside corners, and the cover plate support platform.

[0099] Step 2: Visible light image acquisition and image recognition.

[0100] During the movement of the mobile spraying robot platform A, a visible light camera B continuously acquires images of the inside of the cable trench and sends the images and corresponding timestamps to the image recognition module G. The image recognition module G first performs distortion correction, brightness equalization, shadow suppression, and edge enhancement on the images, and then uses a deep learning detection model, semantic segmentation model, or instance segmentation model to identify the objects in the cable trench.

[0101] The image recognition module G outputs the following objects for recognition: the bottom of the cable trench, the left and right walls of the trench, the inside corner between the bottom and the wall, the top of the wall and the outside corner at the cover support, the cover support, drainage holes, junction boxes, embedded parts and grounding terminals, exposed cables, obstacles, cracks, loose pieces, water accumulation and oil stains, and other base layer abnormalities.

[0102] The image recognition result can be represented as: , in, For visual semantic categories; For visual recognition confidence; This refers to the detection box, segmentation contour, or pixel region of the corresponding object in the image.

[0103] The image recognition module G sends the above two-dimensional semantic category, target boundary and recognition confidence to the fusion modeling module J.

[0104] Step 3: LiDAR point cloud acquisition and point cloud geometric analysis.

[0105] The lidar C works synchronously with the visible light camera B to continuously acquire three-dimensional point clouds inside the cable trench and send the point clouds and timestamps to the point cloud analysis module H.

[0106] Point cloud analysis module H performs the following processing on the raw point cloud: (1) Remove outliers and drift points; (2) Perform voxel downsampling on the point cloud; (3) Remove non-target point clouds such as the robot body and the external structure of the bridge deck; (4) Extract the effective point cloud of the cable trench; (5) Slice the point cloud cross-sections according to the longitudinal direction of the railway or the direction of robot travel; (6) Perform plane fitting on the trench bottom and trench walls; (7) Calculate the local normal vector and the degree of geometric mutation; (8) Extract the three-dimensional boundaries of inside corners, outside corners, steps, holes, protrusions and obstacles.

[0107] The geometric categories output by the point cloud analysis module H include at least the following: horizontal or near-horizontal trench bottom plane, vertical or near-vertical left and right trench wall planes, the inside corner at the junction of the trench bottom and trench wall, the top of the trench wall, the outside corner at the edge of the step, the cover plate support platform and continuous steps, holes or depressions, protruding or protruding components, obstacles, and areas with missing or geometrically abnormal points in the cloud.

[0108] For two adjacent local point cloud regions, their normal vectors are respectively and The angle between the two is: , when When the geometric aberration exceeds a preset threshold, the point cloud analysis module H marks the corresponding region as a candidate region for a concave corner, convex corner, step edge, or protrusion. The point cloud analysis module H then sends the 3D geometric category, 3D boundary, and geometric recognition confidence score to the fusion modeling module J.

[0109] Step 4: Pose acquisition and railway mileage mapping.

[0110] The pose and odometer positioning module E is used to continuously acquire the position, attitude, direction of movement, and cumulative travel of the mobile painting robot platform A. The pose and odometer positioning module E can use one or more of the following positioning information: drive wheel encoder, inertial measurement unit, visual odometer, lidar odometer, RTK or Beidou positioning, railway mileage marker recognition, and known starting mileage of the construction section.

[0111] The pose and odometry localization module E sends the robot's pose information to the odometry mapping module I. The odometry mapping module I calculates the image frames, point cloud frames, and railway mileage coordinates corresponding to each identified object based on the robot's railway mileage and the sensor's installation position relative to the robot platform.

[0112] If the starting mileage of the current construction section is The robot's cumulative longitudinal movement distance along the railway is Then the base mileage corresponding to the current robot can be expressed as: , in, This indicates the railway mileage corresponding to the robot at the current moment; Indicates the starting mileage of the construction section; This represents the cumulative longitudinal distance the robot has moved from its starting point.

[0113] For targets detected by the camera or lidar, the mileage mapping module I further corrects the target mileage based on the target's longitudinal offset relative to the robot's coordinate system. The mileage mapping module I then sends the coordinate unification results and railway mileage information to the fusion modeling module J.

[0114] Step 5: Fusion of image semantics, point cloud geometry, and odometry information.

[0115] The fusion modeling module J receives two-dimensional semantic recognition results from the image recognition module G, three-dimensional geometric recognition results from the point cloud analysis module H, and robot pose and railway mileage information from the mileage mapping module I. Based on the spatial calibration relationship between the camera and the LiDAR, the fusion modeling module J projects the LiDAR point cloud onto the visible light image plane.

[0116] Points in the point cloud in the lidar coordinate system are represented as follows: , Through the external parameter matrix T CL Transform to camera coordinate system: , Then project it onto the image plane using the camera intrinsic parameter matrix: , The fusion modeling module J is based on the projected pixels. Obtain the visual semantic category and visual confidence score for the location, and assign them to the corresponding point cloud point.

[0117] The fusion confidence score is calculated using the following formula: , in, Represents the point cloud points in the lidar coordinate system; Represents the point cloud points in the camera coordinate system; This represents the extrinsic parameter matrix of the lidar to the camera; Represents the camera intrinsic parameter matrix; This represents the pixel coordinates of a point cloud point projected onto an image. This represents the projection scale factor. Indicates the fusion confidence level; Indicates the confidence level of visible light recognition; This represents the confidence level for point cloud geometric recognition; This indicates consistency between visual and geometric categories; Let represent the weight coefficients, and satisfy: .

[0118] The consistency is high when the visual recognition identifies it as a "drain hole" and the point cloud analysis identifies it as a "hole or depression"; the consistency is low when the visual recognition identifies it as a "ditch wall" and the point cloud analysis identifies it as a "protruding obstacle".

[0119] Step 6: Generate semantic point cloud model of cable trench The fusion modeling module J writes the fusion result into the semantic point cloud model K. A single semantic point in the semantic point cloud model K is represented as: , in, This represents the semantic category obtained from visible light recognition. This represents the geometric category obtained from point cloud analysis. Indicates the fusion confidence level. This is represented as railway mileage information.

[0120] Compared to ordinary point clouds, the semantic point cloud model K can not only represent the three-dimensional shape of the cable trench, but also indicate the type of structure each point belongs to, the reliability of the recognition result, and the corresponding railway mileage location. The semantic point cloud model K is divided into continuous construction units according to a preset mileage length or robot movement distance, and output to the irregular structure recognition module L.

[0121] Step 7: Identification of irregular structures.

[0122] The irregular structure recognition module L performs spatial clustering, boundary extraction, and category confirmation on the semantic point cloud model K to identify the following cable trench structures: (1) Ditch bottom; (2) Left and right ditch walls; (3) Left and right inner corners; (4) External corner; (5) Cover plate support platform; (6) Drain hole; (7) Junction box; (8) Embedded parts and grounding terminals; (9) Exposed cables; (10) Obstacles; (11) Abnormalities in the base layer, such as cracks, spalling, water accumulation and oil stains.

[0123] Each irregular construction object includes at least: construction category, three-dimensional boundary, spatial location, fusion confidence, railway start and end mileage, geometric dimensions, and base condition.

[0124] When the fusion confidence level is lower than the preset threshold, the irregular structure recognition module L marks the object as a low-confidence object, which is then divided into a re-inspection area by the subsequent spraying area partitioning module M.

[0125] Step 8: Divide the spraying area into zones.

[0126] The spraying area partitioning module M divides the cable trench into different spraying functional areas based on the structural category, three-dimensional boundary, fusion confidence, railway mileage and base condition output by the irregular structure identification module L.

[0127] Partitioning rules include: (1) Standard spraying area: The bottom of the trench, the trench wall, and the conventional spraying surface with complete structure and normal base condition are divided into the standard spraying area.

[0128] (2) Enhanced spraying area: The inside corner, cracks, beam end joints, expansion joints, historical leakage areas and chip repair areas are divided into enhanced spraying areas.

[0129] (3) No-spray zone: Drain holes, junction boxes, embedded parts, grounding terminals, exposed cables, nameplates and equipment interfaces are designated as no-spray zones.

[0130] (4) Edge compensation zone: The outer corner, the edge of the cover plate support platform, the top edge of the trench wall and the perimeter of the obstacle are divided into edge compensation zones.

[0131] (5) Re-inspection area: The area where the fusion confidence is lower than the preset threshold, or where the visual semantic category and the point cloud geometric category conflict with each other, is designated as the re-inspection area.

[0132] (6) Manual verification area: areas with water accumulation, oil stains, severe damage, debris obstruction, severe image obstruction, or severe point cloud loss are designated as manual verification areas.

[0133] The spraying area partitioning module M sends the three-dimensional boundary, partitioning category, fusion confidence level, and railway mileage of each area to the spraying control module N.

[0134] Step 9: Zoned spraying control.

[0135] The spraying control module N generates corresponding control commands based on the output of the spraying area partitioning module M, and sends the control commands to the spray gun and the spraying actuator D.

[0136] (1) Standard spraying control: For the standard spraying area, the spraying control module N controls the spray gun and the spraying actuator D to spray according to the standard flow rate, standard travel speed, standard spray width and target wet film thickness.

[0137] (2) Enhanced spraying control: For areas requiring enhanced spraying, the spraying control module N improves the coverage quality of areas such as corners and cracks by reducing the robot's walking speed, increasing the spraying overlap rate, increasing the spraying flow rate, or increasing the number of spraying passes.

[0138] (3) No-spray zone control: For the no-spray zone, the spraying control module N safely expands the three-dimensional boundary of the no-spray object to form a no-spray protection zone, and calculates the advance spray gun closing distance based on the robot's walking speed, spray gun closing response time, and control delay. When the spray gun approaches the no-spray protection zone, the spraying control module N controls the spray gun and the spraying actuator D to close the spray gun in advance, and drives the shielding protection device to operate as needed. After passing through the no-spray zone, the spraying control module N controls the spray gun to resume spraying.

[0139] (4) Edge compensation control: For the edge compensation area, the spraying control module N controls the spray gun to adjust the spraying trajectory, spray gun posture, local speed or overlap rate to reduce the risk of missed spraying at the corners and steps.

[0140] (5) Re-inspection and manual verification control: For the re-inspection area, the spraying control module N controls the mobile spraying robot platform A to slow down and re-acquire images and point clouds. For the manual verification area, the spraying control module N pauses automatic spraying and outputs a manual confirmation prompt through the control unit F.

[0141] Step 10: The spray gun and spraying actuator move.

[0142] The spray gun and spraying actuator D perform the following actions according to the instructions issued by the spraying control module N: opening or closing the spray gun, adjusting the spraying flow rate, adjusting the spraying pressure, adjusting the spray gun horizontally, vertically or at an angle, extending or retracting the masking protection device, performing edge compensation actions, and cleaning or anti-clogging actions for the spray gun.

[0143] The spray gun and spraying actuator D feed back the actual execution status to the control unit F and the quality recording module O. The feedback data includes the actual spray gun start / stop status, actual spraying flow rate, spray gun posture, masking device status, and abnormal alarm information.

[0144] Step 11: The quality record module generates milestone construction records.

[0145] The quality recording module O receives data from the control unit F, the semantic point cloud model K, the irregular structure recognition module L, the spraying area zoning module M, and the spraying control module N, and establishes a mileage-based construction quality record. The construction quality recording process is as follows: Figure 7 As shown.

[0146] The quality record should include at least the following: construction line and section, cable trench number, railway start and end mileage, visible light image, lidar point cloud, semantic point cloud model of cable trench, irregular structure category and three-dimensional boundary, spraying area zoning category, zoning fusion confidence, spraying control strategy, spray gun start and stop status, actual spraying flow rate, robot walking speed, no-spray zone protection and shielding device action record, re-inspection area and manual review area record, operation time and abnormal alarm information.

[0147] The quality record for a single mileage segment can be represented as: in, For the quality record of the k-th mileage segment, For railway mileage range, Visible light image, For lidar point clouds, For semantic point cloud data, The results of the irregular structure identification are as follows: The results of the spraying area zoning are as follows. For spray control strategies, Record the actual movements of the spray gun and actuator. This refers to the time allotted for the assignment.

[0148] Step 12: Coordination and troubleshooting of the control unit.

[0149] Control unit F, acting as the system coordination center, manages the operational status of each module. When the visible light camera B or image recognition module G malfunctions, control unit F can reduce the robot speed and maintain limited recognition capability based on the LiDAR point cloud analysis results, while simultaneously dividing the corresponding area into a re-inspection zone or a manual verification zone. When the LiDAR C or point cloud analysis module H malfunctions, control unit F can stop automatic spraying that relies on 3D boundaries and mark the current area as a manual verification zone. When the pose and mileage positioning module E or mileage mapping module I malfunctions, control unit F suspends automatic spraying to prevent misalignment of the identified object, spraying area, and railway mileage. When the spray gun and spraying actuator D fail to operate as instructed, or when the actual spraying flow rate deviates from the set flow rate by more than a threshold, control unit F stops spraying and records the fault. When the quality recording module O experiences a storage malfunction, control unit F can temporarily store the data in a local cache and continue writing it after the storage is restored.

[0150] In this embodiment, each system module forms a complete closed-loop work link: the mobile spraying robot platform carries and moves → visible light camera and lidar perceive simultaneously → image recognition, point cloud analysis and odometer mapping are processed in parallel → fusion modeling generates semantic point cloud → irregular structure recognition → spraying area partitioning → spraying control module drives spray gun and actuator → quality record module forms odometer archive.

[0151] This embodiment fully demonstrates the functions and data relationships of each module in the system composition diagram, and can realize the identification of irregular structures in high-speed railway cable trenches, automatic division of spraying areas, differentiated spraying of polyurethane materials, protection of prohibited spraying objects, and traceability of construction quality.

[0152] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for partitioning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar, characterized in that, include: S1. Acquire visible light images and lidar 3D point clouds of the cable trench; S2. Perform geometric analysis based on the 3D point cloud and deep learning recognition based on the visible light image to identify the geometric structure and object category of the cable trench; S3. Based on the spatial calibration relationship between the visible light camera and the lidar, the visual recognition results are mapped to the three-dimensional point cloud to generate a semantic point cloud model of the cable trench with semantic and geometric categories. S4. Identify irregular structures based on the semantic point cloud model of the cable trench, and divide the cable trench into multiple spraying zones based on the identification results of the irregular structures and the constraints of the spraying process. S5. Generate a spraying control strategy based on the spraying zoning results and execute the spraying, while simultaneously associating and storing data to form a milestone-based construction quality record.

2. The method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar according to claim 1, characterized in that, In S1, the three-dimensional point cloud is represented as: ， in, This represents a set of lidar point clouds; Represents the i-th point in the point cloud; ) represents the three-dimensional coordinates of a point in the point cloud.

3. The method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar according to claim 1, characterized in that, In S2, when performing geometric analysis on a 3D point cloud, the normal vector of a local region of the point cloud is denoted as... n i The normal vector of the adjacent region is denoted as n j The angle between the two is: ， in, This represents the normal vector of a local region between two adjacent point clouds. Indicates the geometric angle between adjacent regions; Indicates the threshold for determining geometric mutations; when If so, then a significant geometric abrupt change is considered to exist.

4. The method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar according to claim 1, characterized in that, In S2, the deep learning recognition includes: setting t Visible light image at any time The visual recognition model output is: ， in, , To identify the object category, To identify confidence levels, This refers to the target boundary, detection box, or segmentation region in the image.

5. The method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar according to claim 1, characterized in that, In step S3, the method for mapping the visual recognition result to the 3D point cloud includes: Point cloud points in the lidar coordinate system Through the external parameter matrix T CL Transform to camera coordinate system Then, it is projected onto the image plane through the camera intrinsic parameter matrix K: ,in, Represents the point cloud points in the lidar coordinate system; Represents the point cloud points in the camera coordinate system; This represents the extrinsic parameter matrix of the lidar to the camera; Represents the camera intrinsic parameter matrix; This represents the pixel coordinates of a point cloud point projected onto an image. This represents the projection scale factor.

6. The method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar according to claim 1, characterized in that, In S3, the semantic point cloud model includes fusion confidence. The calculation method is as follows: ， in: ， The consistency function is: ， in, Indicates the fusion confidence level; Indicates the confidence level of visible light recognition; This represents the confidence level for point cloud geometric recognition; This indicates consistency between visual and geometric categories; This represents the weighting coefficient.

7. The method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar according to claim 1, characterized in that, In S4, for the first k Each area is a designated spraying zone. According to semantic category Geometric categories Fusion confidence and grassroots status Sure: ,and Selected from standard spraying area, enhanced spraying area, no-spray area, edge compensation area, re-inspection area and manual verification area.

8. The method for zoning polyurethane spraying in high-speed railway cable trenches based on visible light and lidar according to claim 1, characterized in that, In step S5, the original area of ​​the prohibited spraying target... Create a no-spraying-protected zone ,in , These are respectively: recognition error compensation, robot positioning error compensation, spray diffusion compensation, and wind disturbance or scattering compensation; Let the robot's walking speed be... v The spray gun shut-off response time is t q The control system delay is t c The safety margin is The distance for prematurely stopping the gun is: 。

Images