Intelligent obstacle avoidance planning and dynamic recovery method in hull block automatic measurement

Abstract

The embodiment of the application discloses a kind of intelligent obstacle avoidance planning and dynamic recovery method in hull section automatic measurement.The fixed optical measurement system is prone to interruption and low efficiency in complex scaffold shielding environment for automated measurement, and the application constructs the closed-loop system of "global perception and planning→intelligent decision and positioning→local execution and recovery".The method comprises: establishing a global coordinate system by laying out global reference light targets on the hull section, and generating a global three-dimensional obstacle avoidance model using motion structure recovery technology;Based on parallel computing, the number of measurable points and the comprehensive accuracy evaluation value are independently calculated for the preset candidate parking point, and the only optimal parking point is automatically selected according to the two-level rules of "first coverage maximization, then accuracy optimization";The reference light target is re-measured at the optimal parking point to establish a local high-precision coordinate system, and the batch measurement is executed after the global model is accurately aligned and converted;Only when the specific feature point measurement fails, the event-driven local dynamic recovery is triggered, and the measurement is completed by fine scanning, incremental modeling and online re-planning of the failed point.The application realizes high continuity, high success rate and strong robustness of automatic measurement in complex shielding environment with minimum calculation cost through "two-level modeling and two-level decision" architecture.

Description

Technical Field

[0001] This application relates to the fields of industrial automation and precision measurement technology, specifically to an intelligent obstacle avoidance planning and dynamic recovery method for achieving automatic, continuous, and reliable measurement of fixed optical measurement systems in complex scaffolding environments at manufacturing sites such as shipbuilding and large steel structure manufacturing sites. In particular, it relates to an intelligent obstacle avoidance planning and dynamic recovery method for automatic measurement of ship hull sections. Background Technology

[0002] In discrete manufacturing industries such as shipbuilding and large steel structures, automated 3D dimensional inspection of segments using fixed optical measurement systems (such as photogrammetry systems) is crucial for ensuring construction accuracy. These systems typically mount high-resolution cameras on gantry frames or columns, and the measurement is performed by moving the measuring head along a predetermined track. However, the dense scaffolding erected for subsequent work creates a complex network of static obstructions, posing a fundamental challenge to automated measurement: The environment is highly unstructured: the position, angle and combination of scaffolding members are varied and difficult to describe with a precise fixed model before measurement. This causes the preset measurement path based on the ideal CAD model to fail frequently due to obstructed view.

[0003] Limitations of existing obstacle avoidance strategies: Traditional solutions mainly fall into two categories: one is to rely entirely on human experience for online remote control or re-teaching, which is inefficient and prone to automation interruptions; the other is to use simple two-dimensional collision avoidance or preset safety zones, which cannot be combined with the specific observation angles, distances, and other constraints required for three-dimensional measurement tasks, often resulting in "avoiding obstacles but failing to measure the target".

[0004] The problem of discrepancies between the model and reality: Even if an initial environmental model is established through 3D scanning, its accuracy and completeness are limited, and there are unavoidable minor differences from the actual object. These differences are enough to cause occlusion during fine measurement, and existing systems lack the ability to detect and correct such discrepancies online and in real time during task execution.

[0005] Therefore, there is an urgent need for a measurement method that can understand the environment online, intelligently plan unobstructed observation paths, and dynamically correct model errors to ensure the continuity and success rate of the fully automated process. Summary of the Invention

[0006] The purpose of this invention is as follows: Establish a global spatial benchmark and a coarse environmental model: By setting benchmark targets on the hull sections, a global coordinate system is established, and a global coarse obstacle avoidance model containing the simplified structure of the hull and scaffolding is generated using fixed camera image reconstruction technology, which serves as the basis for intelligent decision-making.

[0007] Intelligent single-point decision-making for measurement start position: Based on the global model and preset candidate docking points, the system evaluates the "number of measurable points" and "comprehensive measurement accuracy potential" of each docking point in a forward-looking manner through parallel computing, and automatically selects the unique optimal docking point according to clear priority rules (coverage first, accuracy second), instructing the gantry to move into place in one go, replacing inefficient traversal search.

[0008] Achieve high-precision repositioning and model alignment in the local workspace: At the optimal docking point, establish a local high-precision measurement coordinate system that strictly corresponds to the current gantry pose by re-measuring the benchmark target, and transform the global model to this coordinate system to eliminate the errors accumulated due to gantry movement and model transformation, providing accurate spatial reference for the measurement of the current station position.

[0009] Achieve event-driven, localized fine-grained perception and online recovery: Localized fine-grained scanning of the failed point is triggered only when a planned measurement of a specific feature point fails. By discovering and modeling unforeseen minor obstacles, the obstacle avoidance model for the current station is dynamically updated, and the camera angle is fine-tuned to complete the remeasurement, ensuring process continuity with minimal cost.

[0010] This invention provides an intelligent obstacle avoidance planning and dynamic recovery method for automatic measurement of ship hull sections, comprising the following steps: Step 1: Establishing a global coordinate system and coarse environment modeling based on a benchmark target; Step 2: Global optimal station placement decision based on parallel computing; Step 3: Fine registration of the local coordinate system and execution of the measurement task; Step 4: Event-driven local dynamic recovery.

[0011] In some embodiments, step one, establishing a global coordinate system based on a reference target and coarse environment modeling, includes: At least three non-linear locations are selected at the rigid parts of the hull section structure, and coded markers with unique IDs are affixed as global reference targets. A gantry that can move along the track is controlled to move to an initial position with a wide field of view. The three-dimensional coordinates of the global reference targets are observed and accurately measured from multiple angles by a camera with at least two degrees of freedom mounted on the gantry. An initial global measurement coordinate system W-global is established through spatial analytical calculation, and the accurate transformation relationship with the hull section design coordinate system is established. At the same initial position, the camera is controlled to capture multiple images covering the entire workstation. The image sequence is processed based on the Structure for Motion Restoration (SfM) technique to generate global 3D point cloud data G-cloud covering the workstation. The G-cloud is then processed in real time, including background filtering and ROI extraction, hull segmentation and simplification, and scaffolding identification and modeling, ultimately generating a 3D obstacle avoidance model M-global for the current workstation.

[0012] In some embodiments, step two, the global optimal station location decision based on parallel computing, includes: Input the M-global model, the theoretical coordinates of all the feature points to be measured, and the N candidate stopping points {S1, S2, ..., SN} preset on the gantry track. The position of each candidate stopping point Sj in the gantry mechanical coordinate system has been calibrated and can be accurately transformed to the W-global coordinate system to obtain its corresponding theoretical camera position Tj-world. For each candidate docking point Sj, perform the following calculations independently and in parallel: Calculation of the measurable point set Fj: For each point Pi to be measured, based on the Tj-world and camera field of view model, the line of sight from the camera optical center to Pi is simulated in the M-global. If there exists at least one set of gimbal angles such that the line of sight does not intersect with any model primitive, then Pi is determined to be theoretically measurable at point Sj. All measurable points constitute the set Fj. Calculation of comprehensive accuracy evaluation value Qj: For each point Pi in Fj, among all feasible lines of sight, calculate the minimum Euclidean distance dij-min from the camera optical center to Pi. The comprehensive accuracy evaluation value Qj for this docking point is Qj = Σ{Pi ∈ Fj} (dij-min). The optimal stopping point S-optimal is selected from N candidate points according to the following two-level rules: Level 1 rule: Compare the number of measurable points |Fj| at each stop point, and select the stop point with the largest |Fj| value; Second-level rule: If multiple docking points have the same |Fj| value and all of them are the maximum, then select the docking point with the smallest comprehensive accuracy evaluation value Qj among these points; Generate control commands to drive the measuring gantry to move directly and in one go from its current position to the selected S-optimal position.

[0013] In some embodiments, step three, the execution of the local coordinate system fine registration and measurement task, includes: After the gantry is stabilized in S-optimal, the camera is controlled to re-photograph and accurately measure the global reference light target visible in the current field of view. Based on the known precise coordinates of the global reference light target in W-global and their high-precision observation values ​​in the current camera image, a new local measurement coordinate system W-local that strictly corresponds to the current gantry pose is recalculated and established through the spatial resection algorithm. Calculate the coordinate transformation matrix from W-global to W-local. Use this matrix to transform all geometric data stored in the M-global model under W-global to the W-local coordinate system to obtain the M-global-local model. Based on the pre-calculated image angles of each feature point in step two for S-optimal, the M-global-local model is used as the final obstacle avoidance basis to automatically and sequentially image and measure the feature points in the measurable point set F-optimal.

[0014] In some embodiments, step four, event-driven local dynamic recovery, includes: When the system takes a picture of a feature point Pk in the measurable point set F-optimal according to the plan, if the feature point is occluded or the image quality is poor due to slight deviations between the M-global-local model and the actual object on site, and the recognition fails, the recovery process is triggered immediately. Control the camera gimbal to take detailed photos of a small area where point Pk is located from multiple angles with high overlap. Quickly generate a local high-precision point cloud L-cloud for this area through stereo vision or local SfM. Analyze the L-cloud and identify new obstacles O-new that cause occlusion through geometric segmentation and fitting. Simplify them into geometric primitives to form a local incremental obstacle avoidance model M-local-increment. The M-local-increment is fused into the current M-global-local model in real time to generate an updated temporary fused obstacle avoidance model M-fused-local; based on M-fused-local, a feasible unobstructed shooting angle is quickly searched for feature point Pk at the current position of S-optimal. The camera gimbal is adjusted to the newly calculated angle, and Pk is re-measured. Regardless of whether the remeasurement is successful or not, the system records the state of the point. Then, the temporary fusion model M-fused-local is cleared, the M-global-local model is restored, and the measurement process of the next feature point in the F-optimal set is continued.

[0015] In some embodiments, the real-time point cloud processing of G-cloud in step one specifically includes: Background filtering and ROI extraction: Based on the known approximate location of the segments, a three-dimensional spatial region of interest (ROI) is set, and the point cloud outside the ROI is directly filtered out; Hull segmentation and simplification: The point cloud within the ROI is quickly and coarsely registered with the imported segmented CAD model. Point clouds with high matching degree with the model are classified as "hull" and the rest are classified as "non-hull". The detailed CAD model is simplified into a lightweight shell model that retains only the main outer plates and key port structures for subsequent collision detection and line-of-sight analysis. Scaffolding identification and modeling: For "non-hull" point clouds, using the prior knowledge of the regular grid structure of scaffolding, wireframe extraction and model fitting algorithms are used to simplify it into a simplified frame model composed of main members; Dynamic obstacle avoidance model generation: The simplified ship hull model and the scaffolding frame model are merged in a unified coordinate system to generate the current workstation's three-dimensional obstacle avoidance model M-global.

[0016] In some embodiments, the preset method for the candidate docking points {S1, S2, ..., SN} in step two is as follows: Based on the current position of the gantry, a stopping point is set at regular intervals along the direction of the gantry's movement track, eliminating the need for operators to set it every time.

[0017] In some embodiments, the photographing angles of each feature point pre-calculated for S-optimal in step two, as described in step three, need to be fine-tuned according to the W-local coordinate system.

[0018] In some embodiments, the step four of simplifying the new obstacle O-new into a geometric primitive includes a thin cylinder, a small thin plate, or a combination thereof.

[0019] In some embodiments, step four involves quickly searching for unobstructed feasible shooting angles for feature point Pk based on M-fused-local; The M-local-increment can be saved to an experience base for future optimization of the initial model at similar workstations.

[0020] Compared with the prior art, the present invention has the following advantages: High continuity and success rate: Through a two-layer mechanism of "intelligent global planning" and "precise local recovery", the measurement interruption problem in complex occlusion environment is effectively solved, and the near-fully automated operation without human intervention is achieved.

[0021] Extremely high decision-making efficiency: Positioning decisions based on parallel computing and explicit optimization rules enable the gantry to reach the optimal position with the fewest moves (usually 1-2 times), greatly reducing idle waiting time and improving overall efficiency.

[0022] Strong adaptability: The event-triggered local recovery mechanism enables the system to automatically handle unforeseen minor occlusions, significantly enhancing robustness and reducing dependence on the absolute accuracy of environmental modeling.

[0023] Highly practical: The core algorithm is based on the system's inherent camera, eliminating the need for expensive dedicated sensors, making the solution easy to implement and promote. Attached Figure Description

[0024] The accompanying drawings illustrate, by way of example and not limitation, the various embodiments discussed herein.

[0025] Figure 1 This is a schematic diagram of the measurement system; Figure 2 This is a flowchart of a partial dynamic response process; Figure 3 A schematic diagram of the intelligent obstacle avoidance planning and dynamic recovery method in the automatic measurement of ship hull sections. Detailed Implementation

[0026] In order to gain a more detailed understanding of the features and technical content of the embodiments of this application, the implementation of the embodiments of this application will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for reference and illustration only and are not intended to limit the embodiments of this application.

[0027] In the embodiments described in this application, it should be noted that, unless otherwise stated and limited, the term "connection" should be interpreted broadly. For example, it can be an electrical connection, or a connection between two internal components. It can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above term according to the specific circumstances.

[0028] It should be noted that the terms "first," "second," and "third" used in the embodiments of this application are merely used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first," "second," and "third" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first," "second," and "third" can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in an order other than those illustrated or described herein.

[0029] This invention provides an intelligent obstacle avoidance planning and dynamic recovery method for automatic measurement of ship hull sections, such as... Figure 3 As shown, it includes the following steps: Step 1: Establishing a global coordinate system and coarse environment modeling based on the benchmark target.

[0030] Step 2: Global optimal station location decision based on parallel computing.

[0031] Step 3: Local coordinate system fine registration and measurement task execution.

[0032] Step 4: Event-driven local dynamic recovery.

[0033] In some embodiments, step one, establishing a global coordinate system and coarse environment modeling based on a reference target, includes: At least three non-linear locations are selected on the rigid parts of the hull section structure, and coded markers with unique IDs are affixed as global reference targets. A gantry that can move along a track is controlled to move to an initial position with a wide field of view. The three-dimensional coordinates of the global reference targets are observed and accurately measured from multiple angles using cameras with at least two degrees of freedom mounted on the gantry. An initial global measurement coordinate system W-global is established through spatial analytical calculations, and a precise transformation relationship with the hull section design coordinate system is established.

[0034] At the same initial position, the camera captures multiple images covering the entire workstation. The image sequence is processed using Structure for Motion Restoration (SfM) technology to generate global 3D point cloud data G-cloud covering the workstation. Real-time point cloud processing is performed on G-cloud, including background filtering and ROI extraction, hull segmentation and simplification, and scaffolding identification and modeling, ultimately generating a 3D obstacle avoidance model M-global for the current workstation.

[0035] In some embodiments, step two, the global optimal station location decision based on parallel computing, includes: Input the M-global model, the theoretical coordinates of all the feature points to be measured, and the N candidate stopping points {S1, S2, ..., SN} preset on the gantry track. The position of each candidate stopping point Sj in the gantry mechanical coordinate system has been calibrated and can be accurately transformed to the W-global coordinate system to obtain its corresponding theoretical camera position Tj-world.

[0036] For each candidate docking point Sj, perform the following calculations independently and in parallel: Calculation of the measurable point set Fj: For each point Pi to be measured, based on the Tj-world and camera field of view model, the line of sight from the camera optical center to Pi is simulated in the M-global. If there exists at least one set of gimbal angles such that the line of sight does not intersect with any model primitive, then Pi is determined to be theoretically measurable at point Sj. All measurable points constitute the set Fj.

[0037] Calculation of the overall accuracy evaluation value Qj: For each point Pi in Fj, calculate the minimum Euclidean distance dij-min from the camera optical center to Pi among all feasible lines of sight. The overall accuracy evaluation value of the docking point is Qj = Σ{Pi ∈ Fj} (dij-min).

[0038] The optimal stopping point S-optimal is selected from N candidate points according to the following two-level rules: Level 1 rule, maximizing coverage: compare the number of measurable points |Fj| at each stop and select the stop with the largest |Fj| value.

[0039] The second-level rule, accuracy optimization: If multiple docking points have the same |Fj| value and all of them are the maximum, then select the docking point with the smallest comprehensive accuracy evaluation value Qj among these points.

[0040] Generate control commands to drive the measuring gantry to move directly and in one go from its current position to the selected S-optimal position.

[0041] In some embodiments, step three, the execution of local coordinate system fine registration and measurement tasks, includes: After the gantry stabilizes in S-optimal, the camera is controlled to re-photograph and accurately measure the global reference target visible in the current field of view. Based on the known precise coordinates of the global reference target in W-global and their high-precision observations in the current camera image, a new local measurement coordinate system W-local, which strictly corresponds to the current gantry pose, is recalculated and established through the spatial resection algorithm.

[0042] Calculate the coordinate transformation matrix from W-global to W-local. Use this matrix to transform all geometric data stored in the M-global model under W-global to the W-local coordinate system, thus obtaining the M-global-local model.

[0043] Based on the pre-calculated image angles of each feature point in step two for S-optimal, the M-global-local model is used as the final obstacle avoidance basis to automatically and sequentially image and measure the feature points in the measurable point set F-optimal.

[0044] In some embodiments, step four, event-driven local dynamic recovery, includes: When the system takes a picture of a feature point Pk in the measurable point set F-optimal according to the plan, if the feature point is occluded or the image quality is poor due to slight deviations between the M-global-local model and the actual object on site, and the recognition fails, the recovery process is immediately triggered.

[0045] The camera gimbal is controlled to take detailed, multi-angle, high-overlapping photos of a small area containing point Pk. A high-precision local point cloud (L-cloud) for this area is quickly generated using stereo vision or local SfM. The L-cloud is analyzed, and through geometric segmentation and fitting, new obstacles (O-new) causing occlusion are identified and simplified into geometric primitives to form a local incremental obstacle avoidance model (M-local-increment).

[0046] The M-local-increment is fused into the current M-global-local model in real time to generate an updated temporary fused obstacle avoidance model, M-fused-local. Based on M-fused-local, a feasible unobstructed shooting angle is quickly searched for feature point Pk at the current S-optimal position.

[0047] The camera gimbal is adjusted to the newly calculated angle, and Pk is re-measured. Regardless of whether the remeasurement is successful or not, the system records the state of that point. Then, the temporary fusion model M-fused-local is cleared, the M-global-local model is restored, and the measurement process for the next feature point in the F-optimal set continues.

[0048] In some embodiments, step one, which involves real-time point cloud processing of G-cloud, specifically includes: Background filtering and ROI extraction: Based on the known approximate location of the segments, a 3D spatial region of interest (ROI) is set, and the point cloud outside the ROI is directly filtered out.

[0049] Hull segmentation and simplification: The point cloud within the ROI is quickly and coarsely registered with the imported segmented CAD model. Point clouds with high matching degree to the model are classified as "hull" and the rest are classified as "non-hull". The detailed CAD model is simplified into a lightweight shell model that retains only the main outer plates and key port structures for subsequent collision detection and line-of-sight analysis.

[0050] Scaffolding identification and modeling: For "non-hull" point clouds, using the structural prior of regular grid-like scaffolding, wireframe extraction and model fitting algorithms are used to simplify them into a simplified frame model composed of main members.

[0051] Dynamic obstacle avoidance model generation: The simplified ship hull model and the scaffolding frame model are merged in a unified coordinate system to generate the current workstation's three-dimensional obstacle avoidance model M-global.

[0052] In some embodiments, the preset method for selecting candidate docking points {S1, S2, ..., SN} in step two is as follows: Based on the current position of the gantry, a stopping point is set at regular intervals along the direction of the gantry's movement track, eliminating the need for operators to set it every time.

[0053] In some embodiments, in step three, the image angles of each feature point pre-calculated for S-optimal in step two need to be fine-tuned according to the W-local coordinate system.

[0054] In some embodiments, in step four, the newly added obstacle O-new is simplified into a geometric primitive, which includes a thin cylinder, a small thin plate, or a combination thereof.

[0055] In some embodiments, step four involves quickly searching for unobstructed feasible shooting angles for feature point Pk based on M-fused-local.

[0056] M-local-increment can be saved to an experience library for use in optimizing initial models for similar workstations in the future.

[0057] Compared with the prior art, the present invention has the following advantages: High continuity and success rate: Through a two-layer mechanism of "intelligent global planning" and "precise local recovery", the measurement interruption problem in complex occlusion environment is effectively solved, and the near-fully automated operation without human intervention is achieved.

[0058] Extremely high decision-making efficiency: Positioning decisions based on parallel computing and explicit optimization rules enable the gantry to reach the optimal position with the fewest moves (usually 1-2 times), greatly reducing idle waiting time and improving overall efficiency.

[0059] Strong adaptability: The event-triggered local recovery mechanism enables the system to automatically handle unforeseen minor occlusions, significantly enhancing robustness and reducing dependence on the absolute accuracy of environmental modeling.

[0060] Highly practical: The core algorithm is based on the system's inherent camera, eliminating the need for expensive dedicated sensors, making the solution easy to implement and promote.

[0061] This invention constructs a closed-loop system consisting of "global perception and planning → intelligent decision-making and localization → local execution and recovery." For example... Figure 1 As shown, a fixed measurement system includes at least a gantry that can move along a track, and a camera with at least two degrees of freedom (pitch and yaw) mounted on the gantry. The specific technical approach is as follows: global coordinate system establishment and coarse environment modeling based on a benchmark target → global optimal station location decision based on parallel computing → local coordinate system fine registration and measurement task execution → event-driven local dynamic recovery.

[0062] I. Establishment of Global Coordinate System and Coarse Environment Modeling Based on Reference Target After the hull sections are in place, at least three non-aligned locations are selected at the structural rigid parts of each section, and coded markers with unique IDs are affixed as global reference targets. The control gantry is moved to an initial position with a wide field of view, and the three-dimensional coordinates of these global reference targets are observed and precisely measured from multiple angles using a camera pan-tilt unit. Through spatial analytical calculations, an initial global measurement coordinate system (denoted as W-global) is established, and a precise transformation relationship is established between this coordinate system and the design coordinate system (CAD coordinate system) of the hull sections.

[0063] At the same initial position, the camera captures multiple images covering the entire workstation (including the hull and scaffolding). Based on Structure of Motion (SfM) technology, the image sequence is processed to generate global 3D point cloud data (denoted as G-cloud) covering the workstation. Subsequently, real-time point cloud processing is performed on the G-cloud. ① Background filtering and ROI extraction: Based on the approximate location of the known segments, a 3D spatial region of interest (ROI) is defined. Point clouds outside the ROI (such as distant walls and the ground) are directly filtered out.

[0064] ②. Hull Segmentation and Simplification: The point cloud within the ROI is quickly and coarsely registered with the imported segmented, detailed CAD model. Point clouds with high model matching are categorized as "hull," while the rest are categorized as "non-hull (mainly scaffolding)." To further optimize calculations, the detailed CAD model is simplified into a lightweight, empty shell model retaining only the main outer plates, ports, and other key structural elements for subsequent collision detection and line-of-sight analysis.

[0065] ③. Scaffolding identification and modeling: For the "non-hull" point cloud, using the structural prior of the regular grid of the scaffolding, wireframe extraction and model fitting algorithms are used to simplify it into a simplified frame model composed of the main members (uprights, horizontals, and diagonal braces).

[0066] ④. Dynamic obstacle avoidance model generation: The simplified ship hull model and scaffolding frame model are merged in a unified coordinate system to generate a 3D obstacle avoidance model for the current workstation. This model clearly defines the inaccessible physical obstacle areas.

[0067] II. Global Optimal Station Position Decision Based on Parallel Computing The purpose of this step is to use a global model to determine the optimal working position that the gantry should move to through calculation rather than traversal.

[0068] 1. Input: M-global model, theoretical coordinates of all feature points to be measured (the location information of the feature points is included in the structural model), and N candidate stopping points {S1, S2, ..., SN} preset on the gantry track (the stopping points are preset based on the current position of the gantry, and a stopping point is set at regular intervals along the direction of the gantry movement track (forward and backward), without requiring the operator to set them each time). The position of each stopping point Sj in the gantry mechanical coordinate system has been calibrated and can be accurately transformed to the W-global coordinate system to obtain its corresponding theoretical camera position Tj-world.

[0069] 2. Parallelization measurability analysis and accuracy pre-assessment: For each candidate docking point Sj, the system performs the following calculations independently and in parallel: ①. Calculation of the measurable point set Fj: For each point Pi to be measured, based on the Tj-world and camera field-of-view model, simulate the line of sight from the camera optical center to Pi in the M-global. If there exists at least one set of gimbal angles such that the line of sight does not intersect with any model primitive, then Pi is determined to be theoretically measurable at point Sj. All measurable points constitute the set Fj.

[0070] ②. Calculation of the overall accuracy assessment value Qj: For each point Pi in Fj, calculate the minimum Euclidean distance dij-min from the camera's optical center to Pi across all feasible lines of sight. The overall accuracy assessment value Qj for this docking point is Qj = Σ{Pi ∈ Fj} (dij-min). The smaller the Qj value, the shorter the expected average observation distance when measuring the measurable point set at that point, the higher the theoretical overall accuracy, and the shorter the image capture time.

[0071] 3. Optimal docking point decision: The system selects a unique optimal point from N candidate points based on the following two-level rules: ①. First-level rule (maximizing coverage): Compare the number of measurable points |Fj| at each stop and select the stop with the largest |Fj| value.

[0072] ②. Second-level rule (accuracy optimization): If multiple stops have the same |Fj| value and all of them are the maximum, then select the stop with the smallest comprehensive accuracy evaluation value Qj among these stops.

[0073] The decision output is the optimal stopping point S-optimal.

[0074] 4. Single positioning movement of the gantry: The system generates control commands to drive the measuring gantry to move directly and in one go from its current position to the selected S-optimal position. After this step, the gantry no longer needs to traverse other docking points.

[0075] III. Local Coordinate System Fine-tuning and Measurement Task Execution This step is performed after the gantry arrives at S-optimal. The core of this step is to establish a high-precision, error-free local working reference at the new physical location and then begin measurement.

[0076] 1. Establishment of a local high-precision coordinate system: After the gantry stabilizes in S-optimal, the system controls the camera to re-image and precisely measure the global reference targets (at least three) visible within the current field of view. Based on the known precise coordinates of these targets in W-global and their high-precision observations in the current camera image, a completely new local measurement coordinate system (denoted as W-local) that strictly corresponds to the current gantry pose is recalculated and established using algorithms such as spatial resection. W-local has higher local alignment accuracy than the theoretical position obtained through mechanical coordinate transformation.

[0077] 2. Precise alignment conversion of obstacle avoidance model: The system calculates the coordinate transformation matrix from W-global to W-local. Using this matrix, all geometric data stored in the M-global model under W-global is uniformly transformed to the W-local coordinate system, resulting in M-global-local. This operation eliminates the accumulated errors caused by gantry movement and model transformation, enabling precise alignment between the environmental model and the current actual measurement space.

[0078] 3. Guided batch measurement: Based on the pre-calculated image angles of each feature point in step two for S-optimal (these angles need to be fine-tuned according to the W-local coordinate system), the system uses the M-global-local model as the final obstacle avoidance basis and begins to automatically and sequentially take images of the feature points in the measurable point set F-optimal.

[0079] IV. Event-driven local dynamic recovery This step serves as a safeguard mechanism for step three and is dynamically triggered during the measurement execution process.

[0080] 1. Failure detection and triggering: When the system takes a picture of a feature point Pk in the measurable point set F-optimal according to the plan, if the feature point is occluded or the image quality is poor due to slight deviations between the M-global-local model and the actual site (such as scaffolding structures not captured by the global scan), resulting in recognition failure, this recovery process will be triggered immediately. If the measurement is successful, this process will not be initiated.

[0081] 2. Local fine-grained perception and incremental modeling: The system controls the camera gimbal to take detailed, multi-angle, high-overlapping photos of a small area containing point Pk. Using stereo vision or local SfM, a high-precision local point cloud (L-cloud) is quickly generated for this area. Analyzing the L-cloud, through geometric segmentation and fitting, new obstacles O-new causing occlusion are identified and simplified into a geometric primitive (such as a thin cylinder, a small thin plate, or a combination thereof), forming a local incremental obstacle avoidance model M-local-increment.

[0082] 3. Dynamic model fusion and online replanning: The M-local-increment is fused into the current M-global-local model in real time to generate an updated and more accurate temporary fused obstacle avoidance model, M-fused-local. Based on M-fused-local, the system quickly searches for an unobstructed feasible shooting angle for feature point Pk at the current S-optimal position (usually only requiring minor adjustments to the gimbal's pitch and yaw angles).

[0083] 4. Supplementary testing and continued process: like Figure 2 As shown, the camera gimbal is adjusted to the newly calculated angle, and Pk is re-measured. Regardless of whether the remeasurement is successful or not, the system records the state of that point. Subsequently, the temporary fused model M-fused-local is cleared, and the M-global-local model is restored to continue the measurement process for the next feature point in the F-optimal set. M-local-increment can be saved to the experience base for future optimization of the initial model at similar workstations.

[0084] The embodiments of the present invention provide an overall architecture of "two-level modeling and two-level decision-making": Global coarse modeling and pre-decision-making: At the starting point, a simplified obstacle model for prospective site selection is quickly constructed through a single large-scale photograph and simplified fitting. Local fine calibration and dynamic recovery: At the optimal site, the model is precisely aligned through local relocalization; ultra-local fine sensing and incremental updates are triggered only when measurement fails. This architecture balances efficiency and accuracy. It uses a "lightweight global model" to support efficient computation and "precise local actions" to solve practical problems, avoiding reliance on high-precision global scanning and the burden of detailed computation throughout the entire process.

[0085] The above embodiments of the present invention provide an intelligent station location decision-making method based on parallelized look-ahead computing: two quantitative indicators, "number of measurable points (Fj)" and "comprehensive accuracy coefficient (Qj)," are calculated in parallel for multiple preset docking points, and a unique optimal station location is automatically selected according to the determination rule of "maximizing Fj first, then minimizing Qj." This elevates gantry movement from "manual trial and error" or "simple traversal" to data-driven intelligent optimal decision-making, representing a crucial leap towards achieving efficient automation.

[0086] The embodiments of the present invention provide a "failure-triggered" local dynamic recovery mechanism: only when the planned measurement of a specific point fails, is a local fine-scan, incremental modeling, online replanning, and supplementary measurement initiated for that point. After completion, the original model is immediately restored. This is an extremely efficient strategy of "providing accuracy on demand." It solves the real-world problem of models always having subtle errors, achieving strong robustness and self-healing capabilities of the system with minimal computational cost, and is fundamentally different from "periodic global rescanning" or "full-process local fine-scanning."

[0087] The embodiments of the present invention provide a simplified geometric representation of obstacles suitable for collision detection: a combined model that fits a chaotic point cloud (especially scaffolding) into simple geometric primitives such as cylinders and cuboids in real time. This is the computational foundation for all subsequent "fast line-of-sight collision detection" and "real-time path planning". The lightweight model ensures the real-time performance of the entire closed-loop system (especially parallel computing and online replanning).

[0088] The technical solutions described in the embodiments of this application can be combined arbitrarily without conflict.

[0089] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent obstacle avoidance planning and dynamic recovery in automatic measurement of ship hull sections, characterized in that, Includes the following steps: Step 1: Establishing a global coordinate system and coarse environment modeling based on a benchmark target; Step 2: Global optimal station placement decision based on parallel computing; Step 3: Fine registration of the local coordinate system and execution of the measurement task; Step 4: Event-driven local dynamic recovery.

2. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of ship hull sections according to claim 1, characterized in that, Step one, establishing a global coordinate system and coarse environment modeling based on a benchmark target, includes: At least three non-linear locations are selected at the rigid parts of the hull section structure, and coded markers with unique IDs are affixed as global reference targets. A gantry that can move along the track is controlled to move to an initial position with a wide field of view. The three-dimensional coordinates of the global reference targets are observed and accurately measured from multiple angles by a camera with at least two degrees of freedom mounted on the gantry. An initial global measurement coordinate system W-global is established through spatial analytical calculation, and the accurate transformation relationship with the hull section design coordinate system is established. At the same initial position, the camera is controlled to capture multiple images covering the entire workstation. The image sequence is processed based on the Structure for Motion Restoration (SfM) technique to generate global 3D point cloud data G-cloud covering the workstation. The G-cloud is then processed in real time, including background filtering and ROI extraction, hull segmentation and simplification, and scaffolding identification and modeling, ultimately generating a 3D obstacle avoidance model M-global for the current workstation.

3. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of hull sections according to claim 2, characterized in that, Step two, the global optimal station location decision based on parallel computing, includes: Input the M-global model, the theoretical coordinates of all the feature points to be measured, and the N candidate stopping points {S1, S2, ..., SN} preset on the gantry track. The position of each candidate stopping point Sj in the gantry mechanical coordinate system has been calibrated and can be accurately transformed to the W-global coordinate system to obtain its corresponding theoretical camera position Tj-world. For each candidate docking point Sj, perform the following calculations independently and in parallel: Calculation of the measurable point set Fj: For each point Pi to be measured, based on the Tj-world and camera field of view model, the line of sight from the camera optical center to Pi is simulated in the M-global. If there exists at least one set of gimbal angles such that the line of sight does not intersect with any model primitive, then Pi is determined to be theoretically measurable at point Sj. All measurable points constitute the set Fj. Calculation of comprehensive accuracy evaluation value Qj: For each point Pi in Fj, among all feasible lines of sight, calculate the minimum Euclidean distance dij-min from the camera optical center to Pi. The comprehensive accuracy evaluation value Qj for this docking point is Qj = Σ{Pi ∈ Fj} (dij-min). The optimal stopping point S-optimal is selected from N candidate points according to the following two-level rules: Level 1 rule: Compare the number of measurable points |Fj| at each stop point, and select the stop point with the largest |Fj| value; Second-level rule: If multiple docking points have the same |Fj| value and all of them are the maximum, then select the docking point with the smallest comprehensive accuracy evaluation value Qj among these points; Generate control commands to drive the measuring gantry to move directly and in one go from its current position to the selected S-optimal position.

4. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of hull sections according to claim 3, characterized in that, Step three, the local coordinate system fine registration and measurement task execution, includes: After the gantry is stabilized in S-optimal, the camera is controlled to re-photograph and accurately measure the global reference light target visible in the current field of view. Based on the known precise coordinates of the global reference light target in W-global and their high-precision observation values ​​in the current camera image, a new local measurement coordinate system W-local that strictly corresponds to the current gantry pose is recalculated and established through the spatial resection algorithm. Calculate the coordinate transformation matrix from W-global to W-local. Use this matrix to transform all geometric data stored in the M-global model under W-global to the W-local coordinate system to obtain the M-global-local model. Based on the pre-calculated image angles of each feature point in step two for S-optimal, the M-global-local model is used as the final obstacle avoidance basis to automatically and sequentially image and measure the feature points in the measurable point set F-optimal.

5. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of hull sections according to claim 4, characterized in that, Step four, event-driven local dynamic recovery, includes: When the system takes a picture of a feature point Pk in the measurable point set F-optimal according to the plan, if the feature point is occluded or the image quality is poor due to slight deviations between the M-global-local model and the actual object on site, and the recognition fails, the recovery process is triggered immediately. Control the camera gimbal to take detailed photos of a small area where point Pk is located from multiple angles with high overlap. Quickly generate a local high-precision point cloud L-cloud for this area through stereo vision or local SfM. Analyze the L-cloud and identify new obstacles O-new that cause occlusion through geometric segmentation and fitting. Simplify them into geometric primitives to form a local incremental obstacle avoidance model M-local-increment. The M-local-increment is fused into the current M-global-local model in real time to generate an updated temporary fused obstacle avoidance model M-fused-local; based on M-fused-local, a feasible unobstructed shooting angle is quickly searched for feature point Pk at the current position of S-optimal. The camera gimbal is adjusted to the newly calculated angle, and Pk is re-measured. Regardless of whether the remeasurement is successful or not, the system records the state of the point. Then, the temporary fusion model M-fused-local is cleared, the M-global-local model is restored, and the measurement process of the next feature point in the F-optimal set is continued.

6. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of hull sections according to claim 5, characterized in that, The real-time point cloud processing of G-cloud described in step one specifically includes: Background filtering and ROI extraction: Based on the known approximate location of the segments, a three-dimensional spatial region of interest (ROI) is set, and the point cloud outside the ROI is directly filtered out; Hull segmentation and simplification: The point cloud within the ROI is quickly and coarsely registered with the imported segmented CAD model. Point clouds with high matching degree with the model are classified as "hull" and the rest are classified as "non-hull". The detailed CAD model is simplified into a lightweight shell model that retains only the main outer plates and key port structures for subsequent collision detection and line-of-sight analysis. Scaffolding identification and modeling: For "non-hull" point clouds, using the prior knowledge of the regular grid structure of scaffolding, wireframe extraction and model fitting algorithms are used to simplify it into a simplified frame model composed of main members; Dynamic obstacle avoidance model generation: The simplified ship hull model and the scaffolding frame model are merged in a unified coordinate system to generate the current workstation's three-dimensional obstacle avoidance model M-global.

7. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of hull sections according to claim 5, characterized in that, The preset method for the candidate docking points {S1, S2, ..., SN} mentioned in step two is as follows: Based on the current position of the gantry, a stopping point is set at regular intervals along the direction of the gantry's movement track, eliminating the need for operators to set it every time.

8. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of hull sections according to claim 5, characterized in that, The image angles for each feature point, which were pre-calculated for S-optimal in step two, as described in step three, need to be fine-tuned according to the W-local coordinate system.

9. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of hull sections according to claim 5, characterized in that, Step four describes simplifying the newly added obstacle O-new into geometric primitives, which include thin cylinders, small thin plates, or combinations thereof.

10. The intelligent obstacle avoidance planning and dynamic recovery method in automatic measurement of hull sections according to claim 5, characterized in that, Step four describes the method of quickly searching for unobstructed feasible shooting angles for feature point Pk based on M-fused-local. The M-local-increment can be saved to an experience base for future optimization of the initial model at similar workstations.

Images