A Method for Intelligent Posture Correction of Inspection Robots Based on AI Vision

By using Spherelets and a visual response dynamic correction method, the problem of accurate posture correction for inspection robots in complex terrain was solved. This enabled accurate differentiation and correction of terrain-adaptive postures, improving the environmental adaptability and visual observation stability of the inspection robot.

CN123118764APending Publication Date: 2026-11-17XINXINXING (TIANJIN) CLOUD TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611362322.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing inspection robots struggle to accurately distinguish between terrain-adaptive posture changes and target observation deviations in complex terrains, leading to overcorrection or repeated adjustments, and lacking effective posture correction methods.

Method used

Spherelets are used to represent robot posture changes. Combined with visual response, the posture is dynamically corrected. An effective posture channel is established through the posture Spherelet sphere chain, the effective spherical posture range is defined, and posture projection is performed while keeping the position of the terrain process unchanged. Lateral posture offset is extracted to form inspection interference posture components for posture correction.

Benefits of technology

It improves the accuracy of posture correction and environmental adaptability of inspection robots in complex terrain, reduces the probability of miscorrection of posture when adapting to normal terrain, and enhances the stability of visual observation and the reliability of posture control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118764A_ABST
    Figure CN123118764A_ABST
Patent Text Reader

Abstract

This invention discloses an AI vision-based intelligent posture correction method for inspection robots, comprising the following steps: acquiring continuous inspection images and robot motion data to form a visual posture observation sequence; establishing a terrain process, dividing posture curvature intervals, and forming a posture curvature interval sequence; constructing a local posture sphere to form a posture Spherelet chain; establishing a mapping between inspection observation relationships and spherical postures to form an effective posture channel; establishing a posture cross-section in the same process, performing Spherelet posture projection to form inspection interference posture components; determining the target correction posture, performing posture correction, and correcting the posture Spherelet chain and effective posture channel to form a posture correction execution result; and obtaining the intelligent posture correction result of the inspection robot based on the posture correction execution result. This invention improves the posture correction accuracy and environmental adaptability of inspection robots.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and in particular to an AI vision-based intelligent posture correction method for inspection robots. Background Technology

[0002] With the increasing demand for automated inspections in substations, industrial plants, mines, warehouse passages, and equipment rooms, inspection robots are increasingly employing visual perception, posture detection, and motion control technologies to complete target identification, path movement, and posture adjustment. Existing inspection robots typically acquire robot posture and surrounding environment information through inertial measurement units, encoders, visual cameras, or terrain sensing devices. They then generate posture correction parameters based on pitch angle, roll angle, heading angle, or the target's positional deviation in the image, ensuring the robot body or visual acquisition device maintains a preset posture. Some solutions also utilize the visual target position, ground normal direction, or local slope to adjust posture control parameters, improving motion stability and target observation capabilities in complex terrain.

[0003] However, during the movement of inspection robots on slopes, across undulating terrain, over obstacles, and turning, they naturally undergo posture migrations that adapt to the changing terrain. Existing methods typically compare the current posture with a fixed reference posture, making it difficult to distinguish between normal posture changes resulting from terrain adaptation and abnormal posture changes that actually cause deviations in the observed inspection target. This can easily lead to misjudging normal pitch and roll as deviations to be corrected, resulting in overcorrection or repeated adjustments. Furthermore, current posture correction methods often lack a geometric structure that continuously describes the posture curvature relationship along the robot's motion. The relationship between visual observation and robot posture is often treated using instantaneous deviations, making it difficult to define the permissible posture range that meets the effective observation requirements of the inspection target at the same terrain location. When the robot's posture is simultaneously affected by terrain changes and inspection observation needs, it is difficult to accurately separate the terrain-adapted posture that needs to be retained from the inspection interference posture that needs to be eliminated.

[0004] Therefore, how to provide an AI vision-based intelligent posture correction method for inspection robots is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI vision-based intelligent posture correction method for inspection robots. This invention uses a posture Spherelet chain and effective posture channels to represent the robot's posture migration caused by terrain changes. It also utilizes in-process posture projection to separate terrain-adaptive postures from inspection-interference postures, and combines this with dynamic correction of the posture representation structure based on visual response. This reduces the probability of miscorrection of normal terrain-adaptive postures, improving the accuracy of posture correction, environmental adaptability, and visual observation stability of inspection targets. An AI vision-based intelligent posture correction method for inspection robots according to an embodiment of the present invention includes the following steps: Collect continuous inspection images and robot motion data during the inspection process of the inspection robot, construct visual posture association information, and form a visual posture observation sequence; The terrain process is established based on the visual attitude observation sequence, and the attitude curvature interval is divided according to the change relationship of visual attitude correlation information along the terrain process to form an attitude curvature interval sequence. For the attitude curvature interval sequence, Spherelets are called to construct local attitude spheres. The position of the sphere center and the radius of the sphere are constrained according to the curvature change. The tangential continuity of adjacent local attitude spheres is maintained to form an attitude Spherelet sphere chain. Based on the attitude Spherelet spherical chain, a mapping between the inspection observation relationship and the spherical attitude is established, the effective spherical attitude range is defined and connected along the terrain process to form an effective attitude channel; Based on the effective attitude channel, establish the same-process attitude section, perform Spherelets attitude projection while keeping the terrain process position unchanged, retain the terrain-adaptive attitude, extract the lateral attitude offset, and form the inspection disturbance attitude component. The target correction posture is determined based on the inspection interference posture components and effective posture channels, posture correction is performed, and the posture Spherelet chain and effective posture channels are modified according to the corrected visual response to form the posture correction execution result. The intelligent posture correction result of the inspection robot is obtained based on the posture correction execution result.

[0006] Optionally, obtaining the visual pose observation sequence specifically includes: Collect continuous inspection images and robot motion data during the inspection process of the inspection robot, perform time-series correspondence between the continuous inspection images and robot motion data, and extract the inspection target area and local support area from the continuous inspection images. Based on the inspection target area, local support area and robot motion data, the visual orientation change of the inspection target, the orientation change of the local support surface and the robot posture change are determined, the correspondence between the three is established and the visual posture association information is obtained. Visual attitude association information is arranged according to the order of movement of the inspection robot along the inspection path, and visual attitude association information corresponding to adjacent inspection movement positions is associated to form a visual attitude observation sequence.

[0007] Optionally, obtaining the attitude bending interval sequence specifically includes: Based on the order of the movement positions of the inspection robot in the visual posture observation sequence, the inspection start position is selected as the starting point of the terrain process. The movement displacement between adjacent movement positions is extracted along the inspection path. The movement displacement is continuously accumulated according to the direction of the inspection path to determine the terrain process position corresponding to each movement position and establish the terrain process. The visual pose association information of adjacent positions is compared along the terrain process, and the interval connection position is determined based on the changing trends of local support surface direction, robot pose and visual orientation of the inspection target. The terrain process is segmented based on the connection position of the intervals. The posture curvature intervals are determined by combining the continuous change characteristics of visual posture association information within each segment. The continuity of the boundary between adjacent posture curvature intervals is checked, and the posture curvature intervals are organized according to the order of the terrain process to form a posture curvature interval sequence.

[0008] Optionally, obtaining the attitude Spherelet chain specifically includes: According to the terrain process order of the attitude curvature interval sequence, extract the visual attitude association information corresponding to the first attitude curvature interval, call Spherelets to determine the local attitude subspace and perform spherical initialization to obtain the initial local attitude sphere; Extract the exit attitude point and exit tangent corresponding to the interval connection position of the previous local attitude sphere, and use the exit attitude point and exit tangent as the entry constraint of the adjacent attitude bending interval. At the same time, establish curvature constraints based on the degree and direction of rotation of the local support surface along the terrain process in the adjacent attitude bending interval. Use the entry constraint and curvature constraint to limit the solution range of the center and radius of the local attitude sphere, and form the local sphere constraint conditions. Based on the local spherical constraints, conditional Spherelets spherical solutions are performed on adjacent attitude curvature intervals. Local spherical candidates are generated within the allowable spherical radius range and the allowable direction of the spherical center. The fitting status of the local spherical candidates and the attitude data is compared with the visual attitude association information of the corresponding attitude curvature interval. Local spherical candidates that violate the entry constraints, curvature constraints and tangential continuity requirements are filtered out to determine the current local attitude spherical. Extract the exit attitude point and exit tangent corresponding to the connection position of the current local attitude sphere in the next attitude curvature interval, transfer the exit attitude point and exit tangent to the adjacent attitude curvature interval, and repeat the local sphere constraint and condition Spherelet sphere solution in combination with the terrain curvature corresponding to the adjacent attitude curvature interval. Determine the local attitude sphere recursively along the attitude curvature interval sequence, and connect the local attitude spheres in the order of terrain progress to form the attitude Spherelet sphere chain.

[0009] Optionally, obtaining the effective attitude channel specifically includes: Based on the attitude Spherelet spherical chain, the attitude position on the local attitude sphere is sequentially associated with the visual orientation of the corresponding inspection target, and the association results are organized along the attitude migration direction of the local attitude sphere to establish a spherical observation mapping. Based on the spherical observation mapping, the visual orientation of the inspection target is tracked along the attitude migration direction of each local attitude sphere, and the spherical boundary position corresponding to the inspection target entering and leaving the effective observation area is determined. The effective spherical attitude range on each local attitude sphere is defined by the spherical boundary position. Connect the effective spherical attitude ranges on adjacent local attitude spheres in the order of terrain progress, take the connection attitude corresponding to the interval connection position as the range transfer reference, track the continuous migration of the two sides of the effective spherical attitude range along the attitude Spherelet spherical chain, check the effective range connection status of adjacent local attitude spheres at the connection position, connect the effective spherical attitude ranges that maintain the continuous boundary migration, and form an effective attitude channel.

[0010] Optionally, obtaining the inspection interference attitude components specifically includes: Based on the current movement position of the inspection robot, the current terrain process position is determined in the effective attitude channel, and the local attitude sphere to which the current terrain process position belongs is located. The attitude migration position corresponding to the terrain process position is fixed on the local attitude sphere, and the same process attitude section is established along the spherical direction that is different from the attitude Spherelet spherical chain migration direction. Extract the overlap range between the same process attitude section and the effective attitude channel, determine the overlap range as the same process effective attitude range corresponding to the current terrain process position, and limit the spherical search area of ​​Spherelets attitude projection with the same process effective attitude range, project the robot's current attitude to the spherical attitude position that satisfies the effective attitude channel constraint, and obtain the same process projected attitude. Compare the robot's current pose with the projected pose of the same process. Separate the pose differences according to the migration direction of the pose Spherelet sphere chain at the current terrain process position. Retain the pose changes that are consistent with the migration direction of the pose Spherelet sphere chain and extract the lateral pose offsets that deviate from the effective pose range of the same process. The offset direction and degree are determined based on the lateral attitude offset, and the attitude deviation state is determined in combination with the effective attitude channel boundary corresponding to the current terrain process position. The attitude changes that deviate from the effective attitude range of the same process and cause the inspection observation state offset are organized into inspection interference attitude components.

[0011] Optionally, obtaining the attitude correction execution result specifically includes: Based on the inspection interference attitude components, determine the correction direction of the current attitude deviation from the effective attitude channel, keep the terrain adaptation attitude corresponding to the current terrain process position unchanged, and search along the correction direction within the effective attitude channel for the spherical attitude position that can eliminate the inspection interference attitude components and maintain the effective observation of the inspection target, and determine the target correction attitude. Compare the robot's current posture with the target's corrected posture, generate a posture correction control quantity according to the posture deflection direction and degree, send the posture correction control quantity to the posture actuator of the inspection robot to complete the posture adjustment, reacquire the visual orientation of the inspection target, and obtain the corrected visual response. Based on the spherical position of the target's corrected posture in the posture Spherelet spherical chain and the position change in the effective posture channel, the expected visual migration of the inspected target is determined. The expected visual migration is compared with the corrected visual response in terms of direction and degree of migration, and the visual response deviation is extracted. Based on the visual response deviation, the local attitude spheres that generate observational inconsistencies are determined. When the visual response deviation is manifested as an inconsistency between the actual migration degree and the expected migration degree, the spherical curvature of the corresponding local attitude sphere is adjusted. When the visual response deviation is manifested as the actual migration direction deviating from the expected migration direction, the attitude migration direction of the corresponding local attitude sphere is adjusted. Based on the adjusted local attitude spheres, the effective spherical attitude range within the corresponding terrain process range is redefined, the attitude Spherelet spherical chain and the effective attitude channel are updated, and the attitude correction execution result is formed.

[0012] Optionally, obtaining the intelligent posture correction result of the inspection robot specifically includes: Extract the robot's posture state and the visual observation state of the inspected target from the posture correction execution result, associate the effective posture channel corresponding to the current terrain process position, and determine the spherical posture position after posture correction. Based on the spherical posture position and the visual observation status of the inspection target, perform posture correction status verification to determine that the robot posture satisfies the effective posture channel constraint and maintains a terrain-adaptive posture. The posture correction status verification results and posture correction execution results are organized to obtain the intelligent posture correction results of the inspection robot.

[0013] The beneficial effects of this invention are: This invention addresses the issue of continuous bending characteristics in the posture changes of inspection robots during slope travel, terrain undulations, turning, and obstacle crossing. It introduces Spherelets to represent robot posture changes locally on a spherical surface. Compared to describing robot posture using a fixed reference posture, local planar relationships, or a single posture angle threshold, Spherelets can characterize the nonlinear bending process of posture changes with terrain using local spherical curvature, making it more suitable for describing the posture migration relationships formed when an inspection robot moves continuously along an inspection path. Furthermore, this invention does not directly use conventional Spherelets to independently fit spherical surfaces to each local posture interval. Instead, it constrains the sphere center position and sphere radius based on the curvature changes of the local support surface, and uses the exit posture point and exit tangent of the previous local posture sphere as the entry constraint for adjacent local posture spheres. This allows each local posture sphere to be recursively connected according to the terrain progress and maintains the continuity of the posture migration direction, thereby forming a posture Spherelet spherical chain. This improvement enables the geometry of the local attitude sphere to simultaneously reflect robot attitude data, terrain curvature, and previous attitude transition relationships, reducing problems such as local attitude breakage, curvature direction distortion, and inaccurate terrain adaptation relationships caused by independent spherical fitting, and improving the continuity and realism of robot attitude modeling under complex terrain conditions.

[0014] This invention further establishes a correspondence between the visual orientation of the inspection target and the spherical posture position on the posture Spherelet spherical chain, and forms an effective posture channel based on the spherical posture range corresponding to the effective observation of the inspection target. This makes Spherelets no longer just used to describe the robot's posture itself, but further constrained by the inspection vision task. Addressing the problem that conventional Spherelet projection may freely search for neighboring posture positions along local spheres and easily mix the robot's normal terrain-adapted posture with inspection observation deviations, this invention establishes a same-process posture section. While keeping the terrain process position unchanged, it restricts the Spherelet posture projection to the effective posture range corresponding to the same terrain process position, thereby preserving the terrain-adapted posture generated along the posture Spherelet spherical chain, extracting only the lateral posture offset deviating from the effective posture channel, and forming the inspection interference posture component. This enables the differentiation between reasonable pitch and roll caused by slope changes and terrain undulations and posture deviations that truly affect the visual observation of the inspection target, reducing the probability of misjudging and miscorrecting the normal terrain-adapted posture, reducing overcorrection and repeated posture adjustments, and improving the targeting and accuracy of posture correction.

[0015] After determining the target correction posture based on the inspection interference posture components and effective posture channels, this invention does not solely rely on whether the robot posture has been restored to complete control judgment. Instead, it compares the actual visual response after posture correction with the expected visual migration, and corrects the spherical curvature, posture migration direction, and effective spherical posture range of the corresponding local posture sphere based on the visual response deviation. This allows the actual inspection results to participate in the updating of the posture Spherelet spherical chain and effective posture channels. By introducing this visual response feedback, the posture geometry model and the effective visual posture range can be continuously corrected according to changes in actual terrain, robot execution errors, and changes in the inspection target observation. This prevents the fixed posture model from gradually deviating from the real motion state during long-term operation, thereby improving the posture correction accuracy, environmental adaptability, visual observation stability, and posture control reliability of the inspection robot in complex inspection environments. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an AI vision-based intelligent posture correction method for inspection robots proposed in this invention. Figure 2 This is a schematic diagram illustrating the collaborative processing of the posture Spherelet chain, effective posture channel, and posture projection in the same process in an AI vision-based intelligent posture correction method for inspection robots proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-2 A method for intelligent posture correction of inspection robots based on AI vision includes the following steps: Collect continuous inspection images and robot motion data during the inspection process of the inspection robot, construct visual posture association information, and form a visual posture observation sequence; The terrain process is established based on the visual attitude observation sequence, and the attitude curvature interval is divided according to the change relationship of visual attitude correlation information along the terrain process to form an attitude curvature interval sequence. For the attitude curvature interval sequence, Spherelets are called to construct local attitude spheres. The position of the sphere center and the radius of the sphere are constrained according to the curvature change. The tangential continuity of adjacent local attitude spheres is maintained to form an attitude Spherelet sphere chain. Based on the attitude Spherelet spherical chain, a mapping between the inspection observation relationship and the spherical attitude is established, the effective spherical attitude range is defined and connected along the terrain process to form an effective attitude channel; Based on the effective attitude channel, establish the same-process attitude section, perform Spherelets attitude projection while keeping the terrain process position unchanged, retain the terrain-adaptive attitude, extract the lateral attitude offset, and form the inspection disturbance attitude component. The target correction posture is determined based on the inspection interference posture components and effective posture channels, posture correction is performed, and the posture Spherelet chain and effective posture channels are modified according to the corrected visual response to form the posture correction execution result. The intelligent posture correction result of the inspection robot is obtained based on the posture correction execution result.

[0019] In this embodiment, obtaining the visual pose observation sequence specifically includes: The inspection process involves collecting continuous inspection images and robot motion data during the inspection of the inspection robot. A temporal correspondence is established between the continuous inspection images and the robot motion data. The inspection target area and local support area are extracted from the continuous inspection images. The continuous inspection images are image data collected by the vision acquisition device as the inspection robot moves continuously and arranged according to the acquisition time. The robot motion data includes the inspection robot's position, posture, direction of travel, speed, and displacement at each acquisition time. The temporal correspondence is based on the acquisition time of the continuous inspection images, determining the motion state corresponding to the same acquisition time in the robot motion data, thus establishing a temporal association between each inspection image and the corresponding robot motion state. The local support area is an image area corresponding to the current travel position of the inspection robot and representing the terrain support state where the robot is located. Based on the inspection target area, local support area and robot motion data, the visual orientation change of the inspection target, the orientation change of the local support surface and the robot posture change are determined, and the correspondence between the three is established to obtain visual posture association information. The visual posture association information is the association information obtained after establishing the correspondence between the visual orientation change of the inspection target, the orientation change of the local support surface and the robot posture change at the same motion position. Visual attitude association information is arranged according to the movement sequence of the inspection robot along the inspection path, and the visual attitude association information corresponding to adjacent inspection movement positions is associated to form a visual attitude observation sequence. The association corresponding to adjacent inspection movement positions is used to record the continuous changes in the visual orientation of the inspection target, the direction of the local support surface, and the robot's posture from the previous inspection movement position to the next inspection movement position. The visual attitude observation sequence is a sequence formed by continuously organizing the visual attitude association information corresponding to each inspection movement position according to the movement sequence.

[0020] The image resolution of the visual acquisition device is set to 1920×1080, the continuous inspection image acquisition frequency is set to 30 frames / s, the robot motion data sampling frequency is set to 100Hz, and the allowable acquisition time deviation when performing time-series correspondence between continuous inspection images and robot motion data is no more than 20ms. When extracting the inspection target area, the target recognition confidence threshold is set to 0.80. Visual recognition results below 0.80 are not included in the construction of visual posture association information. The local support area is selected from the visible ground area corresponding to the direction of travel of the inspection robot, and the longitudinal range of the area is set to the lower 40% of the height of the continuous inspection image. The robot posture data records pitch, roll and heading changes at an angular resolution of 0.1°, so that visual orientation changes, local support surface orientation changes and robot posture changes can be established in correspondence according to a unified acquisition position.

[0021] In this embodiment, obtaining the posture bending interval sequence specifically includes: Based on the order of the inspection robot's movement positions in the visual attitude observation sequence, the inspection start position is selected as the starting point of the terrain process. The movement displacement between adjacent movement positions is extracted along the inspection path. The movement displacement is continuously accumulated according to the direction of the inspection path to determine the terrain process position corresponding to each movement position and establish the terrain process. The terrain process is the position progress benchmark when the inspection robot moves along the inspection path, with the inspection start position as the benchmark and continuously increasing according to the actual movement displacement. It is used to convert the visual attitude association information in the visual attitude observation sequence into attitude observation information continuously distributed along the inspection path. The terrain process position continuously advances as the inspection robot moves forward along the inspection path and retains the visual attitude association information of the corresponding movement position. The visual pose association information of adjacent positions is compared along the terrain process. The interval connection position is determined according to the changing trend of the local support surface direction, robot pose and visual orientation of the inspection target. The interval connection position is the position where the visual pose association information changes along the terrain process. When determining the position, the changing direction of the local support surface direction and the robot pose direction are compared. The position where the original continuous change state changes is marked in combination with the migration trend of the visual orientation of the inspection target. The terrain process is segmented based on the connection points of the intervals. The posture curvature intervals are determined by combining the continuous change characteristics of visual posture association information within each segment. The continuity of the boundaries of adjacent posture curvature intervals is checked, and the posture curvature intervals are organized according to the order of the terrain process to form a posture curvature interval sequence. The posture curvature interval is a terrain process segment with continuous posture change characteristics between the connection points of adjacent intervals. The boundary continuity check is used to compare the change direction and degree of visual posture association information on both sides of the interval connection point, so that the interval connection point can represent the transformation between adjacent posture change trends. The posture curvature interval sequence retains the terrain process range, interval connection point and visual posture association information corresponding to the posture curvature interval, providing interval data for the continuous construction of the local posture sphere.

[0022] The terrain progress is accumulated according to the actual movement displacement of the inspection robot. The resampling interval between adjacent terrain progress positions is set to 0.05m. When a single movement displacement is less than 0.01m, a new terrain progress position is not established separately. When the local support surface direction change between adjacent positions reaches 3°, the robot posture change reaches 2°, and the visual orientation change of the inspection target reaches 4°, the corresponding change state is recorded. When three consecutive terrain progress positions maintain the same change direction and then a direction change occurs, the change position is used as a candidate for interval connection position. The minimum terrain progress length of the posture curvature interval is set to 0.30m. Segments with a length less than 0.30m and consistent posture change directions on both sides are merged into adjacent posture curvature intervals to reduce interval fragmentation caused by short-distance visual disturbances.

[0023] In this embodiment, obtaining the attitude Spherelet chain specifically includes: Following the terrain progression sequence of the posture curvature intervals, the visual posture association information corresponding to the first posture curvature interval is extracted. Spherelets are then called to determine the local posture subspace and perform spherical initialization to obtain the initial local posture sphere. The Spherelets are a method that approximates the nonlinear data manifold using a local spherical structure. By establishing a low-dimensional subspace and constructing a local sphere within the local data region, the spherical curvature structure represents the nonlinear changes of data within the local range. The local posture subspace is centered on the visual posture association information within the first posture curvature interval to extract the main directions that can centrally represent the continuous changes in robot posture. The corresponding posture data is then mapped to the local space formed by the main directions. The spherical initialization determines the initial sphere center within the local posture subspace based on the distribution center of the posture data and the initial sphere radius based on the overall distance distribution of the posture data around the initial sphere center, so that the posture data can be approximated by the local sphere. The initial local posture sphere serves only as the initial solution reference for the posture Spherelet spherical chain and retains spherical geometric information for the recursive construction of adjacent local posture spheres at the corresponding interval connection positions. Extract the exit posture point and exit tangent corresponding to the interval connection position of the previous local posture sphere. Use the exit posture point and exit tangent as the entry constraint for the adjacent posture bending interval. Simultaneously, establish curvature constraints based on the degree and direction of rotation of the local support surface along the terrain process within the adjacent posture bending interval. Use the entry constraint and curvature constraint to limit the solution range of the center and radius of the local posture sphere, forming the local spherical constraint conditions. The exit posture point is the spherical posture position corresponding to the previous local posture sphere moving along the terrain process to the interval connection position. The exit tangent is the spherical tangent determined by the robot posture migration direction at the exit posture point of the previous local posture sphere. The entry constraint fixes the exit posture point as the entry position of the adjacent local posture sphere and sets the exit tangent... As the allowed attitude migration direction at the entry position of adjacent local attitude spheres, the solution of adjacent local attitude spheres starts from the exit geometry of the previous local attitude sphere, rather than re-fitting mutually independent spheres for adjacent attitude curvature intervals. The curvature constraint limits the sphere curvature strength based on the degree of rotation generated by the local support surface direction within a unit terrain process. When the degree of rotation increases, the allowable sphere radius range is compressed, and when the degree of rotation decreases, the allowable sphere radius range is expanded. The sphere center is limited to the curvature side of the attitude migration trajectory based on the rotation direction of the local support surface. The sphere center position with the curvature direction opposite to the actual terrain change direction is excluded from the sphere solution stage. The local sphere constraint conditions are composed of the entry position, entry direction, allowable sphere radius range, and allowable sphere center direction. Based on local spherical constraints, conditional Spherelets spherical solving is performed on adjacent attitude curvature intervals. Local spherical candidates are generated within the allowable spherical radius range and the allowable direction of the sphere center. The fit between the local spherical candidates and the attitude data is compared using visual attitude association information of the corresponding attitude curvature interval. Local spherical candidates that violate entry constraints, curvature constraints, and tangential continuity requirements are filtered out to determine the current local attitude sphere. This conditional Spherelets spherical solving confines the process of independently determining the sphere center and sphere radius for local data in conventional Spherelets within the local spherical constraints. Local spherical candidates first satisfy the entry position and entry direction passed from the previous local attitude sphere, and then satisfy the current... The range of spherical radius and the allowed direction of the sphere center corresponding to the attitude curvature interval are then determined based on the degree of fit of the visual attitude association information on the candidate spheres. The local spherical candidates that violate the entry constraint include spherical candidates that cannot pass through the entry position and whose attitude migration direction deviates from the exit tangent at the entry position. The local spherical candidates that violate the curvature constraint include spherical candidates whose spherical curvature exceeds the current terrain curvature range and whose sphere center is located on the opposite side of the terrain curvature. The tangential continuity requirement is that the attitude migration direction of the current local attitude sphere at the entry position is continuous with the exit tangent of the previous local attitude sphere, so that the determination of the local attitude sphere is simultaneously constrained by the geometric state of the previous sphere, the current terrain curvature, and the current visual attitude data. Extract the exit attitude point and exit tangent corresponding to the connection position of the current local attitude sphere in the next attitude bending interval. Transfer the exit attitude point and exit tangent to the adjacent attitude bending interval. Combine the local spherical constraints and conditional Spherelets spherical solutions with the terrain curvature corresponding to the adjacent attitude bending intervals. Recursively determine the local attitude spheres along the attitude bending interval sequence. Connect the local attitude spheres according to the terrain progress sequence to form an attitude Spherelet spherical chain. The recursive determination process is as follows: use the solution result of the previous local attitude sphere as the pre-order geometric condition of the adjacent attitude bending interval, and repeatedly execute the establishment of local spherical constraints and the solution of conditional Spherelets spherical solutions until the determination of each local attitude sphere in the attitude bending interval sequence is completed. The attitude Spherelet spherical chain is a local attitude sphere structure that is recursively generated according to the terrain progress sequence and maintains the continuity of the attitude migration direction at the interval connection position. It is used to continuously characterize the attitude bending and attitude migration of the inspection robot caused by terrain changes.

[0024] In this embodiment, obtaining the effective attitude channel specifically includes: Based on the attitude Spherelet spherical chain, the attitude position on the local attitude sphere is sequentially associated with the visual orientation of the corresponding inspection target, and the association results are organized along the attitude migration direction of the local attitude sphere to establish a spherical observation mapping. The inspection observation relationship is the correspondence between the robot's attitude change and the visual orientation change of the inspection target in the visual attitude association information. The spherical observation mapping uses the attitude position on the local attitude sphere as the spherical position reference and records the visual orientation of the inspection target when the inspection robot is in the corresponding attitude position, so that the continuous attitude migration of the local attitude sphere and the continuous orientation migration of the inspection target in the visual acquisition field of view are established to correspond. Based on the spherical observation mapping, the visual orientation of the inspection target is tracked along the attitude migration direction of each local attitude sphere. The spherical boundary positions corresponding to the inspection target entering and leaving the effective observation area are determined. The effective spherical attitude range on each local attitude sphere is defined by the spherical boundary positions. The effective observation area is the visual acquisition area where the inspection target can maintain visual recognition and spatial position resolution. When the visual orientation of the inspection target is inside the effective observation area, the corresponding spherical attitude position is kept within the effective spherical attitude range. When the visual orientation of the inspection target reaches the boundary of the effective observation area, the corresponding spherical attitude position is determined as the boundary of the effective spherical attitude range. The effective spherical attitude range is a continuous spherical area on the same local attitude sphere that can maintain effective observation of the inspection target. This allows the local spherical structure in the attitude Spherelet spherical chain to simultaneously have terrain attitude curvature representation and inspection visual observation constraints. The effective spherical attitude ranges on adjacent local attitude spheres are connected sequentially according to the terrain progress. The connection attitude corresponding to the interval connection position is used as the range transfer benchmark. The continuous migration of the boundaries on both sides of the effective spherical attitude range is tracked along the attitude Spherelet sphere chain. The effective range connection status of adjacent local attitude spheres at the connection position is checked. Effective spherical attitude ranges that maintain continuous boundary migration are connected to form an effective attitude channel. The range transfer benchmark is the attitude position that the adjacent local attitude spheres share at the interval connection position. The effective range boundary of the previous local attitude sphere is transferred to the next local attitude sphere through the connection attitude. The effective range connection status is determined according to the attitude migration direction of the effective range boundary on both sides of the connection position and the visual orientation migration direction of the inspection target. The corresponding range connection is maintained when the boundary migration remains continuous and the visual orientation of the inspection target does not cross the effective observation area. The effective attitude channel is an effective spherical attitude region that is continuously distributed on the attitude Spherelet sphere chain along the terrain progress and retains the effective range boundary corresponding to each terrain progress position, providing spherical constraints for determining the allowable attitude range under a fixed terrain progress position.

[0025] This invention introduces visual orientation constraints of the inspection target into the Spherelet spherical chain of the posture. It limits the spherical posture range on the continuous posture transfer structure to maintain effective visual recognition and spatial position resolution. This expands the robot posture representation from simply reflecting the curvature of the terrain to an effective posture channel that simultaneously meets the inspection observation requirements. It avoids the target deviating from the effective observation area due to correction based solely on the terrain posture, thereby improving the accuracy of inspection posture constraints, the stability of visual observation, and the reliability of posture correction under complex terrain conditions.

[0026] In this embodiment, obtaining the inspection interference attitude components specifically includes: Based on the current movement position of the inspection robot, the current terrain process position is determined in the effective attitude channel, and the local attitude sphere to which the current terrain process position belongs is located. The attitude migration position corresponding to the terrain process position is fixed on the local attitude sphere. A same-process attitude section is established along the spherical direction that is different from the migration direction of the attitude Spherelet spherical chain. The same-process attitude section is a local section that keeps the terrain process position fixed on the local attitude sphere and allows the robot's attitude to change near the corresponding spherical position. When the section is established, the migration direction of the attitude Spherelet spherical chain at the current terrain process position is used as the terrain adaptation direction. The position advancement generated along the terrain adaptation direction is excluded from the attitude changes within the section, so that the attitude changes retained within the section reflect the attitude deflection that the robot can generate when it is at the same terrain process position. The overlapping range of the same-process attitude section and the effective attitude channel is extracted, and the overlapping range is determined as the same-process effective attitude range corresponding to the current terrain process position. The spherical search area of ​​Spherelets attitude projection is limited by the same-process effective attitude range. The robot's current attitude is projected to the spherical attitude position that satisfies the effective attitude channel constraint to obtain the same-process projected attitude. The same-process effective attitude range is the spherical attitude area in the same-process attitude section that can maintain effective observation of the inspection target. The Spherelets attitude projection is different from directly finding the spherical position that is close to the robot's current attitude within the local attitude spherical range. During the projection process, the current terrain process position is locked and the projection position is searched only within the same-process effective attitude range. This ensures that the projection result does not return to the previous terrain process position along the attitude Spherelet spherical chain, nor does it migrate to the next terrain process position. This avoids including the normal attitude migration of the inspection robot caused by slope changes and terrain undulations in the attitude to be corrected. Compare the robot's current posture with the projected posture of the same process. Separate the posture differences according to the migration direction of the posture Spherelet sphere chain at the current terrain process position. Retain the posture changes consistent with the migration direction of the posture Spherelet sphere chain and extract the lateral posture offsets that deviate from the effective posture range of the same process. The posture changes consistent with the migration direction of the posture Spherelet sphere chain correspond to the terrain adaptation posture of the inspection robot as the local support surface changes. The terrain adaptation posture changes along the continuous migration direction of the current local posture sphere and maintains the posture trend corresponding to the current terrain change. The lateral posture offset is the part of the posture change in which the robot's current posture deviates from the effective posture channel relative to the projected posture of the same process. The determination of the lateral posture offset keeps the current terrain process position and terrain adaptation posture migration unchanged, and only extracts the posture differences that deviate from the effective observation constraints. The offset direction and degree are determined based on the lateral attitude offset, and the attitude deviation state is determined in combination with the effective attitude channel boundary corresponding to the current terrain process position. The attitude changes that deviate from the effective attitude range of the same process and cause the inspection observation state to shift are organized into inspection interference attitude components. The inspection interference attitude components are attitude changes that deviate from the effective attitude channel obtained from the robot's current attitude after retaining the terrain adaptation attitude. They are used to characterize the direction and degree of attitude correction that needs to be implemented, and together with the effective attitude channel corresponding to the same terrain process position, they define the target correction attitude.

[0027] This invention limits the Spherelets attitude projection range at a fixed terrain progress location, separating the normal attitude migration caused by slope changes and terrain undulations from the lateral attitude offset that affects inspection observation. This ensures that attitude correction is only applied to inspection interference attitudes that are truly deviating from the effective attitude channel, avoiding the misjudgment of terrain-adaptive attitudes as abnormal deviations, reducing the probability of overcorrection and repeated adjustments, thereby improving the accuracy of inspection robot attitude recognition, the targeting of attitude correction, and the inspection stability in complex terrain environments.

[0028] In this embodiment, obtaining the posture correction execution result specifically includes: Based on the inspection interference posture components, the correction direction of the current posture deviating from the effective posture channel is determined. The terrain adaptation posture corresponding to the current terrain process position remains unchanged. Within the effective posture channel, the spherical posture position that can eliminate the inspection interference posture components and maintain the effective observation of the inspection target is searched along the correction direction. The target correction posture is determined. The target correction posture is located within the effective posture range corresponding to the current terrain process position. The posture adjustment direction is to eliminate the lateral posture offset. The terrain adaptation posture migration represented by the posture Spherelet spherical chain is not changed. The inspection robot completes the posture return to the inside of the effective posture channel while maintaining the current terrain adaptation state. The robot's current posture is compared with the target's corrected posture. An attitude correction control quantity is generated according to the attitude deflection direction and degree. The attitude correction control quantity is sent to the attitude execution mechanism of the inspection robot to complete the attitude adjustment. The visual orientation of the inspection target is reacquired to obtain the corrected visual response. The corrected visual response is the actual migration state of the inspection target relative to the visual orientation before correction after the attitude adjustment is completed. The actual migration direction and actual migration degree are retained to verify the inspection observation changes corresponding to the target's corrected posture. The expected visual migration of the inspection target is determined based on the spherical position of the target correction posture in the posture Spherelet spherical chain and the position change in the effective posture channel. The expected visual migration is compared with the corrected visual response in terms of direction and migration degree, and the visual response deviation is extracted. The expected visual migration is the expected change of the inspection target obtained by the robot from the current posture to the target correction posture according to the posture migration relationship represented by the posture Spherelet spherical chain and the inspection observation state defined by the effective posture channel. The visual response deviation is used to characterize the degree of inconsistency between the visual migration of the inspection target caused by the actual posture correction and the visual migration corresponding to the current spherical posture structure. Based on visual response deviation, the local attitude spheres that cause observational inconsistencies are identified. When the visual response deviation manifests as an inconsistency between the actual and expected migration degrees, the spherical curvature of the corresponding local attitude sphere is adjusted. When the visual response deviation manifests as an actual migration direction deviating from the expected migration direction, the attitude migration direction of the corresponding local attitude sphere is adjusted. Based on the adjusted local attitude spheres, the effective spherical attitude range within the corresponding terrain process is redefined, and the attitude Spherelet sphere chain and effective attitude channels are updated to form the attitude correction execution result. The correction of the attitude Spherelet sphere chain by the visual response deviation is limited to the terrain process range that causes observational inconsistencies, so that the actual visual response participates in the update of the local attitude sphere geometry. The attitude correction execution result includes the robot attitude state after attitude correction and the attitude Spherelet sphere chain and effective attitude channel state after actual visual response correction.

[0029] This invention incorporates the actual visual response after posture correction into the correction process of the posture Spherelet chain and the effective posture channel, enabling the posture geometry to be continuously corrected with the actual execution results. This reduces the correction error caused by the deviation between the fixed posture model and the actual inspection state, enhances the matching degree between the target corrected posture and the inspection observation requirements, and thus improves the accuracy of posture correction, environmental adaptability and visual observation stability of the inspection robot.

[0030] In this embodiment, obtaining the intelligent posture correction result of the inspection robot specifically includes: Extract the robot's posture state and the visual observation state of the inspection target from the posture correction execution result, associate them with the effective posture channel corresponding to the current terrain process position, and determine the spherical posture position after posture correction. The spherical posture position is the mapping position of the robot's posture after posture correction in the effective spherical posture range corresponding to the current terrain process position, which is used to characterize the position state of the robot relative to the effective posture channel after posture correction. Based on the spherical posture position and the visual observation status of the inspection target, the posture correction status verification is performed to determine that the robot posture satisfies the effective posture channel constraint and maintains the terrain-adaptive posture. The posture correction status verification includes determining whether the spherical posture position is within the effective spherical posture range corresponding to the current terrain process position, and confirming that the inspection target maintains effective observation, while retaining the terrain-adaptive posture migration relationship represented by the posture Spherelet spherical chain. The posture correction state verification results and posture correction execution results are organized to obtain the inspection robot posture intelligent correction results. The inspection robot posture intelligent correction results are used to record the robot posture state after posture correction, the visual observation state of the inspection target, and the corresponding posture correction state, which are used to characterize the final execution state of the inspection robot after completing posture correction at the current terrain process position.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a visual inspection scenario involving a patrol robot, which included flat road sections, continuous slopes, laterally inclined road sections, partially undulating road sections, and turning road sections. The patrol robot moved continuously along a preset inspection path and performed visual inspections of instruments, equipment status indicator areas, and structural components located at different heights on both sides of the inspection path. This example focuses on verifying whether the robot can avoid mistaking normal terrain-adapting postures for abnormal postures when traversing slopes and partially undulating areas, and whether it can accurately extract the posture components that need correction after visual deviation of the inspected target occurs. The following data serves as exemplary verification data in this example to illustrate the implementation process and technical effects of this invention.

[0032] The total length of the verification path was set to 120m, and the average robot speed was set to 0.45m / s. The path included approximately 28m of continuous slope, 18m of lateral incline, and several areas of continuous terrain undulation. The slope angle ranged from 3° to 12°, and the lateral terrain angle ranged from 2° to 7°. A total of 36 visual inspection targets were deployed during the inspection. The lateral distance between the targets and the robot's path ranged from 1.5m to 4.0m, and the center height of the targets ranged from 0.8m to 2.2m. The visual acquisition device captured continuous inspection images at a resolution of 1920×1080 and a frame rate of 30 frames per second. Robot motion data was acquired at a frequency of 100Hz, and the time correspondence error between the images and motion data was controlled within 20ms. The target visual recognition confidence threshold was set to 0.80.

[0033] After the inspection robot enters a continuous slope, its pitch attitude changes continuously with the direction of the local support surface. Within the section where the slope gradually increases from 3° to 10°, the robot's actual pitch angle gradually increases from approximately 1.2° to 8.7°. When using a conventional attitude correction method that uses a fixed horizontal attitude as a reference, this attitude change is easily treated as an attitude deviation. In the test, 17 unnecessary pitch corrections occurred, 9 of which caused a significant reverse movement of the vertical position of the inspection target in the image. With this invention, the visual attitude observation sequence correlates the direction of the local support surface, the robot's attitude change, and the visual orientation change of the inspection target, and divides the attitude curvature interval according to the terrain progress, so that the continuous attitude changes caused by the slope are preserved as terrain-adaptive attitudes.

[0034] During the construction of the attitude Spherelet chain, at least 12 valid attitude sampling points are retained for each attitude curvature interval, and the local attitude subspace is set to 3D. The exit attitude point and exit tangent of adjacent local attitude spheres are used to restrict the entry position and entry direction of the next local attitude sphere, and the change in the orientation of the local support surface is further used to restrict the sphere radius and the orientation of the sphere center. During the verification process, the average tangential angle between the connection positions of adjacent local attitude spheres is 2.1°, and the maximum is 4.6°, all within 5°; if independent ordinary local Spherelets are used for fitting, the average tangential angle between the connection positions of adjacent spheres under the same data conditions is about 7.8°, and the maximum reaches 14.2°. It can be seen that the present invention, by using the previous exit geometry and terrain curvature to jointly constrain the subsequent local attitude sphere, can reduce the abrupt attitude transitions between local spheres.

[0035] During the construction of the effective attitude channel, the center of the inspection target is kept within 15%–85% of the horizontal and 15%–85% of the vertical area of ​​the image as an effective visual observation condition, while requiring a target recognition confidence level of not less than 0.80. When the robot passes through a horizontally inclined road section, although some postures meet the terrain adaptation requirements, the inspection target has gradually moved towards the edge of the image. Therefore, relying solely on the terrain posture cannot guarantee the quality of visual inspection. This invention uses spherical observation mapping to map the posture position on the local posture sphere to the visual orientation of the inspection target, and further limits the effective spherical posture range on the posture Spherelet spherical chain, so that postures that meet the terrain adaptation requirements but have caused the inspection target to approach the edge of the field of view are no longer included in the effective attitude channel.

[0036] In the Spherelets attitude projection verification within the same process, the allowable matching deviation for the current terrain process position is set to ±0.02m, and the attitude search step size within the effective attitude range of the same process is set to 0.5°. When the robot traverses a slope of approximately 8°, the current robot pitch angle reaches 7.4°, while simultaneously experiencing a 3.6° roll deviation due to local lateral disturbances. If the current attitude is directly compared with the horizontal reference attitude, the total attitude deviation is large, making it easy to correct both pitch and roll simultaneously. This invention performs Spherelets attitude projection after fixing the current terrain process position. The pitch change of approximately 6.9° is retained as the terrain-adaptive attitude, and the actual lateral attitude offset that needs correction is approximately 3.3°. Therefore, attitude correction is mainly performed on the roll portion that affects visual observation. After correction, the center position of the inspected target recovers from approximately 82% of the position near the lateral boundary of the image to approximately 63%, while the robot maintains a pitch state that matches the slope, without any erroneous control that pulls back to a horizontal attitude.

[0037] During attitude correction, the maximum single attitude correction angle was set to 8°, and the attitude correction angular velocity was limited to 10° / s. After correction, five consecutive inspection images were acquired to determine the actual visual response, and the actual visual migration was compared with the expected visual migration given by the spherical observation mapping. When the migration degree deviation exceeded 15% or the migration direction angle exceeded 5°, local correction was performed on the corresponding local attitude sphere. After three rounds of visual response feedback, the average degree deviation between the expected and actual visual migration in the test section decreased from the initial 18.6% to 7.4%, and the average deviation of the visual migration direction decreased from 6.2° to 2.5°, indicating that the actual inspection results can correct the attitude Spherelet spherical chain and effective attitude channel in reverse.

[0038] To further compare the technical effects, a fixed attitude threshold correction method, a standard Spherelets method without prior constraints and in-process projection, and the method of this invention were compared. The fixed attitude threshold correction method uses a preset reference attitude and attitude deviation threshold as the judgment criteria. When the robot's pitch angle, roll angle, or yaw angle deviates from the reference attitude by more than a set range, attitude adjustment is performed, without distinguishing whether the attitude change is caused by terrain adaptation or abnormal offset. The standard Spherelets method uses Spherelets to perform spherical fitting on each local attitude data and projects the robot's current attitude onto the corresponding local sphere, but does not introduce the exit state constraint of the previous local attitude sphere, nor does it perform in-process attitude projection while fixing the current terrain process position. Ten complete inspections were performed on the same inspection path, the same inspection target, and similar initial robot states. The statistical average results are shown in Table 1. Table 1. Performance Comparison of Different Attitude Correction Methods in Complex Terrain Inspection Scenarios

[0039] As shown in Table 1, the method of the present invention has good attitude correction accuracy and visual observation stability under complex terrain conditions. The average number of attitude corrections is reduced from 31.6 times in the fixed attitude threshold correction method to 15.8 times, the number of terrain-adaptive attitude miscorrections is reduced from 12.4 times to 1.6 times, and the number of repeated attitude adjustments is reduced to 1.9 times. This indicates that the present invention can better distinguish between normal terrain-adaptive attitudes and inspection interference attitudes that need correction. The effective observation rate of the inspection target reaches 98.1%, and the average lateral offset after attitude correction is reduced to 0.8°. At the same time, compared with the ordinary Spherelets method, the present invention reduces the average tangential angle between adjacent local attitude spheres from 7.8° to 2.1%, the average deviation of visual migration degree from 16.9% to 7.4%, and the average deviation of visual migration direction from 5.8° to 2.5%. This shows that the preceding geometric constraints, terrain curvature constraints, and visual response feedback can improve the continuity of the attitude Spherelet sphere chain and the matching degree between attitude changes and actual inspection visual responses.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent posture correction of an inspection robot based on AI vision, characterized in that, Includes the following steps: Collect continuous inspection images and robot motion data during the inspection process of the inspection robot, construct visual posture association information, and form a visual posture observation sequence; The terrain process is established based on the visual attitude observation sequence, and the attitude curvature interval is divided according to the relationship of visual attitude correlation information along the terrain process to form an attitude curvature interval sequence. For the attitude curvature interval sequence, Spherelets are called to construct local attitude spheres. The position of the sphere center and the radius of the sphere are constrained according to the curvature change. The tangential continuity of adjacent local attitude spheres is maintained to form an attitude Spherelet sphere chain. Based on the attitude Spherelet spherical chain, a mapping between the inspection observation relationship and the spherical attitude is established, the effective spherical attitude range is defined and connected along the terrain process to form an effective attitude channel; Based on the effective attitude channel, establish the same-process attitude section, perform Spherelets attitude projection while keeping the terrain process position unchanged, retain the terrain-adaptive attitude, extract the lateral attitude offset, and form the inspection disturbance attitude component. The target correction posture is determined based on the inspection interference posture components and effective posture channels, posture correction is performed, and the posture Spherelet chain and effective posture channels are modified according to the corrected visual response to form the posture correction execution result. The intelligent posture correction result of the inspection robot is obtained based on the posture correction execution result.

2. The method for intelligent posture correction of an inspection robot based on AI vision according to claim 1, characterized in that, The visual pose observation sequence is obtained specifically through: Collect continuous inspection images and robot motion data during the inspection process of the inspection robot, perform time-series correspondence between the continuous inspection images and robot motion data, and extract the inspection target area and local support area from the continuous inspection images. Based on the inspection target area, local support area and robot motion data, the visual orientation change of the inspection target, the orientation change of the local support surface and the robot posture change are determined, the correspondence between the three is established and the visual posture association information is obtained. Visual attitude association information is arranged according to the order of movement of the inspection robot along the inspection path, and visual attitude association information corresponding to adjacent inspection movement positions is associated to form a visual attitude observation sequence.

3. The method for intelligent posture correction of an inspection robot based on AI vision according to claim 1, characterized in that, The specific steps to obtain the attitude bending interval sequence include: Based on the order of the movement positions of the inspection robot in the visual posture observation sequence, the inspection start position is selected as the starting point of the terrain process. The movement displacement between adjacent movement positions is extracted along the inspection path. The movement displacement is continuously accumulated according to the direction of the inspection path to determine the terrain process position corresponding to each movement position and establish the terrain process. The visual pose association information of adjacent positions is compared along the terrain process, and the interval connection position is determined based on the changing trends of local support surface direction, robot pose and visual orientation of the inspection target. The terrain process is segmented based on the connection position of the intervals. The posture curvature intervals are determined by combining the continuous change characteristics of visual posture association information within each segment. The continuity of the boundary between adjacent posture curvature intervals is checked, and the posture curvature intervals are organized according to the order of the terrain process to form a posture curvature interval sequence.

4. The method for intelligent posture correction of an inspection robot based on AI vision according to claim 1, characterized in that, The specific steps involved in obtaining the attitude Spherelet chain are as follows: According to the terrain process order of the attitude curvature interval sequence, extract the visual attitude association information corresponding to the first attitude curvature interval, call Spherelets to determine the local attitude subspace and perform spherical initialization to obtain the initial local attitude sphere; Extract the exit attitude point and exit tangent corresponding to the interval connection position of the previous local attitude sphere, and use the exit attitude point and exit tangent as the entry constraint of the adjacent attitude bending interval. At the same time, establish curvature constraints based on the degree and direction of rotation of the local support surface along the terrain process in the adjacent attitude bending interval. Use the entry constraint and curvature constraint to limit the solution range of the center and radius of the local attitude sphere, and form the local sphere constraint conditions. Based on the local spherical constraints, conditional Spherelets spherical solutions are performed on adjacent attitude curvature intervals. Local spherical candidates are generated within the allowable spherical radius range and the allowable direction of the spherical center. The fitting status of the local spherical candidates and the attitude data is compared with the visual attitude association information of the corresponding attitude curvature interval. Local spherical candidates that violate the entry constraints, curvature constraints and tangential continuity requirements are filtered out to determine the current local attitude spherical. Extract the exit attitude point and exit tangent corresponding to the connection position of the current local attitude sphere in the next attitude curvature interval, transfer the exit attitude point and exit tangent to the adjacent attitude curvature interval, and repeat the local sphere constraint and condition Spherelet sphere solution in combination with the terrain curvature corresponding to the adjacent attitude curvature interval. Determine the local attitude sphere recursively along the attitude curvature interval sequence, and connect the local attitude spheres in the order of terrain progress to form the attitude Spherelet sphere chain.

5. The method for intelligent posture correction of an inspection robot based on AI vision according to claim 1, characterized in that, Obtaining the effective attitude channel specifically includes: Based on the attitude Spherelet spherical chain, the attitude position on the local attitude sphere is sequentially associated with the visual orientation of the corresponding inspection target, and the association results are organized along the attitude migration direction of the local attitude sphere to establish a spherical observation mapping. Based on the spherical observation mapping, the visual orientation of the inspection target is tracked along the attitude migration direction of each local attitude sphere, and the spherical boundary position corresponding to the inspection target entering and leaving the effective observation area is determined. The effective spherical attitude range on each local attitude sphere is defined by the spherical boundary position. Connect the effective spherical attitude ranges on adjacent local attitude spheres in the order of terrain progress, take the connection attitude corresponding to the interval connection position as the range transfer reference, track the continuous migration of the two sides of the effective spherical attitude range along the attitude Spherelet spherical chain, check the effective range connection status of adjacent local attitude spheres at the connection position, connect the effective spherical attitude ranges that maintain the continuous boundary migration, and form an effective attitude channel.

6. The method for intelligent posture correction of an inspection robot based on AI vision according to claim 1, characterized in that, The acquisition of the inspection interference attitude components specifically includes: Based on the current movement position of the inspection robot, the current terrain process position is determined in the effective attitude channel, and the local attitude sphere to which the current terrain process position belongs is located. The attitude migration position corresponding to the terrain process position is fixed on the local attitude sphere, and the same process attitude section is established along the spherical direction that is different from the attitude Spherelet spherical chain migration direction. Extract the overlap range between the same process attitude section and the effective attitude channel, determine the overlap range as the same process effective attitude range corresponding to the current terrain process position, and limit the spherical search area of ​​Spherelets attitude projection with the same process effective attitude range, project the robot's current attitude to the spherical attitude position that satisfies the effective attitude channel constraint, and obtain the same process projected attitude. Compare the robot's current pose with the projected pose of the same process. Separate the pose differences according to the migration direction of the pose Spherelet sphere chain at the current terrain process position. Retain the pose changes that are consistent with the migration direction of the pose Spherelet sphere chain and extract the lateral pose offsets that deviate from the effective pose range of the same process. The offset direction and degree are determined based on the lateral attitude offset, and the attitude deviation state is determined in combination with the effective attitude channel boundary corresponding to the current terrain process position. The attitude changes that deviate from the effective attitude range of the same process and cause the inspection observation state offset are organized into inspection interference attitude components.

7. The method for intelligent posture correction of an inspection robot based on AI vision according to claim 1, characterized in that, The specific steps to obtain the posture correction execution result include: Based on the inspection interference attitude components, determine the correction direction of the current attitude deviation from the effective attitude channel, keep the terrain adaptation attitude corresponding to the current terrain process position unchanged, and search along the correction direction within the effective attitude channel for the spherical attitude position that can eliminate the inspection interference attitude components and maintain the effective observation of the inspection target, and determine the target correction attitude. Compare the robot's current posture with the target's corrected posture, generate a posture correction control quantity according to the posture deflection direction and degree, send the posture correction control quantity to the posture actuator of the inspection robot to complete the posture adjustment, reacquire the visual orientation of the inspection target, and obtain the corrected visual response. Based on the spherical position of the target's corrected posture in the posture Spherelet spherical chain and the position change in the effective posture channel, the expected visual migration of the inspected target is determined. The expected visual migration is compared with the corrected visual response in terms of direction and degree of migration, and the visual response deviation is extracted. Based on the visual response deviation, the local attitude spheres that generate observational inconsistencies are determined. When the visual response deviation is manifested as an inconsistency between the actual migration degree and the expected migration degree, the spherical curvature of the corresponding local attitude sphere is adjusted. When the visual response deviation is manifested as the actual migration direction deviating from the expected migration direction, the attitude migration direction of the corresponding local attitude sphere is adjusted. Based on the adjusted local attitude spheres, the effective spherical attitude range within the corresponding terrain process range is redefined, the attitude Spherelet spherical chain and the effective attitude channel are updated, and the attitude correction execution result is formed.

8. The method for intelligent posture correction of an inspection robot based on AI vision according to claim 1, characterized in that, The specific methods for obtaining the intelligent posture correction results of the inspection robot include: Extract the robot's posture state and the visual observation state of the inspected target from the posture correction execution result, associate the effective posture channel corresponding to the current terrain process position, and determine the spherical posture position after posture correction. Based on the spherical posture position and the visual observation status of the inspection target, perform posture correction status verification to determine that the robot posture satisfies the effective posture channel constraint and maintains a terrain-adaptive posture. The posture correction status verification results and posture correction execution results are organized to obtain the intelligent posture correction results of the inspection robot.