A navigation matching correction method based on a patrol robot

Abstract

The application discloses a navigation matching correction method based on a patrol robot, and establishes a feature database by deploying physical calibration objects and identifying natural feature objects to provide reliable positioning reference for the robot. In the positioning process, visual and laser radar data are fused to realize coarse positioning, and a dynamic trust degree evaluation mechanism is introduced to quantitatively position the reliability in real time through an exponential decay model. When the trust degree is lower than a threshold value, the system automatically triggers a compensation behavior to search for features again. In the aspect of multi-robot cooperation, secondary positioning correction is realized through trajectory matching and data fusion, high-confidence reference data is screened by using a clustering algorithm, and the group positioning accuracy is improved. For key inspection areas, multi-angle image matching is adopted to realize fine positioning, and positioning errors are dynamically corrected through a sliding window. Through closed-loop correction and adaptive optimization, the application significantly improves the continuity and accuracy of robot navigation in a complex environment, and is suitable for intelligent patrol inspection requirements of railway trains.

Description

Technical Field

[0001] This invention relates to the technical field of train inspection, and more specifically to a navigation matching and correction method based on an inspection robot. Background Technology

[0002] With the increasing demand for railway train inspection, the application of inspection robots in self-service inspection is gradually becoming more widespread. However, existing inspection robot navigation technologies still have shortcomings in terms of path planning accuracy, environmental adaptability, and navigation matching and correction capabilities, making it difficult to meet actual needs in terms of inspection efficiency and accuracy.

[0003] A search revealed a patent with publication number CN112729302B, which discloses a navigation method, device, robot, and storage medium for an inspection robot. The publication date is March 29, 2024. This patent determines the robot's current position and performs navigation by acquiring an inspection area map and inspection path, and combining the matching degree between an estimated area reference map and the actual area map. However, the accuracy of navigation matching in this solution depends on the division of the estimated area and the construction of the reference map. If the estimated area is too large or too small, it may increase the matching error, thus affecting the accuracy of navigation. Furthermore, this solution does not consider a real-time correction mechanism in dynamic environments, which may lead to positioning deviations due to environmental changes in complex scenarios.

[0004] A search revealed a navigation method and device for an inspection robot, published on June 20, 2023, with publication number CN115979249B. This patent uses pre-built visual SLAM technology to generate a global map and selects the most energy-efficient inspection route based on an optimization function and energy loss calculation model. It also incorporates yaw angle for navigation control. However, this technical solution places high demands on sensor accuracy and the richness of environmental features in the global map construction. If environmental features are sparse or sensor data is noisy, the map construction may be inaccurate, affecting the accuracy of navigation path planning. Furthermore, this solution lacks a real-time response mechanism for dynamic obstacles, which may lead to path deviations or navigation failures in complex dynamic environments.

[0005] The aforementioned problems indicate that existing inspection robot navigation technologies still have certain shortcomings in terms of navigation matching accuracy, dynamic environment adaptability, and real-time correction capabilities. Therefore, this invention provides a navigation matching and correction method based on inspection robots, aiming to improve navigation matching accuracy, enhance real-time correction capabilities in dynamic environments, and optimize the flexibility and accuracy of path planning, thereby meeting the needs of the railway train inspection field for efficient and accurate navigation. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of this invention is to overcome the over-reliance on static environment models in the existing inspection robot navigation methods, as well as the lack of dynamic quantitative management and smoothing correction mechanisms when dealing with the continuous pose uncertainty caused by factors such as sensor drift and odometer cumulative error. Therefore, this invention provides a navigation matching and correction method based on inspection robots. This method aims to establish a closed-loop control system capable of dynamically quantifying, modeling, tracking, and correcting the confidence and accuracy of the robot's own pose information. The system introduces a dual time decay model to manage the lifecycle of confidence in the coarse positioning stage and accuracy in the fine positioning stage, respectively. Combined with high-confidence information fusion and adaptive decay parameter correction within a multi-robot collaborative framework, this fundamentally solves the navigation failure problem caused by the continuous accumulation of uncertainty, significantly improving the continuity, accuracy, and reliability of autonomous navigation for inspection robots in complex industrial environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A navigation matching and correction method based on an inspection robot includes the following steps: The inspection scene feature calibration step involves deploying physical calibrators at preset locations within the inspection area and identifying inherent natural features in the environment. The three-dimensional spatial coordinates and attitude parameters of the physical calibrators and natural features in the world coordinate system are obtained through three-dimensional scanning. The geometric type, three-dimensional coordinates, and attitude parameters of the features are used to form feature records, which are stored in the inspection robot's memory to form a feature and physical coordinate database. The robot coarse localization step involves synchronously acquiring image data streams and 3D point cloud data streams of the environment through visual sensors and lidar sensors. When at least one of the features is identified, the perspective n-point algorithm is used to calculate the 3D translation vector and rotation matrix of the current camera coordinate system relative to the identified feature. Combined with the feature and the coordinates and orientation of the feature in the physical coordinate database, the relative pose in the camera coordinate system is converted into the absolute pose of the inspection robot in the world coordinate system through a homogeneous transformation of the coordinate system. The quantitative calculation step of coarse localization confidence is to calculate an initial confidence level to characterize the reliability of the coarse localization result after each successful execution of the robot coarse localization step. The coarse positioning confidence decay step involves, after obtaining the initial confidence, dynamically changing the initial confidence over time according to a preset exponential decay model to obtain a real-time confidence that evolves over time, which is used to dynamically quantify and track the uncertainty of the inspection robot's positioning information.

[0008] Furthermore, the quantitative calculation step of the coarse positioning confidence level includes a basic score calculation sub-step, a basic score adjustment sub-step, a confidence level correction sub-step, and a sensor verification sub-step. The basic score calculation sub-step determines the basic score that is positively correlated with the number of SIFT feature matching point pairs used in the perspective n-point algorithm in the robot coarse localization step. The basic score adjustment sub-step determines the feature integrity weight based on the feature outline integrity or occlusion degree, and then performs a product adjustment on the basic score. The trust level correction sub-step determines the environmental adaptability correction factor based on the real-time light intensity or image clarity, and then performs product correction again on the base score after product adjustment. The sensor verification sub-step compares the feature positions calculated by the visual sensor and the lidar sensor, and adds or subtracts the confidence value based on whether the three-dimensional distance deviation between the two is within a preset threshold to obtain the initial confidence level.

[0009] Furthermore, the mathematical expression for the exponential decay model is as follows: ,in The attenuation coefficient is linearly mapped to the moving speed of the inspection robot obtained by the wheel speed odometer. Furthermore, the attenuation coefficient is also affected by the feature density of the area where the inspection robot is located. When the number of calibrated features within a preset radius of the current position of the inspection robot is greater than a preset threshold, the attenuation coefficient is multiplied by a first compensation coefficient less than 1. When the number of calibrated features within the preset radius is less than another preset threshold, the attenuation coefficient is multiplied by a second compensation coefficient greater than 1.

[0010] Furthermore, it also includes real-time trust monitoring and compensation triggering steps, and trust update and decay reset steps under multi-feature recognition; The real-time trust level monitoring and compensation triggering step monitors the value of the real-time trust level in real time. When the value of the real-time trust level decays to below the preset compensation trigger threshold, an active coarse positioning compensation behavior is automatically triggered to control the inspection robot to scan and search for new usable features. In the multi-feature recognition trust update and decay reset step, when the inspection robot continuously recognizes multiple features, if the newly calculated initial trust is higher than the current decayed real-time trust, the current trust is directly updated to the new initial trust, and the decay time is reset.

[0011] Furthermore, it also includes a secondary localization fusion correction step for multi-robot interaction. Each inspection robot transmits a data packet containing its own ID, timestamp, coarse localization pose, real-time trust level, and historical trajectory point sequence to other inspection robots at a preset period. When an inspection robot receives a data packet from another robot, it extracts the historical trajectory feature sequences of its own machine and the other robot, and uses the Dynamic Time Warping (DTW) algorithm to calculate the similarity between the two trajectory feature sequences. When the similarity meets a preset condition, it is determined that the other robot and the machine have a path overlap, and the localization data points transmitted by the other robot in the overlapping segment are used as reference data to be processed.

[0012] Furthermore, the secondary localization fusion correction step for multi-robot interaction also includes an effective cluster selection sub-step and a secondary localization correction sub-step. The effective cluster screening sub-step processes all the reference data to be processed using the density-based spatial clustering DBSCAN algorithm. It sets the neighborhood radius and minimum number of samples, divides the spatially close location data into clusters, and removes isolated data points that are judged as noise by the algorithm. Only the largest cluster with a data volume ratio exceeding the preset proportion is retained as effective reference data. The secondary positioning correction sub-step calculates the secondary positioning correction result of the local machine using a weighted fusion algorithm. The final fused coordinates of the local machine are determined by its own coarse positioning result and all valid reference data. The calculation formula is as follows: ,in This is the coarse positioning coordinate for the machine. For effective reference data coordinates, and Each represents its corresponding weight.

[0013] Furthermore, the secondary positioning correction sub-step also includes dividing the real-time trust level into multiple levels, and weighting them accordingly. and Dynamically correlated with their respective real-time trust levels, before the weighted average calculation, abnormal reference data points are checked and identified, and the weights of reference data points whose deviation from the cluster mean exceeds a preset multiple of the standard deviation are directly set to zero.

[0014] Furthermore, it also includes a fine positioning step. When the inspection robot confirms that it has entered the preset key component inspection area through positioning, the fine positioning step is activated. The inspection robot's camera acquires images of the pre-selected fine positioning features from at least three different spatial positions and angles, and extracts the two-dimensional key points of the fine positioning features. The two-dimensional key points are matched with the three-dimensional model of the feature stored in the feature and physical coordinate database. The relative pose transformation matrix between the current camera coordinate system and the fine positioning feature's own coordinate system is calculated. Combined with the coordinates of the feature in the world coordinate system in the database, the current fine positioning coordinates of the inspection robot are obtained.

[0015] Furthermore, it also includes dynamic attenuation and correction steps. The initial accuracy is calibrated based on the image matching error of the two-dimensional key points, and the initial accuracy is dynamically adjusted using the vibration acceleration amplitude collected by the inspection robot. The initial accuracy deteriorates over time according to a linear attenuation rule, and the attenuation expression for the accuracy value is: ,in for Precision value of time, For accuracy attenuation factor, This represents the initial precision.

[0016] Furthermore, the dynamic attenuation and correction step also includes storing recent fine positioning results through a sliding window. Whenever a new fine positioning result is obtained, the difference between the result and the historical results in the sliding window is calculated to obtain a difference sequence. The average value of the difference sequence is calculated, and the accuracy attenuation factor is adjusted up or down by analyzing the changing trend of the average value over multiple consecutive calculation cycles.

[0017] The beneficial effects of this invention are as follows: 1. By introducing a dual time decay model of coarse positioning confidence and fine positioning accuracy, dynamic quantification and lifecycle management of robot positioning information reliability are achieved. Combined with real-time monitoring and compensation triggering mechanisms, a closed-loop control system is formed, effectively solving the problem of continuously expanding positioning deviations caused by sensor drift or accumulated errors in traditional methods. This significantly improves the continuity and stability of navigation. Furthermore, by fusing multimodal data from visual sensors and LiDAR, and dynamically adjusting the confidence level in conjunction with real-time parameters such as ambient lighting and feature occlusion, the robustness of robot positioning in complex or dynamic environments is enhanced. Even if some features are occluded or lighting conditions change, the system can still maintain high positioning reliability through weighted fusion and attenuation compensation. 2. By sharing data and matching trajectories among multiple robots, and then using clustering and weighted fusion to achieve secondary positioning correction, the system avoids positioning failures caused by missing local features in a single robot. The group collaboration mechanism not only improves positioning accuracy but also enhances the system's fault tolerance in sparse feature environments. In addition, the trust decay coefficient and accuracy decay factor are dynamically correlated with the robot's motion state and environmental feature density, and can be adaptively adjusted according to the actual scenario. At the same time, by analyzing the difference trend of historical positioning results through a sliding window, the decay parameters are corrected in real time, further optimizing the long-term positioning accuracy. High-precision image matching positioning is triggered in key inspection areas to ensure accurate detection of important components. When features are missing or communication is interrupted, the system automatically switches to inertial navigation emergency mode and slows down trust decay, ensuring basic navigation functions under extreme conditions and improving the overall reliability of the system. Attached Figure Description

[0018] Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a flowchart of the sub-steps for quantifying the coarse positioning trust level in this invention; Figure 3 This is a flowchart of the secondary localization fusion correction process for multi-robot interaction in this invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0020] Because existing navigation technologies for inspection robots still have certain shortcomings in terms of navigation matching accuracy, dynamic environment adaptability, and real-time correction capabilities, this invention provides a navigation matching and correction method based on an inspection robot, applied to at least one inspection robot. The inspection robot is equipped with a central processing unit, a memory, at least one set of visual sensors, one set of lidar sensors, one set of inertial measurement units, and a wireless communication module. Figure 1 As shown, the method includes the following steps: First, the inspection scene's characteristic shapes are calibrated and 3D spatial modeled. This step provides a stable and identifiable geometric reference for the subsequent positioning process. Specifically, physical calibrators with unique geometric shapes and high optical contrast are deployed at preset key locations within the inspection area. Components inherent in the inspection environment and possessing stable geometric features are identified as natural features. The physical calibrators are red equilateral triangle calibration plates with sides of 8 cm. The natural features include hexagonal bolt groups on the train body and rectangular inspection port covers. A 3D laser scanner scans all physical calibrators and natural features to obtain their precise 3D spatial coordinates in a pre-established world coordinate system. The geometric type, three-dimensional coordinates, and attitude parameters of the feature are used to construct a feature record, and all feature records are stored in the local memory of the inspection robot to form a feature and physical coordinate database.

[0021] Meanwhile, an industrial camera is used to collect a set of image samples of various types of features under different lighting conditions and observation angles. The image samples are then input into a pre-built regional convolutional neural network target detection model for training. In addition, for each type of feature, its scale-invariant feature transform (SIFT) key points are extracted, and a standard SIFT feature descriptor template is generated and stored in the feature-physical coordinate database.

[0022] Secondly, coarse localization of the inspection robot based on feature recognition is performed. This step aims to provide the robot with an initial or periodic pose estimate with a controlled error range. The inspection robot is equipped with an industrial camera and a LiDAR. During the robot's movement, the industrial camera and LiDAR synchronously and continuously acquire image data streams and 3D point cloud data streams of the environment. The central processing unit performs Gaussian filtering for noise reduction and histogram equalization on the acquired image frames in real time, and then calls the pre-trained Faster algorithm. The R-CNN model detects the presence of labeled features in the processed image frames. Simultaneously, it applies Euclidean clustering to the acquired point cloud data stream to extract point cloud clusters with preset geometric dimensions matching the labeled features. When visual detection and point cloud segmentation collaboratively identify at least one feature, coarse localization calculation is immediately triggered. Specifically, using the camera intrinsic matrix and distortion coefficients, and employing a perspective n-point algorithm, the 3D translation vector and rotation matrix of the current camera coordinate system relative to the identified feature are calculated. Subsequently, the coordinates and pose of the feature in the world coordinate system are retrieved from the feature and physical coordinate database. Through the multiplication of homogeneous transformation matrices, the relative pose in the camera coordinate system is converted into the robot's absolute pose in the world coordinate system. .

[0023] Furthermore, such as Figure 2 As shown, a quantitative calculation and time decay mechanism for coarse localization confidence is introduced. This mechanism aims to quantitatively evaluate the reliability of each coarse localization result and simulate its time-dependent decay. After each successful coarse localization, an initial confidence level C0 is immediately calculated. The calculation of the initial confidence level incorporates multiple evaluation metrics. First, a base score is set based on the number of feature matching points. When the number of SIFT feature matching point pairs used for the perspective n-point algorithm is greater than or equal to 10, the base score is 0.9; when the number is between 5 and 9, the base score is 0.6; and when the number is less than 5, the base score is 0.3. Secondly, a feature integrity weight is introduced. If the integrity of the identified feature outline is higher than 90%, 0.1 is added to the base score; if there is less than 30% occlusion, 0.1 is deducted from the base score. Furthermore, an environmental adaptability correction factor is introduced. Real-time environmental parameters are obtained through the robot's ambient light and humidity sensors. When the light intensity is greater than N1 or less than N2, the current confidence level is multiplied by a correction factor of 0.8; when an image clarity degradation exceeding a preset threshold is detected, it is multiplied by a correction factor of 0.7. Finally, sensor consistency verification is performed. The positions of the features calculated by the visual n-point algorithm and the centroid matching of the LiDAR point cloud are compared. If the three-dimensional spatial distance deviation between the two is less than or equal to 5 cm, the confidence level is increased by 0.05; if the deviation is greater than 10 cm, the confidence level is decreased by 0.15. The final value calculated by all the above factors is used as the initial confidence level C0 for this coarse localization.

[0024] After obtaining the initial trust level C0, this trust level is not static, but rather changes dynamically with time t according to an exponential decay model. Its mathematical expression is: Among them, the attenuation coefficient It is a dynamic parameter whose value is related to the robot's state; specifically, A linear mapping relationship is established between the value and the robot's moving speed v; furthermore, the attenuation coefficient... It is also affected by the feature density of the area where the robot is located. If there are more than or equal to 3 calibrated features within a 5-meter radius of the robot's current position, then the currently calculated features will be... The value is multiplied by a compensation coefficient of 0.8; if there is less than one feature within a 10-meter radius, then... The value is multiplied by a compensation coefficient of 1.2.

[0025] When the value of C(t) decays to below the preset compensation trigger threshold of 0.3, the robot control system automatically triggers an active coarse positioning compensation action, controlling the robot gimbal to perform a 360-degree scan to search for new available features. On the monitoring interface of the robot control system, the decay curve of C(t) over time is plotted in real time, and the warning line (C=0.5) and the compensation trigger line (C=0.3) are marked. When the robot continuously identifies multiple features, the current confidence level is updated using the maximum value priority principle. That is, if the newly calculated C0 is higher than the current decayed C(t), the current confidence level of the system is directly updated to the new C0, and the decay time t is reset.

[0026] Furthermore, such as Figure 3 As shown, it also includes a secondary localization fusion correction method for multi-robot interaction. This method is applied to a cluster of multiple inspection robots. Each robot in the cluster constructs a low-latency, high-bandwidth real-time communication network through its onboard communication module or Mesh self-organizing network module. Each robot transmits a pose calculated from coarse localization to other robots in the network every 0.5 seconds, including its own ID, current timestamp, and position. The data packet contains the current trust level C(t) and the historical trajectory point sequence within the last 60 seconds.

[0027] When a robot (hereinafter referred to as "the local robot") receives a positioning data packet from another robot (hereinafter referred to as "the other robot"), it performs a secondary positioning fusion calculation. This calculation first performs path matching and reference data filtering. The local robot extracts its own historical trajectory feature sequence from the most recent 60 seconds. This sequence consists of a feature point containing a turning angle and velocity recorded every 20 centimeters. The local robot uses the Dynamic Time Warping (DTW) algorithm to calculate the similarity between its own trajectory feature sequence and the received trajectory feature sequence of the other robot. When the distance calculated by DTW is less than a preset threshold (indicating that the two paths are highly similar in shape, with a similarity greater than 80%), it is determined that the other robot has a path overlap with the local robot in the corresponding time period. Subsequently, broadcasts are made to all other robots with overlapping paths within the overlapping segment. The location data points were processed using the density-based spatial clustering (DBSCAN) algorithm. The neighborhood radius eps of DBSCAN was set to 2 meters, and the minimum number of samples minPts was set to 3. The spatially close location data were divided into clusters, and isolated data points judged as noise by the algorithm were removed. Only the largest cluster, whose data volume accounted for more than 70% of the total number of data points in the overlapping segment, was retained. The mean three-dimensional coordinates of all location points in the largest cluster were used as the effective reference data benchmark for this fusion calculation. Before fusion, all the selected data from other machines were time-synchronized and calibrated. For data with a timestamp deviation within 100 milliseconds, linear interpolation was used to correct the pose to the position aligned with the current timestamp of the local machine based on the velocity vector of the other machine.

[0028] After obtaining the filtered and calibrated valid reference data, a weighted fusion algorithm based on the idea of federated filtering is used to calculate the secondary positioning correction result of the local machine, and the final fused coordinates of the local machine are determined by its own rough positioning result and all valid reference data jointly, and its calculation formula is: , where the and are not fixed values, but are dynamically associated with their respective confidence levels C(t). Specifically, the weights and are positively correlated with the confidence level C.

[0029] Introduce a hierarchical incentive mechanism: divide the confidence level into three levels: high (C≥0.7), medium (0.3<C<0.7), and low (C≤0.3). When calculating the weights, if the confidence level of a data source (whether it is the local machine or other machines) belongs to the high level, its weight is multiplied by an incentive coefficient of 1.2 based on its confidence level value; if it belongs to the low level, it is multiplied by an inhibition coefficient of 0.5. Before performing the weighted average calculation, an outlier rejection based on Grubbs' test is also performed, and the weights of the reference data points whose deviation from the mean of the clustering cluster exceeds 3 times the standard deviation are directly set to zero. After the fusion calculation, not only the corrected coordinates are output, but also its 95% confidence interval is calculated. If the radius of this interval (i.e., the standard deviation σ) is greater than 50 cm, the system will trigger a request to increase the amount of reference data or shorten the data acquisition cycle for re-fusion.

[0030] Furthermore, the present invention also includes a fine positioning module based on high-precision image calibration. When the robot confirms through secondary positioning that it has entered a preset key component inspection area (for example, less than 1 meter away from the axle end of the target wheel set), this module is activated; the inspection robot is equipped with an industrial camera, and the robot controls the robotic arm to collect images of the preset fine positioning feature (for example, a laser-etched circular mark with a diameter of 10 mm at the center of the axle end of the wheel set) from at least three different spatial positions and angles. At the same time, the lidar performs high-density scanning on this feature to generate a local three-dimensional point cloud model; For each frame of high-precision image collected, the image processing unit extracts the 2D key points of the feature (for example, the exact center pixel coordinates and the edge point set of the circular mark); match these 2D key points with the three-dimensional model of this feature pre-stored in the feature and physical coordinate database, and apply a variant of the perspective-n-point algorithm - the EPnP algorithm to accurately calculate the relative pose transformation matrix between the current camera coordinate system and the feature's own coordinate system. This matrix contains the translation vector And rotation matrix. Combining the precise coordinates of the feature in the world coordinate system in the database, the fine positioning coordinates of the robot in the world coordinate system at the current moment are calculated through coordinate transformation. If multiple fine positioning features are identified at the same time, the multiple relative pose results calculated by EPnP are put into a least squares optimization framework. By minimizing the sum of the reprojection errors of all features, an optimal robot fine pose that integrates multiple feature information is solved.

[0031] Corresponding to the coarse positioning confidence decay model, this invention also establishes a dynamic decay and correction model for the accuracy of fine positioning. This model aims to quantify the accuracy degradation of fine positioning results over time caused by factors such as minor vibrations and IMU drift. After each fine positioning is completed, the system calibrates an initial accuracy S0 based on the quality of feature matching. For example, for highly reflective, clearly edged laser-etched marks, if the matching error of their key points in the image is less than 0.1 pixels, then S0 is calibrated to ±2 mm; for natural weld features on the train surface, if the matching error is between 0.1 and 0.3 pixels, then S0 is calibrated to ±5 mm. The initial accuracy S0 is also dynamically adjusted according to environmental parameters. For example, if the vibration acceleration amplitude detected by the IMU is greater than 0.5g, then a compensation of ±2 mm is added to the calibrated S0.

[0032] After obtaining the initial precision S0, this precision value will deteriorate with time t according to a linear decay rule, the mathematical expression of which is: ,in, This is the accuracy attenuation factor, initially set to 0.005 / second. This attenuation factor... It is not static, but rather adaptively corrects itself based on the robot's operating status and historical positioning performance. Specifically, The setting and scene association: When the robot moves smoothly, The value is 0.005 g / second; when the IMU detects that the robot's vibration acceleration is greater than 0.5 g, Automatically increased to 0.01 / second, accelerating accuracy decay; when the robot is stationary performing a fixed-point detection task, Reduced to 0.002 / second, slowing down the decay.

[0033] To achieve the attenuation factor For closed-loop correction, this invention employs an algorithm based on difference feedback. The system uses a sliding window of size 5 and a step size of 1 second to store the most recent 5 fine-tuning results (within 5 seconds). Whenever a new fine-tuning result is obtained, the difference between this result and the historical results within the sliding window is calculated to obtain a difference sequence. The average value of this difference sequence is then calculated. By analyzing multiple consecutive calculation cycles The changing trend, on Make corrections if It shows a monotonically increasing trend, indicating that the current If the value is set too small, the actual accuracy deteriorates faster than the model prediction. Therefore, a preset linear interpolation formula is used to... Adjustments will be made upwards: New Conversely, if A decreasing trend indicates If the setting is too large, then... Adjustments will be made: New ,like If the fluctuation is within a stable range of ±1 mm, then Remain unchanged. (Revised) Will be applied immediately The calculation, when three consecutive corrections When the value changes by more than 20%, the system will automatically trigger the sensor self-test process to check for hardware faults.

[0034] Finally, this invention forms a closed loop of dynamic compensation and optimization throughout the entire process. During the inspection, if the robot's vision system identifies a new feature that is not pre-registered in the database but conforms to preset geometric rules (such as circles or rectangles), the system will trigger a temporary coarse localization compensation. It will calculate the approximate world coordinates of the new feature based on the coordinates of the nearest known feature in the vicinity and use it for real-time localization to compensate for the accumulated error of the odometry. In multi-robot collaborative scenarios, if the coarse localization confidence level of one robot is high... If the value is below 0.3, it will proactively broadcast a location calibration request to the network. The network will then receive the request and, given its own trust level... A high-confidence robot with a confidence level higher than 0.7 will send its current fine-grained localization results and related local point cloud map fragments to the requester. Upon receiving the data, the requester uses the Iterative Closest Point (ICP) algorithm to align its own path point cloud, calculated based on odometry, with the received high-precision local map, thereby correcting its accumulated local positioning deviations significantly in one go. Furthermore, the system includes an emergency handling mechanism: if the robot fails to identify any known features for 5 consecutive seconds and receives no valid positioning data from other robots, the system will switch to a pure inertial navigation emergency mode, relying entirely on wheel speed odometry and the IMU for trajectory calculation, while simultaneously reducing the coarse-grained positioning confidence level. The value is temporarily set to a minimum of 0.001 to maximize the "shelf life" of the previous valid positioning result until the robot re-enters the feature area or restores network communication.

[0035] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A navigation matching and correction method based on an inspection robot, characterized in that: Includes the following steps: The inspection scene feature calibration step involves deploying physical calibrators at preset locations within the inspection area and identifying inherent natural features in the environment. The three-dimensional spatial coordinates and attitude parameters of the physical calibrators and natural features in the world coordinate system are obtained through three-dimensional scanning. The geometric type, three-dimensional coordinates, and attitude parameters of the features are used to form feature records, which are stored in the inspection robot's memory to form a feature and physical coordinate database. The robot coarse localization step involves synchronously acquiring image data streams and 3D point cloud data streams of the environment through visual sensors and lidar sensors. When at least one of the features is identified, the perspective n-point algorithm is used to calculate the 3D translation vector and rotation matrix of the current camera coordinate system relative to the identified feature. Combined with the feature and the coordinates and orientation of the feature in the physical coordinate database, the relative pose in the camera coordinate system is converted into the absolute pose of the inspection robot in the world coordinate system through a homogeneous transformation of the coordinate system. The quantitative calculation step of coarse localization confidence is to calculate an initial confidence level to characterize the reliability of the coarse localization result after each successful execution of the robot coarse localization step. The coarse positioning confidence decay step involves, after obtaining the initial confidence, dynamically changing the initial confidence over time according to a preset exponential decay model to obtain a real-time confidence that evolves over time, which is used to dynamically quantify and track the uncertainty of the inspection robot's positioning information.

2. The navigation matching and correction method based on an inspection robot according to claim 1, characterized in that: The quantitative calculation steps of the coarse positioning confidence level include a basic score calculation sub-step, a basic score adjustment sub-step, a confidence level correction sub-step, and a sensor verification sub-step. The basic score calculation sub-step determines the basic score that is positively correlated with the number of SIFT feature matching point pairs used in the perspective n-point algorithm in the robot coarse localization step. The basic score adjustment sub-step determines the feature integrity weight based on the feature outline integrity or occlusion degree, and then performs a product adjustment on the basic score. The trust level correction sub-step determines the environmental adaptability correction factor based on the real-time light intensity or image clarity, and then performs product correction again on the base score after product adjustment. The sensor verification sub-step compares the feature positions calculated by the visual sensor and the lidar sensor, and adds or subtracts the confidence value based on whether the three-dimensional distance deviation between the two is within a preset threshold to obtain the initial confidence level.

3. The navigation matching and correction method based on an inspection robot according to claim 1 or 2, characterized in that: The mathematical expression for the exponential decay model is: ,in The attenuation coefficient is linearly mapped to the moving speed of the inspection robot obtained by the wheel speed odometer. Furthermore, the attenuation coefficient is also affected by the feature density of the area where the inspection robot is located. When the number of calibrated features within a preset radius of the current position of the inspection robot is greater than a preset threshold, the attenuation coefficient is multiplied by a first compensation coefficient less than 1. When the number of calibrated features within the preset radius is less than another preset threshold, the attenuation coefficient is multiplied by a second compensation coefficient greater than 1.

4. The navigation matching and correction method based on an inspection robot according to claim 3, characterized in that: It also includes real-time trust monitoring and compensation triggering steps, and trust update and decay reset steps under multi-feature recognition; The real-time trust level monitoring and compensation triggering step monitors the value of the real-time trust level in real time. When the value of the real-time trust level decays to below the preset compensation trigger threshold, an active coarse positioning compensation behavior is automatically triggered to control the inspection robot to scan and search for new usable features. In the multi-feature recognition trust update and decay reset step, when the inspection robot continuously recognizes multiple features, if the newly calculated initial trust is higher than the current decayed real-time trust, the current trust is directly updated to the new initial trust, and the decay time is reset.

5. The navigation matching and correction method based on an inspection robot according to claim 1, characterized in that: It also includes a secondary localization fusion correction step for multi-robot interaction. Each inspection robot transmits a data packet containing its own ID, timestamp, coarse localization pose, real-time trust level, and historical trajectory point sequence to other inspection robots at a preset period. When an inspection robot receives a data packet from another robot, it extracts the historical trajectory feature sequences of its own machine and the other robot, and uses the Dynamic Time Warping (DTW) algorithm to calculate the similarity between the two trajectory feature sequences. When the similarity meets a preset condition, it is determined that the other robot and the machine have a path overlap, and the localization data points transmitted by the other robot in the overlapping segment are used as reference data to be processed.

6. The navigation matching and correction method based on an inspection robot according to claim 5, characterized in that: The secondary localization fusion correction step of the multi-robot interaction also includes an effective cluster selection sub-step and a secondary localization correction sub-step. The effective cluster screening sub-step processes all the reference data to be processed using the density-based spatial clustering DBSCAN algorithm. It sets the neighborhood radius and minimum number of samples, divides the spatially close location data into clusters, and removes isolated data points that are judged as noise by the algorithm. Only the largest cluster with a data volume ratio exceeding the preset proportion is retained as effective reference data. The secondary positioning correction sub-step calculates the secondary positioning correction result of the local machine using a weighted fusion algorithm. The final fused coordinates of the local machine are determined by its own coarse positioning result and all valid reference data. The calculation formula is as follows: ,in This is the coarse positioning coordinate for the machine. For effective reference data coordinates, and Each represents its corresponding weight.

7. The navigation matching and correction method based on an inspection robot according to claim 6, characterized in that: The secondary positioning correction sub-step further includes dividing the real-time trust level into multiple levels, and weighting them accordingly. and Dynamically correlated with their respective real-time trust levels, before the weighted average calculation, abnormal reference data points are checked and identified, and the weights of reference data points whose deviation from the cluster mean exceeds a preset multiple of the standard deviation are directly set to zero.

8. The navigation matching and correction method based on an inspection robot according to claim 1, characterized in that: It also includes a fine positioning step. When the inspection robot confirms that it has entered the preset key component inspection area through positioning, the fine positioning step is activated. The inspection robot uses a camera to capture images of the pre-selected fine positioning features from at least three different spatial positions and angles, and extracts the two-dimensional key points of the fine positioning features. The two-dimensional key points are matched with the three-dimensional model of the feature stored in the feature and physical coordinate database. The relative pose transformation matrix between the current camera coordinate system and the fine positioning feature's own coordinate system is calculated. Combined with the coordinates of the feature in the world coordinate system in the database, the current fine positioning coordinates of the inspection robot are obtained.

9. The navigation matching and correction method based on an inspection robot according to claim 8, characterized in that: It also includes dynamic attenuation and correction steps. The initial accuracy is calibrated based on the image matching error of the two-dimensional key points, and the initial accuracy is dynamically adjusted based on the vibration acceleration amplitude collected by the inspection robot. The initial accuracy deteriorates over time according to a linear attenuation rule, and the expression for the accuracy value attenuation is: ,in for Precision value of time, As the accuracy attenuation factor, This represents the initial precision.

10. The navigation matching and correction method based on an inspection robot according to claim 9, characterized in that: The dynamic attenuation and correction steps also include storing recent precise fine positioning results through a sliding window. Whenever a new fine positioning result is obtained, the difference between the result and the historical results in the sliding window is calculated to obtain a difference sequence. The average value of the difference sequence is calculated, and the precision attenuation factor is adjusted up or down by analyzing the changing trend of the average value over multiple consecutive calculation cycles.

Images