Unmanned monorail hoist positioning system based on machine vision

Abstract

The invention relates to the technical field of image processing, and discloses an unmanned monorail hoist positioning system based on machine vision. The system comprises a noise reduction module, a matching module, a filtering module, a tracking module, a sampling module and an evaluation module which are connected in series. The noise reduction module processes the image to obtain track characteristic data, the matching module generates three-dimensional characteristic data, the filtering module fuses multi-source data, the tracking module optimizes position data, the sampling module calculates motion parameters, and the evaluation module outputs correction data to the driving system. According to the method, the positioning efficiency and accuracy of the unmanned monorail hoist based on machine vision are improved.

Description

Technical Field

[0001] The present application relates to the field of image processing, and in particular to an unmanned monorail crane positioning system based on machine vision. Background Art

[0002] At present, underground monorail cranes in coal mines have been widely used in underground material transportation systems. The positioning of traditional monorail cranes mainly relies on manual operation or single sensors for navigation and positioning, such as laser ranging-based positioning systems, RFID-based positioning systems, and inertial navigation-based positioning systems. Among them, the laser ranging positioning system determines the position by measuring the distance between the crane and the tunnel wall, but it is easily affected by environmental factors such as dust and water vapor underground; the RFID positioning system requires a large number of tags to be laid along the track, and the maintenance cost is high; the inertial navigation system is prone to cumulative errors under the interference of strong magnetic fields underground, and the positioning accuracy is significantly reduced after long-term operation.

[0003] The main shortcomings of the existing technology are: the positioning method of a single sensor has poor anti-interference ability, and it is difficult to ensure the reliability and accuracy of positioning in the complex environment of the mine. In particular, in the harsh environment conditions of the mine, such as weak light, dust, humidity, and strong magnetic field, a single sensor is easily interfered or fails, resulting in an unstable positioning system. At the same time, the existing technology lacks an effective fusion mechanism for multi-source sensor data, cannot fully utilize the advantages of various sensors to improve positioning accuracy, and lacks adaptive processing capabilities for special environments underground. Summary of the invention

[0004] This application provides an unmanned monorail crane positioning system based on machine vision, which is used to solve the problem of how to achieve high-precision and reliable positioning of monorail cranes in complex underground environments. The present invention uses multi-sensor information fusion technology, combined with advanced algorithms such as machine vision processing and deep learning, to establish a complete positioning solution that can effectively overcome the adverse effects of the underground environment on the positioning system and significantly improve positioning accuracy and system reliability.

[0005] The present application provides an unmanned monorail crane positioning system based on machine vision, and the unmanned monorail crane positioning system based on machine vision includes: a noise reduction module, a matching module, a filtering module, a tracking module, a sampling module and an evaluation module;

[0006] The output end of the noise reduction module is connected to the input end of the matching module; the output end of the matching module is connected to the input end of the filtering module; the output end of the filtering module is connected to the input end of the tracking module; the output end of the tracking module is connected to the input end of the sampling module; the output end of the sampling module is connected to the input end of the evaluation module; the output end of the evaluation module is connected to the input end of the driving system;

[0007] The noise reduction module is used to reduce noise and enhance contrast of the acquired underground track image, and obtain track profile data, node position data and feature mark data through feature extraction and calculation;

[0008] A matching module, used to perform binocular matching calculation using the track profile data, node position data and feature mark data, and obtain three-dimensional feature data by parallax mapping conversion;

[0009] A filtering module is used to perform data fusion and filtering processing on the three-dimensional feature data, the inertial measurement data, and the distance measurement data to calculate and obtain initial position coordinate data;

[0010] A tracking module, used to track and calculate the track feature points according to the initial position coordinate data, and generate continuous position data through local area optimization processing;

[0011] A sampling module, used for performing trajectory sampling calculation according to the continuous position data, and obtaining motion parameter data through constraint optimization;

[0012] The evaluation module is used to perform real-time state evaluation calculation on the motion parameter data, generate correction data through deviation analysis processing, and send the correction data to the drive system.

[0013] In the technical solution provided by the present application, a high-precision and reliable positioning effect is achieved through the synergy of various modules. The noise reduction module reduces noise and enhances the contrast of the underground track image, effectively overcoming the problems of insufficient lighting and dust interference in the underground environment. The extracted track profile data, node position data and feature mark data provide a reliable basis for subsequent processing. The matching module converts the two-dimensional image features into three-dimensional feature data through binocular matching calculation and parallax mapping conversion, significantly improving the accuracy of spatial position perception. The filtering module fuses and filters the three-dimensional feature data with the inertial measurement data and distance measurement data, improves the robustness of positioning through multi-source data fusion, and the obtained initial position coordinate data has a high degree of credibility. The tracking module continuously tracks the track feature points, effectively eliminates the cumulative error through local area optimization, and the generated continuous position data ensures the continuity of positioning. The sampling module performs trajectory sampling calculation based on the continuous position data, and the motion parameter data obtained by combining the constraint optimization ensures the feasibility of motion planning. The evaluation module detects and corrects positioning deviations in a timely manner through real-time status evaluation and deviation analysis. The output correction data can ensure the continuous and stable operation of the system and realize the precise positioning of the monorail crane in the complex environment underground. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0015] Figure 1 A schematic diagram of an embodiment of an unmanned monorail crane positioning system based on machine vision in an embodiment of the present application;

[0016] Figure 2 Schematic diagram of the structure of the noise reduction module in the unmanned monorail crane positioning system based on machine vision in the embodiment of the present application. DETAILED DESCRIPTION

[0017] The embodiment of the present application provides an unmanned monorail crane positioning system based on machine vision. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0018] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the unmanned monorail crane positioning system based on machine vision includes:

[0019] The noise reduction module 101 is used to reduce noise and enhance contrast of the acquired downhole track image, and obtain track profile data, node position data and feature mark data through feature extraction and calculation;

[0020] A matching module 102 is used to perform binocular matching calculation using track profile data, node position data and feature mark data, and obtain three-dimensional feature data by parallax mapping conversion;

[0021] The filtering module 103 is used to perform data fusion and filtering processing on the three-dimensional feature data, the inertial measurement data, and the distance measurement data, and calculate the initial position coordinate data;

[0022] Tracking module 104, used to track and calculate the track feature points according to the initial position coordinate data, and generate continuous position data through local area optimization processing;

[0023] The sampling module 105 is used to perform trajectory sampling calculation according to the continuous position data and obtain motion parameter data through constraint optimization;

[0024] The evaluation module 106 is used to perform real-time state evaluation calculation on the motion parameter data, generate correction data through deviation analysis processing, and send the correction data to the drive system.

[0025] It is understandable that the execution subject of the present application can be an unmanned monorail crane positioning system based on machine vision, or a terminal or a server, which is not limited here. The present application embodiment is described by taking the server as the execution subject as an example.

[0026] Specifically, the noise reduction module 101 first processes the collected downhole track image, realizes multi-scale decomposition through wavelet transform, decomposes the image into different frequency components, and performs hierarchical processing on the Gaussian noise and salt and pepper noise commonly found in the downhole environment. In practical applications, for an original image with a resolution of 1920×1080, 16 sub-band images are obtained after 4 layers of wavelet decomposition, and each sub-band contains information in a specific frequency range. The coefficients of each sub-band are screened based on the threshold judgment method to generate a noise coefficient matrix, which records the noise distribution characteristics in the image. After obtaining the noise coefficient matrix, the noise reduction module 101 enhances the image through adaptive histogram equalization technology. Adaptive histogram equalization is different from traditional global histogram equalization. Instead, it divides the image into multiple local areas, and calculates and optimizes the contrast of each area separately. For dark areas, its brightness value is increased and the detail information is maintained; for bright areas, it is appropriately suppressed to avoid overexposure. The enhanced image data obtained after this processing has a more uniform grayscale distribution and more obvious features.

[0027] Then, the noise reduction module 101 uses the Canny edge detection operator to perform edge detection on the enhanced image and extract the contour information of the track. The Canny operator first uses a Gaussian filter to smooth the image, then calculates the gradient amplitude and direction of the image, and obtains complete edge information through the double threshold method and edge connection. After edge detection, the initial data of the track contour is obtained, which is represented in the form of a binary image, where the value of the edge point is 255 and the value of the background point is 0. After obtaining the initial data of the track contour, the matching module 102 extracts the key node position through curve fitting technology. First, the Douglas-Peucker algorithm is used to simplify the contour line, retain the key turning points, and then the corner detection algorithm is used to identify the connection points and change points of the track. These points constitute the initial data of the node position. The node position data is transformed into the actual physical coordinate system through spatial coordinate transformation, which is convenient for subsequent three-dimensional reconstruction.

[0028] The converted initial node position data is clustered and analyzed, and the matching module 102 uses the DBSCAN density clustering algorithm to combine similar feature points into feature point groups according to the spatial distance and point density. The DBSCAN algorithm does not need to pre-specify the number of categories, can effectively identify point groups of any shape, and has good robustness to noise points. Finally, the feature point groups are marked and paired, the correlation between the points is calculated, and the final track profile data, node position data and feature marking data are obtained by feature merging. In the matching module 102, the image collected by the binocular camera is firstly corrected for the epipolar line to eliminate the distortion caused by the camera installation error, so that the corresponding points are on the same horizontal line, which is convenient for subsequent stereo matching. The corrected feature data is regionally matched, and the SAD (Sum of Absolute Differences) algorithm is used to calculate the similarity of the corresponding areas of the left and right images to generate the initial disparity data.

[0029] After depth constraint judgment and detail compensation processing, the filtering module 103 removes erroneous matching points and obtains reliable parallax data. Based on the principle of triangulation, the parallax data is converted into three-dimensional space coordinates to establish point cloud data. The point cloud data is statistically filtered to remove outliers and noise points to improve data quality. Finally, a spatial feature descriptor is constructed to complete feature matching and fusion, and accurate three-dimensional feature data is output. The filtering module 103 receives data from the inertial measurement unit, identifies the stationary state of the crane through the zero-speed detection algorithm, and compensates for the zero bias error of the gyroscope and accelerometer. At the same time, the three-dimensional feature data is fused with the inertial data, and the Kalman filter framework is used to calculate the optimal estimate. In practice, the state vector of the filter contains information such as position, speed, and attitude, and the positioning result is continuously optimized through a prediction-update cycle.

[0030] Tracking module 104 uses pyramid optical flow algorithm to track feature points, and improves the robustness and accuracy of tracking through multi-layer image pyramid. The feature point sequence obtained by tracking is optimized by sliding window to eliminate cumulative errors. The Lie group and Lie algebra theory are used for posture interpolation to ensure the continuity and accuracy of posture estimation. Sampling module 105 uses Bezier curve to smooth the trajectory to ensure the continuity and smoothness of the motion trajectory. The boundary conditions of velocity and acceleration are determined by dynamic constraints to ensure that the motion parameters meet the physical limitations of the crane. The recursive least squares method is used to extract the main features of the trajectory, and a polynomial curve family is established to describe the feasible motion path.

[0031] The evaluation module 106 monitors the system status in real time, extracts the intrinsic characteristics of the motion parameters through variational decomposition, and establishes a Gaussian mixture model to evaluate the system status. The conditional entropy criterion is used for decision analysis, and the convergence of the control system is guaranteed by the Lyapunov stability theory. Finally, the control parameters are identified and optimized online by the recursive least squares method.

[0032] For example, in the process of positioning a monorail crane in a mine, the noise reduction module 101 collects the original image with a resolution of 1920×1080, and obtains 16 sub-band images after 4-layer wavelet decomposition. After adaptive histogram equalization processing, the contrast of the image is improved by 40%, and the edge features are clearer. After Canny edge detection, 50 key nodes are extracted, and 15 feature point groups are formed after DBSCAN clustering. In the matching module 102, the matching window size of the SAD algorithm is set to 11×11 pixels, and the accuracy of the generated disparity map reaches the sub-pixel level. The filtering module 103 integrates IMU data and visual features, and the root mean square error of position estimation is controlled within the range of ±5 cm. The tracking module 104 uses a 3-layer image pyramid for feature tracking, and the accuracy of posture estimation is better than 0.5 degrees. The curvature of the motion trajectory generated by the sampling module 105 is continuous, the maximum speed does not exceed 2 meters / second, and the acceleration is limited to 0.5 meters / second squared. The control cycle of the evaluation module 106 is 20 milliseconds, and both real-time performance and stability are guaranteed.

[0033] In the embodiment of the present application, through the synergy of each module, a high-precision and reliable positioning effect is achieved. The noise reduction module 101 performs noise reduction and contrast enhancement on the underground track image, effectively overcoming the problems of insufficient lighting and dust interference in the underground environment. The extracted track profile data, node position data and feature mark data provide a reliable basis for subsequent processing. The matching module 102 converts the two-dimensional image features into three-dimensional feature data through binocular matching calculation and parallax mapping conversion, significantly improving the accuracy of spatial position perception. The filtering module 103 fuses and filters the three-dimensional feature data with the inertial measurement data and the distance measurement data, improves the robustness of positioning through multi-source data fusion, and the obtained initial position coordinate data has a high degree of credibility. The tracking module 104 continuously tracks the track feature points, effectively eliminates the cumulative error through local area optimization, and the generated continuous position data ensures the continuity of positioning. The sampling module 105 performs trajectory sampling calculation based on the continuous position data, and the motion parameter data obtained by combining the constraint condition optimization ensures the feasibility of motion planning. The evaluation module 106 detects and corrects positioning deviations in a timely manner through real-time status evaluation and deviation analysis. The output correction data can ensure the continuous and stable operation of the system and realize the precise positioning of the monorail crane in the complex environment underground.

[0034] In a specific embodiment, if Figure 2 As shown, the noise reduction module 101 specifically includes:

[0035] A decomposition unit 1011 is used to perform multi-scale decomposition on the downhole trajectory image by wavelet transform, and obtain a noise coefficient matrix in combination with threshold judgment;

[0036] A calculation unit 1012, configured to generate enhanced image data through adaptive histogram equalization calculation according to the noise coefficient matrix;

[0037] A tracking unit 1013 is used to perform edge detection and contour tracking using the enhanced image data, and obtain initial track contour data through gradient operator calculation;

[0038] An input unit 1014 is used to input the initial data of the track profile into the feature analyzer, and obtain the initial data of the node position through curve fitting and feature point extraction;

[0039] The screening unit 1015 is used to perform spatial coordinate transformation and feature screening on the initial data of the node positions, and obtain feature point group data through cluster analysis;

[0040] The merging unit 1016 is used to perform label pairing and correlation calculation on the feature point group data, and output track profile data, node position data and feature label data through feature merging operation.

[0041] Specifically, the decomposition unit 1011 first performs wavelet transform to perform multi-scale decomposition on the downhole track image. Wavelet transform is a time-frequency analysis method that realizes multi-resolution analysis of image details by decomposing the image into sub-bands of different scales. Specifically, for the input downhole track image, a two-dimensional discrete wavelet transform is used for decomposition, and the image is decomposed into low-frequency approximate components and high-frequency detail components in turn. The high-frequency component mainly contains noise and edge information, while the low-frequency component retains the main structural information of the image.

[0042] The mathematical expression of the wavelet transform decomposition process is:

[0043] W j,k (x,y)=∑ m ∑ n f(m,n)ψ j,k (mx,ny)

[0044] Where: W j,k (x, y) represents the wavelet coefficient at scale j and direction k; f(m, n) represents the gray value of the original image at point (m, n); ψ j,k represents the wavelet basis function; (x, y) represents the image coordinate position.

[0045] After wavelet decomposition, the calculation unit 1012 performs threshold judgment on the coefficients of each scale to obtain a noise coefficient matrix. The threshold judgment adopts a soft threshold method to adaptively set the threshold according to the noise level of each scale. Next, based on the obtained noise coefficient matrix, an adaptive histogram equalization calculation is performed to generate enhanced image data. Adaptive histogram equalization is different from traditional global histogram equalization. Instead, the image is divided into multiple local areas, and the contrast of each area is calculated and optimized separately.

[0046] The calculation formula of adaptive histogram equalization is:

[0047]

[0048] Where: g(i,j) represents the gray value of the equalized image at position (i,j); L represents the gray level; M×N represents the size of the local area; h(x,y) represents the gray value of the original image at position (x,y).

[0049] After generating enhanced image data, the tracking unit 1013 processes the image using edge detection and contour tracking technology. Edge detection uses an improved Canny operator, which includes four steps: Gaussian filtering, gradient calculation, non-maximum suppression, and double threshold detection. The initial data of the track contour is obtained by gradient operator calculation. The specific calculation of the gradient operator adopts the following formula:

[0050]

[0051] Where: G x ,G y Represent the gradients in the x and y directions respectively; I(x,y) represents the grayscale value of the image at the point (x,y); * represents the convolution operation.

[0052] The gradient amplitude is calculated by calculate.

[0053] For the initial data of the track profile, the input unit 1014 performs curve fitting and feature point extraction through the feature analyzer to obtain the initial data of the node position. The curve fitting adopts the cubic spline interpolation method to ensure the smoothness and continuity of the fitting curve. The feature point extraction is based on the curvature calculation, and the key points are extracted at the position with the larger curvature value. The curvature calculation formula for feature point extraction is:

[0054]

[0055] Where: κ(t) represents the curvature of the curve at parameter t; x-(t), y-(t) represent the first-order derivative of the curve at t; x . (t),y . (t) represents the second derivative of the curve at point t.

[0056] Then, the screening unit 1015 performs spatial coordinate transformation and feature screening on the initial data of the node position. The spatial coordinate transformation converts the two-dimensional image coordinates into three-dimensional physical coordinates, using the following transformation matrix:

[0057]

[0058] Among them: (X3, Y3, Z3) represents the coordinates of the point in the world coordinate system; (x : ,y : ,z : ) represents the point coordinates in the image coordinate system; β ij (i,j=1,2,3) represents the rotation matrix element; γ k (k=1, 2, 3) represents the translation vector component.

[0059] The merging unit 1016 obtains the feature point group data through cluster analysis, and adopts the density-based clustering algorithm DBSCAN, which automatically determines the number and range of clusters by calculating the distance relationship and density distribution characteristics between points. Finally, the feature point group data is marked and paired and the correlation degree is calculated. The pairing process adopts the Hungarian algorithm, and the correlation degree calculation is based on the distance and direction information between the feature points. Finally, the track profile data, node position data and feature mark data are output through feature merging operation.

[0060] For example, in the process of underground monorail crane positioning, a 1920×1080 resolution original image is decomposed by four layers of wavelet to generate 16 subbands. After threshold judgment processing, the size of the noise coefficient matrix is ​​960×540. Adaptive histogram equalization uses a 16×16 local window for processing, and the clarity of the enhanced image is significantly improved. The low threshold of Canny edge detection is set to 50 and the high threshold is set to 150, and the complete track profile is successfully extracted. The feature analyzer extracts about 200 initial feature points, and retains 50 key nodes after curvature screening. The accuracy of spatial coordinate transformation reaches the millimeter level. The final feature point group contains 15 stable feature clusters, each of which contains 3-5 key points. The spatial distribution of these feature points can accurately reflect the geometric characteristics of the track.

[0061] In a specific embodiment, the matching module 102 is specifically configured to:

[0062] (1) Using track profile data, node position data and feature marker data to perform binocular matching calculations, and converting disparity mapping to obtain three-dimensional feature data, including:

[0063] (2) calculating and correcting the track profile data and node position data through polar line constraint optimization to obtain corrected feature data;

[0064] (3) Calculate the similarity between the corrected feature data and the feature tag data based on regional cost aggregation, and output the initial disparity data;

[0065] (4) screening the initial disparity data through depth constraint determination, and generating effective disparity data through detail compensation processing;

[0066] (5) Mapping the effective disparity data to the spatial coordinate system by triangulation to obtain the initial point cloud data;

[0067] (6) Remove outliers from the initial point cloud data according to the density clustering criterion and generate point cloud optimized data through noise reduction processing;

[0068] (7) Construct a spatial feature descriptor based on the point cloud optimization data, complete feature matching and fusion, and output three-dimensional feature data.

[0069] Specifically, the matching module 102 processing process involves multiple key links. First, the track contour data, node position data and feature mark data output by the noise reduction module 101 are received. These data contain the geometric shape of the track, the key node position and the feature mark information used for matching. Epipolar constraint optimization is an important step in binocular stereo vision. The track contour data and node position data are corrected through epipolar constraint optimization to ensure that the corresponding points in the left and right images are on the same horizontal line, reducing the search range of stereo matching. Epipolar constraint optimization uses basic matrix calculation and epipolar equation. During the calculation process, the basic matrix is ​​first estimated, and then the epipolar geometric relationship is calculated based on the basic matrix to correct the epipolar deviation caused by image distortion and camera installation error. For the corrected feature data, the regional cost aggregation method is used to calculate the similarity. Regional cost aggregation is the core step in stereo matching. The matching cost matrix is ​​constructed by calculating the grayscale difference, gradient difference and other features of the corresponding areas of the left and right images. During the calculation process, the sliding window method is used. The window size is usually set to 11×11 or 15×15 pixels, and the cost value is calculated for each candidate matching point. The cost calculation uses a combination of multiple metrics such as SAD (Sum of Absolute Differences) and Census transform to improve the robustness of the matching. The spatial position relationship and color similarity of pixels are considered during cost aggregation, and different weights are assigned to different pixels in the window using an adaptive weight method.

[0070] After obtaining the initial disparity data, it is screened through depth constraint judgment. Depth constraint judgment is based on the geometric features of the scene, and uses prior information such as track installation height and lane width to set a reasonable depth range. For disparity values ​​that exceed the depth range or violate geometric constraints, they are corrected through detail compensation processing. Detail compensation processing includes steps such as left-right consistency check, disparity continuity check, and occlusion area processing. Through iterative optimization, abnormal disparity values ​​are eliminated, disparity holes are filled, and continuous and reliable valid disparity data is generated. Triangulation is a key step in converting disparity data into three-dimensional space coordinates. Based on the calibrated camera parameters, a mapping relationship between two-dimensional image coordinates and three-dimensional space coordinates is established. During the triangulation process, the depth information is calculated through the disparity value, and each feature point is mapped to the world coordinate system in combination with the internal and external parameters of the camera to generate the initial three-dimensional point cloud data. Each point in the point cloud data contains spatial coordinates and corresponding feature attribute information.

[0071] For the initial point cloud data obtained, the density clustering criterion is used to remove outliers. Density clustering is based on the DBSCAN algorithm. By setting the neighborhood radius and the minimum point number threshold, noise points and outliers are identified and removed. In the denoising process, statistical filtering and radius filtering are combined to remove points with abnormal spatial distribution and retain representative point cloud data. In this way, the quality and reliability of point cloud data are significantly improved. After obtaining the optimized point cloud data, it is necessary to construct a spatial feature descriptor to achieve feature matching and fusion. The spatial feature descriptor uses a local feature description method based on normal vectors and curvature to calculate the geometric features and topological relationships of each point. Feature matching uses nearest neighbor search and feature similarity measurement to align and fuse point cloud data acquired at different times or from different perspectives, and finally output complete three-dimensional feature data.

[0072] For example, in the actual monorail crane positioning application, when the binocular camera collects a pair of 1920×1080 resolution stereo images, they are first corrected by epipolar constraint optimization. During the correction process, the basic matrix is ​​estimated by the RANSAC algorithm, more than 100 pairs of feature points are selected for calculation, the number of iterations is set to 1000, and the threshold of the inlier ratio is set to 0.8. The alignment error of the corrected image is controlled within 1 pixel. The regional cost aggregation uses a window size of 15×15, combined with the hybrid cost metric of SAD and Census transform to generate the initial disparity map. The effective depth range set by the depth constraint judgment is 0.5 meters to 5 meters, and the disparity values ​​beyond the range are marked as invalid. Through detail compensation processing, the effective pixel rate of the disparity map is increased from the initial 85% to 95%. During the triangulation process, the disparity data is converted into spatial coordinates based on the calibrated camera focal length of 2000 pixels and baseline length of 120 mm. The neighborhood radius of density clustering was set to 0.05 meters, and the minimum number of points was 10, which successfully eliminated about 5% of outliers. The final generated three-dimensional feature data contains more than 1,000 valid feature points, which are evenly distributed in space and accurately reflect the three-dimensional geometric structure of the track.

[0073] In a specific embodiment, the filtering module 103 is specifically used for:

[0074] (1) Compensate the zero bias error of the inertial measurement data based on zero velocity detection, and generate the initial inertial data through quaternion transformation and integration operation;

[0075] (2) Establishing spatial geometric constraints for the three-dimensional feature data, and outputting three-dimensional calibration data by feature point projection and depth calibration calculation;

[0076] (3) Calculate the weight coefficients of the inertial initial data and the three-dimensional calibration data according to the Kalman filter framework, and generate a fusion weight matrix through optimization iteration;

[0077] (4) Probabilistically resample the distance measurement data and the fusion weight matrix according to the covariance propagation criterion to extract the optimal estimation data;

[0078] (5) Perform time series correlation analysis on the optimal estimation data using Bayesian reasoning, and obtain position estimation data through particle swarm optimization algorithm;

[0079] (6) The uncertainty of the position estimation data is quantified and compensated through the ellipsoid error distribution, and the initial position coordinate data is output in combination with the multi-sensor consistency constraint.

[0080] Specifically, the filtering module 103 performs multi-sensor data fusion and filtering processing, and first processes the data from the inertial measurement unit (IMU). Zero speed detection is a key step in processing IMU data. By analyzing the output signal characteristics of the accelerometer and the gyroscope, it is determined whether the crane is in a stationary state. When the stationary state is detected, the measured value at this moment is used to estimate and compensate the zero bias of the IMU online. The data after zero bias compensation is updated through quaternion transformation, and then the initial inertial data is obtained through integral operation. These data contain the position, speed and attitude information of the crane. When processing three-dimensional feature data, strict spatial geometric constraints need to be established. These constraints include the spatial layout characteristics of the track (such as the standard cross-sectional size of the track, the installation height, etc.) and kinematic constraints (such as the freedom of motion limit of the crane). By projecting the three-dimensional feature points onto these constraint planes and performing depth calibration, it is ensured that the spatial position of the feature points conforms to the actual scene characteristics. During the depth calibration process, the depth information measured by binocular vision is used as a reference to optimize the spatial coordinates of the feature points and output calibrated three-dimensional data.

[0081] The Kalman filter framework is the core algorithm for realizing multi-source data fusion. In this framework, the initial inertial data is used as the system state prediction value, and the three-dimensional calibration data is used as the observation value, and the optimal estimate is calculated by recursion. The calculation of the weight coefficient is based on the measurement noise characteristics and dynamic performance of each sensor, reflecting the credibility of different data sources in the fusion process. Through multiple iterative optimizations, a fusion weight matrix reflecting the reliability of each data source is generated. When the distance measurement data is obtained, it is processed according to the covariance propagation criterion. Covariance propagation describes the transmission law of measurement errors in the coordinate transformation and data fusion process. The distance measurement data is combined with the aforementioned fusion weight matrix for probability resampling to obtain the optimal estimate data. The probability resampling process is based on the Monte Carlo method. By random sampling, high-weight data samples are retained and unreliable measurements are eliminated.

[0082] Bayesian reasoning introduces prior knowledge to perform time series analysis on the optimal estimation data. During the movement of the monorail crane, the positions at adjacent moments have a strong correlation. By establishing a state transition probability model, the correlation between the current state and the historical state is calculated. The particle swarm optimization algorithm is used to search for the optimal solution in the high-dimensional state space. Each particle represents a possible state estimation. By iteratively updating the position and velocity of the particle, it finally converges to the optimal solution and obtains reliable position estimation data. Finally, the ellipsoid error distribution model is used to perform uncertainty analysis on the position estimation data. The ellipsoid error distribution can more accurately describe the position error characteristics in three-dimensional space, which is more in line with the actual situation than the spherical error distribution. The uncertainty of the positioning result is quantified by calculating the main axis direction and size of the error ellipsoid. On this basis, the multi-source data from vision, IMU and distance sensors are combined to establish multi-sensor consistency constraints, and the final initial position coordinate data is output through weighted fusion.

[0083] For example, in actual application scenarios, when the crane speed is lower than 0.01 m / s for more than 1 second, the zero speed detection module determines it as stationary. At this time, 100 sets of IMU data are collected for zero bias estimation. The typical value of the gyroscope zero bias is about 0.02 degrees / s, and the accelerometer zero bias is about 0.01 m / s2. The initial trajectory is obtained by integrating the compensated inertial data. The three-dimensional feature data provided by the visual system contains about 200 feature points, and 150 valid points are retained after spatial geometric constraint processing. The prediction cycle of the Kalman filter is 10 milliseconds. After 20 iterations of optimization, the diagonal element values ​​of the fusion weight matrix are distributed between 0.6 and 0.9, indicating the credibility of different sensors. The laser ranging data is input at a frequency of 50 Hz, with a measurement range of 0.1 m to 10 m and an accuracy of ±2 mm. When performing probability resampling, 1000 particles are kept, and the convergence result is obtained after 50 generations of iterations. The standard deviation of the final output position coordinate data in the horizontal direction is controlled within ±1 cm, and the standard deviation in the vertical direction is controlled within ±0.5 cm, which fully meets the accuracy requirements of crane positioning.

[0084] In a specific embodiment, the tracking module 104 is specifically configured to:

[0085] (1) Extract feature descriptors from the initial position coordinate data through pyramid layering, and generate feature matching data through optical flow tracking algorithm;

[0086] (2) Segment the feature matching data using the sliding window criterion, and output the segmented trajectory data by local optimization calculation;

[0087] (3) Performing posture interpolation operation on the segmented trajectory data according to the Lie algebraic space, and obtaining a key frame sequence after completing key frame selection;

[0088] (4) Construct a local covariance matrix for the key frame sequence, eliminate the accumulated error through the sparse optimization algorithm, and output the optimized trajectory data;

[0089] (5) Using hidden Markov chains to calculate the state transition probability of the optimized trajectory data, and obtaining the position prediction data through maximum likelihood estimation;

[0090] (6) Based on the graph optimization framework, the position prediction data and historical observations are globally constrained, and continuous position data is generated through iterative optimization.

[0091] Specifically, the tracking module 104 processes the initial position coordinate data based on the pyramid layering mechanism. Pyramid layering is a multi-scale image representation method that constructs an image pyramid by downsampling layer by layer, with the bottom layer maintaining the original resolution and each upper layer image size halved. When extracting feature descriptors from the pyramid structure, the FAST (Features from Accelerated Segment Test) corner detection algorithm is used to perform feature detection on each layer of the image. Optical flow tracking uses the Lucas-Kanade algorithm to calculate the motion vector of feature points between adjacent frames based on the assumption of image brightness constancy. The sliding window is an efficient data processing mechanism. When processing feature matching data, a fixed-size time window is set, and only the data within the window is optimized. The selection of the window size requires a trade-off between computational efficiency and optimization accuracy, and usually contains 3-5 frames of data. When performing local optimization within the window, a graph-based optimization method is used to construct the observation constraints and motion model constraints of the feature points into an optimization objective function, and the segmented trajectory data is obtained by least squares solution.

[0092] Lie algebra space provides a theoretical basis for processing continuous posture transformation. The SO(3) Lie group represents three-dimensional rotation, and its corresponding Lie algebra SO(3) is in Euclidean space, which is convenient for posture interpolation calculation. When interpolating segmented trajectory data, the posture data is first mapped to the Lie algebra space, and the intermediate posture is obtained by linear interpolation, and then mapped back to the Lie group space. The selection of key frames is based on the posture change and feature point parallax. When the posture change exceeds the threshold or the average parallax is large, the current frame is marked as a key frame.

[0093] The local covariance matrix reflects the uncertainty of trajectory estimation. For the key frame sequence, the covariance matrix of each key frame pose is calculated to build an error model that includes position and pose uncertainty. Sparse optimization uses the geometric constraint relationship between key frames to build an edge-based graph structure, and uses the g2o (General Graph Optimization) framework for optimization and solution, effectively eliminating cumulative errors and outputting coherent trajectory data. The hidden Markov chain model describes the temporal dependency of the trajectory. The state transition probability matrix represents the transition relationship between the trajectory states at adjacent moments, and the transition probability parameters are estimated by statistical learning methods. Combined with maximum likelihood estimation, the most likely state sequence is calculated under the condition of a given observation sequence to predict the future position of the crane.

[0094] Finally, in the graph optimization framework, the position prediction data is added as a new node to the global pose graph, and a constraint relationship is established with the historical observation data. The nodes in the graph represent the pose state, and the edges represent the observation constraints and motion constraints. Through the iterative optimization method, the errors of all constraints are minimized to generate smooth and continuous position data.

[0095] For example, when processing a 1920×1080 resolution image sequence, a 4-layer image pyramid is constructed, with the top layer having a resolution of 120×68. FAST corner features are extracted at each layer, and the corner response threshold is set to 40, resulting in about 1000 feature points. Optical flow tracking uses a 21×21 search window, with a brightness difference threshold set to 0.3, and a tracking success rate of 95%. The sliding window size is set to 5 frames, and the number of feature points in the window is kept above 800. When performing Lie algebraic interpolation, the rotation threshold for keyframe selection is set to 5 degrees and the translation threshold is set to 0.1 meters. The eigenvalues ​​of the local covariance matrix show that the standard deviation of the position estimate is controlled within ±2 cm in the horizontal direction and ±1 cm in the vertical direction. The number of states of the hidden Markov model is set to 10, and the transition probability matrix is ​​estimated using 1000 sets of training data. The final position tracking accuracy is better than ±5 cm, and the attitude accuracy is better than ±0.5 degrees.

[0096] In a specific embodiment, the sampling module 105 is specifically used to:

[0097] (1) Perform trajectory fitting and interpolation on continuous position data based on Bezier curves, and obtain smooth trajectory data through curvature constraint optimization;

[0098] (2) Calculate the velocity-acceleration boundary of the smooth trajectory data through dynamic constraints, and generate dynamic reference data in combination with the safety distance threshold;

[0099] (3) Decomposing the dynamic reference data into principal component features using the recursive least squares criterion, and outputting the trajectory principal metadata through feature reconstruction;

[0100] (4) Construct a polynomial curve family based on the trajectory master metadata, and calculate the optimal path data using a nonlinear programming solver;

[0101] (5) Apply kinematic constraints to the optimal path data through the Lagrange multiplier method, and generate trajectory control data through iterative optimization;

[0102] (6) The trajectory control data is converted into velocity, acceleration, and angular velocity sequences using a variational reasoning framework, and the motion parameter data is output in combination with dynamic boundary constraints.

[0103] Specifically, the sampling module 105 processes the continuous position data through the Bezier curve. The Bezier curve has good smoothness and continuity and is an ideal tool for trajectory planning. For n control points, an n-1 order Bezier curve is constructed, and the selection of control points is based on the key points of the position data. Curvature constraint optimization ensures that the trajectory meets the steering capability of the crane. By setting the maximum curvature threshold, the curve is locally adjusted so that the curvature of all points is within the safe range, thereby obtaining smooth trajectory data that meets the motion constraints. The dynamic constraints take into account the physical characteristics of the crane, including the maximum speed limit, acceleration limit and deceleration limit. Based on the smooth trajectory data, the speed and acceleration boundaries of each segment of the trajectory are calculated. During the calculation process, the slope, curve and load conditions of the track are considered, and the speed upper limit is dynamically adjusted. The safety distance threshold is determined based on the braking performance of the crane and the obstacle situation ahead. By combining these constraints, dynamic reference data containing time information is generated.

[0104] Recursive least squares is used to extract the main features of the trajectory. The dynamic reference data is regarded as a high-dimensional time series, and the parameter estimation is updated through recursive calculation to reduce the redundancy of the data. The feature extraction process focuses on the key changes in the trajectory, such as velocity change points, acceleration change points, etc. The dimension is reduced through principal component analysis, the most representative feature components are retained, and the simplified trajectory principal metadata is reconstructed. A family of polynomial curves is constructed based on the trajectory principal metadata, and polynomials of different orders are used to describe possible motion paths. The polynomial coefficients are optimized by a nonlinear programming solver, and the optimization objectives include path smoothness, motion continuity, energy consumption and other aspects. The solution process uses the interior point method to find the optimal solution that meets all constraints through iterative calculation.

[0105] The Lagrange multiplier method adds kinematic constraints as hard constraints to the optimization problem. These constraints include the maximum steering angle, maximum angular velocity and other restrictions. By constructing the Lagrange function, the constrained optimization problem is converted into an unconstrained optimization problem. During the iterative optimization process, the original variables and Lagrange multipliers are updated simultaneously until they converge to a solution that satisfies all constraints and obtain executable trajectory control data. The variational reasoning framework converts trajectory control data into specific motion instructions. Through variational calculations, the velocity, acceleration and angular velocity values ​​at each time point are obtained. These motion parameters must meet the dynamic boundary constraints of the crane, including motor torque limits, mechanical strength limits, etc. The final output motion parameter data contains complete timing information to ensure the smoothness and safety of the motion process.

[0106] For example, for a 50-meter-long track motion planning, a 5th-order Bezier curve is used for trajectory fitting, and the control point interval is set to 5 meters. The curvature constraint is set to a maximum value of 0.2 / meter to ensure that the turning radius is not less than 5 meters. In the dynamic constraints, the maximum speed limit is 2 meters / second, the maximum acceleration is 0.5 meters / square second, the maximum deceleration is 0.8 meters / square second, and the safety distance threshold is set to 3 meters. After recursive least squares processing, the first 5 principal components are retained, accounting for more than 95% of the total variance. The polynomial curve family uses a 7th-order polynomial, and the nonlinear programming solver iterates 50 times to reach convergence. The final generated motion parameter data shows that during the 50-meter movement, the maximum speed is 1.8 meters / second, the average acceleration is 0.3 meters / square second, and the maximum angular velocity is 0.15 radians / second. The whole process is smooth and controllable, meeting the requirements of safe operation.

[0107] In a specific embodiment, the evaluation module 106 is specifically configured to:

[0108] (1) Extract the intrinsic modes of motion parameter data through variational decomposition, and obtain state characteristic data through singular spectrum analysis;

[0109] (2) constructing a probability density function for the state feature data based on the Gaussian mixture distribution, and generating state clustering data by density clustering operation;

[0110] (3) Mapping the state clustering data to the decision space using the conditional entropy criterion, and outputting the state evaluation data through information gain calculation;

[0111] (4) Construct a robust controller based on the state evaluation data and obtain the deviation compensation data through Lyapunov stability analysis;

[0112] (5) Perform error propagation analysis on the deviation compensation data based on adaptive iterative learning, and obtain corrected control data by feedback correction calculation;

[0113] (6) Parameter identification and optimization of the modified control data are performed through recursive least squares, and the correction data are output in combination with the system dynamics constraints.

[0114] Specifically, the evaluation module 106 first uses a variational decomposition method to process motion parameter data. Variational decomposition is a signal processing method that can decompose complex signals into components with different frequency characteristics. In the process of intrinsic mode extraction, a series of intrinsic mode functions are obtained by iterative screening and decomposition, and each mode function represents a characteristic component of motion. Singular spectrum analysis further performs spectrum analysis on these modes, extracts the main frequency components and energy distribution, and generates state characteristic data reflecting the dynamic characteristics of the system. The Gaussian mixture model is used to perform probabilistic modeling on the state characteristic data. The probability density function of the state space is constructed by weighted combination of multiple Gaussian distributions. Each Gaussian component represents a state of motion, and the weight coefficient reflects the probability of occurrence of each state. The density clustering operation is based on the EM (expectation maximization) algorithm, which iteratively optimizes the parameters of the Gaussian components until convergence to the optimal state clustering result.

[0115] The conditional entropy criterion evaluates the uncertainty of state clustering and quantifies the degree of certainty of state transition by calculating the predictive power of the current state for the future state. When mapping the state clustering data to the decision space, information gain is used as the evaluation metric to select the most discriminative features for mapping. The information gain calculation is based on the reduction of information entropy and outputs data containing the state evaluation results. The robust controller design adopts H∞ control theory, considering the uncertainty of the system model and external disturbances. The design goal of the controller is to maintain the stability and performance of the system under the worst case. Lyapunov stability analysis proves the stability of the system under disturbance by constructing a suitable Lyapunov function and calculates the required deviation compensation.

[0116] Adaptive iterative learning control is optimized for repetitive tasks. Based on historical operation data, the error propagation law of deviation compensation data is analyzed and an error model is established. During the feedback correction process, the control parameters are adjusted according to the real-time error, the control accuracy is improved through iterative optimization, and the corrected control data is generated.

[0117] The recursive least squares method is used to update the system parameter estimation in real time. The online parameter identification is performed on the modified control data, and the weight of the historical data is processed by the forgetting factor to ensure that the model can adapt to the changes in the system parameters. Finally, the control parameters are optimized in combination with the dynamic constraints of the crane, such as maximum load and speed limit, and the final correction data is output.

[0118] For example, when processing motion parameter data with a sampling frequency of 100 Hz, the variational decomposition obtains 5 main eigenmodes, with an energy share of more than 98%. The Gaussian mixture model uses 3 Gaussian components and converges after 50 EM iterations. The eigenvalue of the covariance matrix of each component is between 0.01 and 0.1. Conditional entropy calculation shows that the uncertainty of state prediction is reduced by 60%, and the information gain reaches 0.8. The H∞ norm of the robust controller is less than 1.5, and the derivative of the Lyapunov function is always negative, proving the stability of the system. After 10 iterations of adaptive iterative learning, the tracking error is reduced from the initial 5 cm to less than 1 cm. The recursive least squares uses a forgetting factor of 0.98, and the parameter error of real-time estimation is kept below 2%. The final output correction data makes the system control accuracy reach ±0.5 cm.

[0119] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0120] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An unmanned monorail crane positioning system based on machine vision, characterized in that: The unmanned monorail crane positioning system based on machine vision includes: a noise reduction module, a matching module, a filtering module, a tracking module, a sampling module and an evaluation module; The output end of the noise reduction module is connected to the input end of the matching module; the output end of the matching module is connected to the input end of the filtering module; the output end of the filtering module is connected to the input end of the tracking module; the output end of the tracking module is connected to the input end of the sampling module; the output end of the sampling module is connected to the input end of the evaluation module; the output end of the evaluation module is connected to the input end of the driving system; The noise reduction module is used to reduce noise and enhance contrast of the acquired underground track image, and obtain track profile data, node position data and feature mark data through feature extraction and calculation; A matching module, used to perform binocular matching calculation using the track profile data, node position data and feature mark data, and obtain three-dimensional feature data by parallax mapping conversion; A filtering module is used to perform data fusion and filtering processing on the three-dimensional feature data, the inertial measurement data, and the distance measurement data to calculate and obtain initial position coordinate data; A tracking module, used to track and calculate the track feature points according to the initial position coordinate data, and generate continuous position data through local area optimization processing; A sampling module, used for performing trajectory sampling calculation according to the continuous position data, and obtaining motion parameter data through constraint optimization; The evaluation module is used to perform real-time state evaluation calculation on the motion parameter data, generate correction data through deviation analysis processing, and send the correction data to the drive system.

2. The unmanned monorail crane positioning system based on machine vision according to claim 1, characterized in that: The noise reduction module comprises: A decomposition unit, used for performing multi-scale decomposition on the downhole trajectory image by wavelet transform, and obtaining a noise coefficient matrix in combination with threshold judgment; A calculation unit, configured to generate enhanced image data through adaptive histogram equalization calculation according to the noise coefficient matrix; A tracking unit, used to perform edge detection and contour tracking using the enhanced image data, and obtain initial track contour data through gradient operator calculation; An input unit, used for inputting the initial data of the track profile into a feature analyzer, and obtaining the initial data of the node position through curve fitting and feature point extraction; A screening unit, used for performing spatial coordinate transformation and feature screening on the initial data of the node positions, and obtaining feature point group data through cluster analysis; The merging unit is used to perform label pairing and correlation calculation on the feature point group data, and output track profile data, node position data and feature label data through feature merging operation.

3. The unmanned monorail crane positioning system based on machine vision according to claim 1, characterized in that: The matching module is specifically used for: The method of performing binocular matching calculation using the track profile data, node position data and feature mark data, and converting the three-dimensional feature data by disparity mapping, includes: The track profile data and the node position data are calculated and corrected through polar line constraint optimization to obtain corrected feature data; Performing similarity calculation on the corrected feature data and the feature marking data based on regional cost aggregation, and outputting initial disparity data; The initial disparity data is screened by depth constraint determination, and effective disparity data is generated by detail compensation processing; Mapping the effective disparity data to a spatial coordinate system by triangulation to obtain initial point cloud data; Eliminate outliers from the initial point cloud data according to a density clustering criterion, and generate point cloud optimized data through noise reduction processing; A spatial feature descriptor is constructed for the point cloud optimization data, and three-dimensional feature data is output after feature matching and fusion.

4. The unmanned monorail crane positioning system based on machine vision according to claim 1, characterized in that: The filtering module is specifically used for: Compensating the zero bias error of the inertial measurement data based on zero speed detection, and generating inertial initial data through quaternion transformation and integral operation; Establishing spatial geometric constraints for the three-dimensional feature data, and outputting three-dimensional calibration data by feature point projection and depth calibration calculation; Calculating weight coefficients for the initial inertial data and the three-dimensional calibration data according to a Kalman filter framework, and generating a fusion weight matrix through optimization iteration; Probabilistically resampling the distance measurement data and the fusion weight matrix according to the covariance propagation criterion to extract optimal estimation data; Performing time series correlation analysis on the optimal estimation data by using Bayesian reasoning, and obtaining position estimation data by screening using a particle swarm optimization algorithm; The uncertainty of the position estimation data is quantified and compensated through ellipsoid error distribution, and the initial position coordinate data is output in combination with multi-sensor consistency constraints.

5. The unmanned monorail crane positioning system based on machine vision according to claim 1, characterized in that: The tracking module is specifically used for: Extracting feature descriptors from the initial position coordinate data through pyramid layering, and generating feature matching data through an optical flow tracking algorithm; The feature matching data is processed in segments according to a sliding window criterion, and segmented trajectory data is output by local optimization calculation; Performing posture interpolation operation on the segmented trajectory data according to Lie algebraic space, and obtaining a key frame sequence after completing key frame selection; Constructing a local covariance matrix for the key frame sequence, eliminating accumulated errors through a sparse optimization algorithm, and outputting optimized trajectory data; Calculating the state transition probability of the optimized trajectory data by using a hidden Markov chain, and obtaining position prediction data by maximum likelihood estimation; The position prediction data and historical observations are globally constrained according to a graph optimization framework, and continuous position data is generated through iterative optimization.

6. The unmanned monorail crane positioning system based on machine vision according to claim 1, characterized in that: The sampling module is specifically used for: Performing trajectory fitting and interpolation on the continuous position data based on the Bezier curve, and obtaining smooth trajectory data through curvature constraint optimization; Calculating the velocity-acceleration boundary of the smooth trajectory data through dynamic constraints, and generating dynamic reference data in combination with a safety distance threshold; Decomposing the dynamic reference data into principal component features by recursive least squares criterion, and outputting trajectory principal metadata by feature reconstruction; Constructing a polynomial curve family for the trajectory master metadata, and calculating the optimal path data by a nonlinear programming solver; Applying kinematic constraints to the optimal path data via a Lagrange multiplier method, and generating trajectory control data through iterative optimization; The trajectory control data is converted into a velocity, acceleration and angular velocity sequence using a variational reasoning framework, and motion parameter data is output in combination with dynamic boundary constraints.

7. The unmanned monorail crane positioning system based on machine vision according to claim 1, characterized in that: The evaluation module is specifically used for: Extracting the intrinsic modes of the motion parameter data by variational decomposition, and obtaining state characteristic data by singular spectrum analysis; Constructing a probability density function for the state feature data according to a Gaussian mixture distribution, and generating state clustering data by density clustering operation; Mapping the state clustering data to the decision space using the conditional entropy criterion, and outputting the state evaluation data through information gain calculation; A robust controller is constructed for the state evaluation data, and deviation compensation data is obtained through Lyapunov stability analysis; Performing error propagation analysis on the deviation compensation data based on adaptive iterative learning, and obtaining corrected control data by feedback correction calculation; The modified control data is subjected to parameter identification and optimization via recursive least squares, and correction data is output in combination with system dynamics constraints.

Images