Measuring cylinder liquid level intelligent identification and analysis system based on computer vision and method thereof

Abstract

The invention discloses a measuring cylinder liquid level intelligent identification and analysis system based on computer vision and a method thereof, and relates to the technical field of industrial monitoring. The system comprises a measuring cylinder type liquid level meter, a fixed camera, a mobile augmented reality terminal, a visual anchor point and a data processing server, the server side calculates the spatial pose of the mobile terminal based on the visual anchor point, calculates ideal observation position coordinates for eliminating parallax, and sends an instruction to guide the mobile terminal to move to the orthogonal position to collect an image, so that a high-precision liquid level numerical value is calculated; the system uses the high-precision numerical value as a reference truth value which is not only used for verifying the sensor state of the distributed control unit, but also used for reversely updating the parallax correction parameter of the fixed camera. According to the invention, observation geometric parallax is eliminated through AR dynamic guidance, and closed-loop online calibration of fixed monitoring equipment is realized by using heterogeneous data.

Description

Technical Field

[0001] This invention relates to the field of industrial monitoring technology, specifically to a computer vision-based intelligent identification and analysis system and method for measuring cylinder liquid level. Background Technology

[0002] In petrochemical and industrial production processes, magnetic float level gauges or glass tube level gauges are widely used for level indication in on-site storage tanks due to their intuitive and visual characteristics. Although modern industrial sites are usually equipped with remote transmitters to transmit level signals to distributed control systems, various electronic sensors are prone to zero drift or nonlinear errors after long-term operation. Therefore, regular comparison and calibration based on the readings of on-site level gauges is a key step in ensuring production safety.

[0003] Current on-site calibration primarily relies on manual inspections, but this method has inherent limitations. Due to variations in the height of inspectors and the haphazard positioning of their positions, it's often difficult to ensure that the line of sight is perpendicular to the level gauge's scale and level with the liquid surface. This non-orthogonal observation angle introduces significant geometric parallax, leading to low accuracy and inconsistent manual readings. On the other hand, to achieve automated monitoring, some sites have deployed machine vision reading solutions based on fixed cameras. However, limited by the complex pipe rack structures and installation space in industrial sites, fixed cameras are often difficult to install in ideal positions directly facing the level gauge, typically requiring installation on high poles or side supports. This fixed, tilted observation angle not only introduces initial image distortion and geometric parallax, but also causes the preset perspective transformation parameters to gradually become ineffective when environmental temperature changes or equipment vibrations cause minor deformations in the camera bracket, resulting in accumulated monitoring errors over time. Existing technical solutions lack a mechanism that can effectively guide the acquisition of parallax-free true values ​​and use these values ​​for online adaptive calibration of fixed monitoring equipment. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a computer vision-based intelligent identification and analysis system and method for graduated cylinder liquid levels. This system solves the technical problems in existing technologies, such as large geometric parallax caused by inconsistent observation angles during manual inspections, and inherent parallax due to installation space limitations of fixed monitoring equipment, as well as the difficulty in online calibration of long-term operating parameter drift.

[0005] This invention provides a computer vision-based intelligent identification and analysis system for graduated cylinder liquid levels. The system includes a graduated cylinder liquid level gauge, a fixed camera, a mobile augmented reality terminal, visual anchor points, a data processing server, and a distributed control unit. The visual anchor points are fixed to the outside of the graduated cylinder liquid level gauge, serving as absolute coordinate references for the local spatial location.

[0006] The data processing server is used to perform the following operations: acquire the video stream of the graduated cylinder level gauge using the fixed camera and obtain the first liquid level value; calculate the spatial pose of the mobile augmented reality terminal relative to the graduated cylinder level gauge based on the visual anchor point, calculate the ideal observation position coordinates to eliminate parallax, and send a guidance command to the mobile augmented reality terminal to instruct the mobile augmented reality terminal to move to the ideal observation position coordinates; identify the target image acquired by the mobile augmented reality terminal at the ideal observation position coordinates, and calculate the high-precision liquid level value; use the high-precision liquid level value as the reference true value, compare it with the first liquid level value identified by the fixed camera, and update the parallax correction parameters of the fixed camera according to the comparison result.

[0007] Furthermore, this invention also provides a computer vision-based intelligent identification and analysis method for graduated cylinder liquid levels, applied to the aforementioned system. This method first acquires a first liquid level value and a remote instrument value through a fixed camera and a distributed control unit, respectively. The data processing server monitors the deviation of these values ​​in real time. When the absolute value of the deviation exceeds a preset safety threshold, it determines that the current monitoring status is abnormal, generates an on-site verification task, and sends it to the mobile augmented reality terminal.

[0008] In response to the verification task, the mobile augmented reality terminal acquires environmental images containing visual anchor points. The data processing server uses the PnP algorithm to calculate the extrinsic parameter matrix of the mobile augmented reality terminal relative to the visual anchor points and establishes a local spatial coordinate system. Based on the position of the common perpendicular between the current camera line-of-sight vector of the mobile augmented reality terminal and the axis of the connecting pipe of the graduated cylinder level gauge, a rough liquid level value is calculated.

[0009] Subsequently, the data processing server determines the target point on the axis of the connecting pipe based on the approximate liquid level value. This target point is then translated a preset observation distance along the normal vector direction of the scale plate plane to obtain the ideal observation position coordinates, eliminating parallax. The mobile augmented reality terminal calculates the spatial deviation between its real-time position and the ideal observation position coordinates, and renders a virtual guide graphic in the display module until its real-time position enters the low parallax tolerance range centered on the ideal observation position coordinates.

[0010] When the mobile augmented reality terminal is within the low parallax tolerance range, it automatically acquires image sequences. The data processing server performs a multi-factor weighted quality evaluation on the image sequences. This evaluation is calculated based on the weighted sum of parallax scoring factors, glare interference factors, and sharpness factors, and the image with the highest comprehensive score is selected as the target image. A homography matrix is ​​constructed using the extrinsic parameter matrix acquired from the target image and the normal vector of the scale plane. An inverse perspective transformation is performed on the target image to generate an orthogonal view, and a high-precision liquid level value is obtained based on a neural network model.

[0011] Finally, the system uses the high-precision liquid level value as the baseline true value. When it is close to the first liquid level value but deviates significantly from the value of the remote instrument, it determines that the sensor corresponding to the distributed control unit has a drift fault. Simultaneously, the system uses the high-precision liquid level value to deduce the current parallax correction coefficient observation value of the fixed camera, and uses an exponentially weighted moving average algorithm to update the historical parallax correction coefficient of the fixed camera using this observation value.

[0012] This invention provides a computer vision-based intelligent identification and analysis system and method for graduated cylinder liquid levels. It has the following beneficial effects: 1. This invention establishes a local spatial coordinate system through visual anchor points and calculates the coordinates of the ideal observation position that can eliminate parallax, guiding the mobile augmented reality terminal to enter the orthogonal observation position. This mechanism forces the line of sight to be perpendicular to the scale plate plane and flush with the liquid surface, eliminating the geometric parallax caused by looking up, down, or side from the physical optical path, ensuring the geometric accuracy of the image data at the source of acquisition, and solving the problems of inconsistent perspectives in traditional manual inspection and large reading deviations caused by the limited installation position of fixed cameras.

[0013] 2. This invention utilizes the high-precision liquid level value obtained by the mobile terminal at an ideal observation position as the reference true value, and reversely derives and updates the parallax correction parameters of the fixed camera. This closed-loop calibration mechanism based on heterogeneous data enables the fixed camera to automatically adapt to parameter drift caused by slight deformation of the equipment, loose installation, or changes in ambient temperature, realizing online self-calibration without the need for manual on-site calibration, and maintaining the data accuracy of the fixed monitoring equipment during long-term operation.

[0014] 3. This invention introduces a multi-factor image quality optimization mechanism based on an evaluation function, which comprehensively considers parallax, reflectivity, and sharpness indicators to automatically select the best frame from a continuous image sequence for recognition. Combined with inverse perspective transformation and neural network algorithms, this scheme effectively reduces the interference of glass tube reflection under complex lighting conditions and the effect of motion blur during handheld shooting, thereby improving the robustness of the system in recognizing fine scales and liquid surface features in harsh industrial environments. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a comparative data graph showing the impact of observation angle deviation on the accuracy of liquid level measurement in embodiments of the present invention; Figure 4 This is a schematic diagram of the adaptive multi-frame optimization process based on the quality evaluation function in an embodiment of the present invention; Figure 5This is a curve showing the accuracy and stability verification of the fixed monitoring terminal under long-term operation in an embodiment of the present invention.

[0016] The components include: 1. Measuring cylinder level gauge; 2. Fixed camera; 3. Mobile augmented reality terminal; 4. Visual anchor point; 5. Data processing server; 51. Video stream access module; 52. Spatial computing module; 53. Image analysis module; 54. Data fusion and verification module; and 6. Distributed control unit. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See attached document Figure 1 , Figure 1 This is a schematic diagram of the overall architecture of a computer vision-based intelligent identification and analysis system for graduated cylinder liquid levels according to an embodiment of the present invention. The present invention provides a computer vision-based intelligent identification and analysis system for graduated cylinder liquid levels. This system is deployed in industrial environments containing graduated cylinder liquid level gauges and achieves accurate reading and multi-dimensional verification of liquid level data through a combination of fixed monitoring and mobile spatial calculation.

[0019] The computer vision-based intelligent liquid level recognition and analysis system for graduated cylinders includes: a graduated cylinder level gauge 1, a fixed camera 2, a mobile augmented reality terminal 3, visual anchor points 4, a data processing server 5, and a distributed control unit 6. All the above hardware components establish data communication connections via industrial Ethernet or wireless LAN.

[0020] The graduated cylinder level gauge 1 is the field instrument to be tested, typically installed vertically on the side of a pressure vessel or process pipeline. The graduated cylinder level gauge 1 includes a transparent connecting tube and a parallel scale indicator. Visual anchor points 4 are affixed or rigidly fixed to the external bracket or adjacent fixed wall of the graduated cylinder level gauge 1. Visual anchor points 4 are high-contrast industrial QR codes or AprilTag tags, whose encoded information is associated with the device tag number and physical geometric parameters of the graduated cylinder level gauge 1. Visual anchor points 4 serve as the absolute coordinate reference for the local field space; the coordinates of their center point in the world coordinate system are predefined, and visual anchor points 4 have a fixed relative positional relationship with the glass tube axis and the scale plane of the graduated cylinder level gauge 1.

[0021] A fixed camera 2 is mounted on a fixed bracket in front of the graduated cylinder level gauge 1, with its optical lens aimed at the area of ​​liquid level change in the graduated cylinder level gauge 1. The fixed camera 2 is used to continuously acquire on-site video images and transmit the video stream to the data processing server 5 via the RTSP protocol. The position and orientation of the fixed camera 2 remain relatively stationary after installation, providing basic daily monitoring data.

[0022] The mobile augmented reality terminal 3 is worn or held by on-site inspection personnel and is a head-mounted display device or smart glasses with spatial positioning capabilities. The mobile augmented reality terminal 3 integrates an image acquisition sensor, an inertial measurement unit, and a near-eye display module. The image acquisition sensor acquires high-resolution environmental texture images; the inertial measurement unit senses the device's acceleration and angular velocity to assist in motion tracking; and the near-eye display module overlays virtual guidance graphics onto the inspection personnel's field of vision. The mobile augmented reality terminal 3 internally runs real-time localization and mapping algorithms, capable of resetting and calibrating its six-degree-of-freedom pose in the world coordinate system by identifying visual anchor points 4.

[0023] The data processing server 5 is the core computing platform of the system, and its physical form may include a local server, an edge computing gateway, or a cloud computing platform. The data processing server 5 is used to receive video stream data from the fixed camera 2, image and spatial pose data from the mobile augmented reality terminal 3, and to interact with the distributed control unit 6.

[0024] The data processing server 5 is logically divided into multiple functional modules, including: video stream access module 51, spatial calculation module 52, image analysis module 53, and data fusion and verification module 54.

[0025] The video stream access module 51 is used to establish an RTSP connection with the fixed camera 2, and to decode and extract frames from the video stream according to a preset time interval to generate a sequence of still images to be analyzed. In addition, the video stream access module 51 is also used to receive high-definition captured images uploaded by the mobile augmented reality terminal 3.

[0026] The spatial calculation module 52 is used to process spatial geometric relationships. Based on the real-time pose data uploaded by the mobile augmented reality terminal 3 and the spatial geometric parameters of the graduated cylinder level gauge 1, the spatial calculation module 52 calculates the parallax angle of the mobile augmented reality terminal 3 relative to the graduated cylinder level gauge 1. The spatial calculation module 52 is also used to calculate the coordinates of the ideal observation position for eliminating parallax based on the current liquid level height data, and generate corresponding virtual guidance commands to be sent to the mobile augmented reality terminal 3.

[0027] Image analysis module 53 is used to run deep learning algorithms and computer vision algorithms. Image analysis module 53 includes a pre-trained object detection network and a semantic segmentation network, used to identify the level gauge region, locate the pixel coordinates of the gas-liquid interface, and identify scale values ​​from the input image. Image analysis module 53 is also used to perform image quality assessment, calculating image sharpness indicators and the proportion of reflective areas to select the optimal recognition frame.

[0028] The data fusion verification module 54 is used to perform comparison and logical judgment of multi-source data. The data fusion verification module 54 connects to the distributed control unit 6 via the MQTT or OPC protocol to read the remote instrument value corresponding to the graduated cylinder level gauge 1. The data fusion verification module 54 performs a weighted comparison of the recognition result from the fixed camera 2, the precise measurement result from the mobile augmented reality terminal 3, and the remote value transmitted by the distributed control unit 6. When the data deviation exceeds a preset threshold, an alarm signal is generated, and the calibration parameters of the fixed camera 2 are calculated based on the high-precision data from the mobile augmented reality terminal 3.

[0029] See attached document Figure 2 , Figure 2 This is a flowchart illustrating a collaborative monitoring and verification method according to an embodiment of the present invention. The present invention provides a computer vision-based intelligent identification and analysis method for graduated cylinder liquid levels. This method is implemented based on the aforementioned computer vision-based intelligent identification and analysis system for graduated cylinder liquid levels and mainly includes the following steps: S100. Fixed-end daily monitoring and anomaly triggering. A fixed camera 2 continuously captures images of the graduated cylinder level gauge 1, transmitting the video data in real-time to the data processing server 5. The video stream access module 51 extracts frames from the video data at a preset frequency to obtain static images. The image analysis module 53 identifies the static images and calculates the current first liquid level value. The data fusion verification module 54 reads the remote instrument value from the distributed control unit 6 and compares the first liquid level value with the remote instrument value. When the absolute value of the difference exceeds a preset alarm threshold, the data processing server 5 generates a field verification task and pushes the task to the mobile augmented reality terminal 3.

[0030] S200, Mobile Terminal Spatial Registration and Device Anchoring. The mobile augmented reality terminal 3 enters the field area where the graduated cylinder level gauge 1 is located. The image acquisition sensor of the mobile augmented reality terminal 3 captures the visual anchor point 4 arranged in the field. Based on the feature information of the visual anchor point 4, the mobile augmented reality terminal 3 establishes a local spatial coordinate system and determines the initial position and attitude of the mobile augmented reality terminal 3 relative to the visual anchor point 4. The spatial calculation module 52 retrieves the physical geometric parameters of the graduated cylinder level gauge 1 from the database according to the coded index of the visual anchor point 4. These parameters include the spatial position equation of the glass tube axis relative to the visual anchor point 4 and the physical height range of the scale plate.

[0031] S300. Rough Liquid Level Estimation and Spatial Inverse Calculation. The mobile augmented reality terminal 3 takes a preliminary picture of the graduated cylinder level gauge 1 from any observation position. The image analysis module 53 identifies the pixel coordinates of the gas-liquid interface in the preliminary picture. The spatial calculation module 52 combines the current pose data of the mobile augmented reality terminal 3 and the camera intrinsic parameters, uses a ray projection algorithm to establish a line-of-sight ray, calculates the intersection point of this ray with the cylindrical model of the graduated cylinder level gauge 1, and thus obtains a rough second liquid level value.

[0032] S400, Dynamic Orthogonal Viewpoint Guidance. The spatial calculation module 52 calculates the ideal observation position coordinates required to eliminate parallax based on the second liquid level value. These ideal observation position coordinates are located on a horizontal plane passing through the height of the second liquid level value and perpendicular to the direction of gravity, and are located in the normal direction of the scale surface of the graduated cylinder level gauge 1. The spatial calculation module 52 calculates the spatial deviation vector between the current position of the mobile augmented reality terminal 3 and the ideal observation position coordinates, generating a virtual guidance graphic instruction. The mobile augmented reality terminal 3 renders the virtual guidance graphic in the near-eye display module, instructing the operator to adjust their head height and horizontal position until the optical axis of the mobile augmented reality terminal 3 enters the preset low parallax tolerance range.

[0033] S500, Adaptive Multi-Frame Optimization and Fine Measurement. When the mobile augmented reality terminal 3 enters the low parallax tolerance range, it automatically starts the multi-frame acquisition mode. The image analysis module 53 performs quality evaluation on the acquired image sequence, calculates the parallax angle factor, reflective area proportion factor, and sharpness factor of each frame, and selects the frame with the highest comprehensive score as the target image. The image analysis module 53 performs fine recognition on the target image and calculates a high-precision third liquid level value.

[0034] S600, Multi-dimensional Data Fusion and Closed-Loop Calibration. The data fusion verification module 54 receives the third liquid level value and uses it as the reference true value. The data fusion verification module 54 then reads the remote instrument value from the distributed control unit 6 and the first liquid level value from the fixed camera 2 at the same time. If the third liquid level value is consistent with the remote instrument value, the data fusion verification module 54 determines that the fixed camera 2 has an observation deviation, and calculates the parallax correction coefficient of the fixed camera 2 based on the difference between the third liquid level value and the first liquid level value. The data processing server 5 stores the parallax correction coefficient for compensation and correction of subsequent data collected by the fixed camera 2.

[0035] In the initialization phase of the collaborative monitoring and verification method, to achieve accurate mapping between physical and digital spaces, the system constructs a unified mathematical model foundation. By defining the cascading relationship between rigid body transformation and perspective projection, the Euclidean geometric space of the three-dimensional world is transformed into the projective geometric space of the two-dimensional image plane. Specifically, this includes the following steps: Constructing a multi-dimensional spatial reference coordinate system: The system establishes three basic coordinate systems to describe spatial geometric relationships. The system defines the geometric center of visual anchor point 4 as the origin of the world coordinate system (WCS). , The axis is perpendicular to the plane where the visual anchor point is located and points outward, forming a local inertial reference frame. The system defines the optical center of the image sensor of the mobile augmented reality terminal 3 as the origin of the camera coordinate system (CCS). , The axis points forward along the principal optical axis. The system defines the upper left corner of the image plane as the origin of the image pixel coordinate system (IPCS), with the u-axis and v-axis corresponding to the row and column directions of the image, respectively. Through the above definition, the system establishes a spatial reference from the physical environment to the sensor's viewpoint.

[0036] Establishing an imaging mapping relationship based on the pinhole model: The system, based on the principle of ideal pinhole imaging, quantifies the geometric process of light reflection from the object surface to the imaging plane as a linear combination of rigid body transformation and perspective projection. The system accurately describes points in three-dimensional space using the following projective transformation formula. With two-dimensional image pixels Mapping relationship between them: ; The parameters are defined as follows: The homogeneous coordinate vector of a point in space in the world coordinate system; This represents the homogeneous coordinate vector of the corresponding point in the image pixel coordinate system; The scale factor corresponds to the depth value of a point in space in the camera coordinate system. ; The extrinsic parameter matrix contains the rotation matrix describing the device's orientation. and the translation vector describing spatial displacement ; This is the camera intrinsic parameter matrix, including focal length. and principal point coordinates The matrix is ​​pre-obtained and stored by the system using the Zhang Zhengyou calibration method.

[0037] Real-time calculation of the pose extrinsic parameter matrix: During the monitoring process, the mobile augmented reality terminal 3 continuously acquires image data of the visual anchor point 4, extracts the two-dimensional coordinate set of its feature corner points in the pixel coordinate system, and retrieves the pre-stored physical dimensions of the visual anchor point and its corresponding three-dimensional coordinate set. Using the established 2D-3D point pair relationship, the system executes the PnP (Perspective-n-Point) pose solving algorithm to calculate the extrinsic parameter matrix at the current moment in real time. This matrix precisely quantifies the real-time position and line of sight of the inspection personnel relative to the graduated cylinder level gauge 1, providing a unique geometric transformation benchmark for subsequent parallax correction, virtual scale overlay, and spatial light projection in the augmented reality environment.

[0038] After constructing the world coordinate system based on visual anchor points, the system is used to establish a rigid geometric mapping between the physical entity of the graduated cylinder level gauge and the digital spatial coordinate system. This embodiment uses database indexing and spatial transformation matrices to dynamically map static equipment parameters to the current world coordinate system, constructing a virtual geometric twin containing axis vectors, pipe wall boundaries, and observation normals. Specifically, the following steps are included: Visual Anchor Point Decoding and Spatial Configuration Retrieval: The mobile augmented reality terminal captures visual anchor point images through its image acquisition sensor and uses a preset decoding algorithm (such as AprilTag or DataMatrix decoder) to parse and extract a unique device identifier (ID). The system uses this identifier as the primary key to retrieve the corresponding device spatial configuration file from a pre-built device attribute database. This configuration file contains key parameter fields defining the device's geometric attributes: relative pose matrix, and physical radius of the connecting tube. Effective range of the scale plate And the optimal observation distance coefficient.

[0039] Constructing the relative pose rigid body transformation matrix: The system reads the rigid transformation parameters from the configuration file and constructs a homogeneous transformation matrix describing the positional relationship between the local coordinate system of the graduated cylinder level gauge and the world coordinate system (the center of the visual anchor point). .

[0040] This matrix is ​​the mathematical foundation for aligning the digital twin of a device with its physical entity in space, and its definition is as follows: ; in: for The orthogonal rotation matrix quantifies the rotational deviation of the inherent axis of the level gauge relative to the world coordinate axis; for Translation vector, representing the three-dimensional coordinates of the local coordinate system origin of the level gauge (such as the center of the bottom flange) in the world coordinate system. The above parameters were obtained through calibration using high-precision measuring equipment during the equipment installation phase and are considered known constants when not under maintenance.

[0041] Generate the equations for the axis parameters of the level gauge: In order to support subsequent line-of-sight projection and intersection calculation, the system abstracts the central axis of the connecting pipe into a spatial straight line model in the world coordinate system.

[0042] The system first performs a rotation transformation on the direction vector, converting the normalized axis vector in the local coordinate system of the level gauge. Mapped to axis direction vectors in world coordinate system .

[0043] Simultaneously, extract the translation vector. As the reference point of the axis .

[0044] Based on this, the system constructs the spatial parametric equations of the central axis. Used to describe the three-dimensional coordinates of any position at any height on a straight line: ; in: Represents the world coordinates of any point on the axis; For scalar parameters along the axis, since As a unit vector, its physical meaning corresponds to the vertical height distance relative to a reference point, and its value range is constrained. In addition, the system combines the read physical radius , construct With the axis, A cylindrical geometric model with radius is used as the solution boundary for the subsequent disparity correction algorithm.

[0045] Determining the scale plane normal vector and optimal observation reference: In order to generate AR guidance graphics to indicate the optimal observation angle, the system needs to calculate the spatial orientation of the scale indicator plane and the normal vector in the local coordinate system. Transformed into an absolute normal vector in the world coordinate system using a rotation matrix. .

[0046] Based on this normal vector, the system further calculates the target liquid level height. Ideal observation point coordinates This coordinate system defines the optimal spatial position where the camera's optical axis is perpendicular to the scale plane and completely covers the observation area. Its calculation formula is as follows: ; in: This is the current liquid level reference value; The optimal observation distance scalar value is based on the camera's vertical field of view. and effective range Calculation yields ( This ensures that the height of the graduated cylinder image occupies a preset proportion of the image frame. Through the above steps, the system transforms simple label recognition into a complete spatial reconstruction of the geometric attributes of the field equipment, providing a physically constrained geometric reference benchmark for subsequent high-precision parallax correction.

[0047] After completing the spatial benchmark construction and device registration, the system performs a first-level coarse detection based on ray projection to quickly locate the liquid surface position in the two-dimensional image space, providing an observation benchmark for subsequent three-dimensional back projection. Specifically, this includes the following steps: Region of Interest (ROI) Adaptive Cropping: The system uses the established spatial mapping relationship to project the connected tube region in three-dimensional space onto the two-dimensional image plane to narrow the search range of image processing.

[0048] The system selects the axis of the graduated cylinder level gauge. The top point, bottom point, and extreme points of the cylindrical surface of the pipe wall in the world coordinate system constitute a set of three-dimensional key points. For each 3D point in the set, the system calculates its projection point on the image plane using the current pose parameters and projection model. : ; in: These are the homogeneous coordinates of the 3D key points; These are the homogeneous coordinates of the image projection points; The scale factor is used. After obtaining the set of projection points, the system calculates their extreme value ranges along the u-axis and v-axis of the image coordinate system, constructs an initial bounding rectangle, and applies a preset dilation coefficient (e.g., 1.2 times) or a fixed edge distance (padding) to generate the final region of interest. The system then uses the coordinates of this region to calculate the region of interest from the original image. Cropping out a local image matrix And scale it to the network input resolution.

[0049] Semantic feature extraction and probabilistic graphical reasoning: The system will extract semantic features from the preprocessed local images. The input is fed into a pre-built convolutional neural network model. This model employs an encoder-decoder architecture based on MobileNet or ShuffleNet backbone networks, extracting high-dimensional semantic features of the image through multi-layer convolutional operations.

[0050] The network output layer uses a Sigmoid activation function to generate a probability heatmap. Any element in the matrix The value represents the confidence probability that the pixel belongs to the liquid phase region (the liquid-filled part of the connecting tube). If the probability value is greater than a preset threshold (such as 0.5), it is determined to be a liquid phase candidate point.

[0051] Liquid surface edge response gradient calculation. To accurately locate the interface between the gas and liquid phases, the system performs vertical gradient analysis on the probability heatmap generated in the above steps. The system first applies Gaussian smoothing filtering to the heatmap to suppress noise, and then calculates the first-order difference operator response along the vertical axis (v-axis) of the image. The calculation formula is as follows: ; In the formula: This represents the gradient response value of the pixel at coordinates (u, v) in the vertical direction in the image. This represents the probability heatmap matrix of liquid level output by the neural network; Represents the horizontal coordinates (column index) of the image; Represents the vertical coordinates (row index) of the image; and They represent the same column. Above, the probability values ​​of the lower and upper neighboring pixels of the current pixel.

[0052] Physical meaning: This formula uses the central difference method to calculate the rate of change of gray level in the vertical direction. The numerical value quantifies the intensity of the abrupt change in the probability distribution in the vertical direction. The larger the value, the more significant the difference in probability values ​​on both sides of the location, indicating a higher probability that the location is a gas-liquid interface. By traversing the entire image to calculate this gradient response, the system can transform the blurred probability distribution into sharpened edge features.

[0053] Subpixel-level liquid level coordinate extraction: In order to overcome the limitation of image raster resolution, the system uses the weighted centroid method to calculate the subpixel position of the liquid surface.

[0054] The system first calculates the cumulative response values ​​of the gradient graph rows to determine the index of the row with the maximum response. The following definition includes Layer calculation window of its neighborhood Calculate the sub-pixel liquid level ordinate within this window. : ; in: The integer value representing the row coordinate of a pixel within the window; These are the gradient response weights at the corresponding positions; This represents the vertical coordinate offset of the ROI cropping region relative to the origin of the original image. Finally, the system uses the equation of the projected line of the connected tube axis to... Substitute the values ​​to solve for the corresponding x-coordinates. This allows us to obtain the precise two-dimensional coordinates of the center of the liquid surface in the original image coordinate system. .

[0055] After obtaining the two-dimensional coordinates of the liquid surface in the pixel coordinate system through image processing, the system executes a three-dimensional spatial inverse mapping algorithm. This process, based on the principle of optical path reversibility and camera calibration parameters, maps the two-dimensional observation data back to the three-dimensional world coordinate system. Specifically, it includes the following steps: Constructing the gaze vector in the camera coordinate system: The system utilizes the camera intrinsic parameter matrix to eliminate the perspective projection effect of the imaging lens. Let the homogeneous form of the center pixel coordinates of the liquid surface be... The system multiplies the inverse of the intrinsic parameter matrix by left multiplication. Calculate the non-normalized line-of-sight vector in the camera coordinate system. Subsequently, the system normalizes the vector to obtain a unit vector describing the direction of light ray flight in the camera's reference frame. .

[0056] Transform the line of sight to the world coordinate system: The system applies the principle of rigid body kinematics and uses an external parameter matrix to transform the light description in the local coordinate system of the camera to the world coordinate system, so as to establish a unified spatial reference with the equipment axis.

[0057] First, using rotation matrix Using the orthogonality property, we can solve for the absolute coordinates of the camera's optical center in the world coordinate system. Secondly, a rotation transformation is applied to the line-of-sight vector to obtain... .

[0058] Based on the above parameters, the system constructs a spatial ray equation describing the line-of-sight trajectory. : ; in: The distance scalar parameter along the ray ( ).

[0059] Solving for the parameters of the common perpendicular of skew lines: The reconstructed line-of-sight ray is affected by measurement noise and calibration residuals. With the central axis of the graduated cylinder In three-dimensional space, these lines are typically skew lines (i.e., they do not intersect). The system calculates the common perpendicular segment between the two skew lines and selects its foot on the equipment axis as the three-dimensional position estimate of the liquid level.

[0060] make Let be the connection vector pointing from the axis reference point to the camera center. The system calculates the liquid level height parameter based on the least squares principle using the following geometric projection formula. : ; Where: the symbol “·” represents the vector dot product operation; and These are the unit direction vectors of the line of sight and the axis, respectively. and The value is 1; the denominator term The square of the sine of the angle between the two lines ensures the uniqueness of the solution under non-parallel conditions.

[0061] Height constraint verification and output: The system verifies the calculated physical height. Perform boundary validity verification. The system retrieves the effective measurement range of the equipment. : If the result exceeds the physical base or top (i.e. or If the result is within the range, the system will forcibly correct it to the corresponding boundary value or trigger an exception message; if the result is within the range, it is considered valid. This height value It serves both as the output of the first-level detection result and as the initial iteration seed value for the subsequent second-level fine disparity correction algorithm.

[0062] After obtaining an approximate value of the liquid level height through a primary coarse detection, the system performs ideal observation pose calculation, aiming to solve for a specific six-degree-of-freedom (6DoF) pose such that the camera optical axis is perpendicular to the scale plane and the optical center is level with the liquid level. Specifically, this includes the following steps: Establishing a zero-parallax observation reference point: In order to eliminate the geometric parallax caused by the tilt of the viewing angle, the system determines the absolute position of the center of the liquid surface in the world coordinate system based on the principle of spatial linear interpolation, which serves as the geometric anchor point of the zero-parallax plane of the line of sight.

[0063] The system combines the axis reference point and normalized axis direction vector Using a rough liquid level height Calculate the target point of the line of sight : ; This formula describes the process of vector superposition in three-dimensional space. Physically, it means that starting from the physical reference point of the device, along the axis of the connecting pipe, a distance equivalent to the current liquid level height is moved, thereby accurately locating the three-dimensional coordinates of the center of the liquid surface in the world coordinate system.

[0064] Calculating the optimal observation distance: The system dynamically calculates the optimal observation distance based on the camera's imaging parameters. This ensures that the scale readings have a pixel density in the image that supports high-precision recognition.

[0065] The system sets the target imaging ratio coefficient. and the physical height of the area of ​​concern Combined with the camera's vertical field of view The optimal depth of field is solved using the following field-of-view constraint formula: ; This formula is based on the triangular geometry of the pinhole imaging model. Physically, it calculates the value when the object (height is...)... The projection on the image sensor occupies a preset proportion. At that time, the optical center of the camera needs to maintain a perpendicular distance from the object.

[0066] in: The preset physical height of the area of ​​interest (e.g., 100mm); The percentage of the target in the vertical direction of the image (e.g., 0.5). This refers to the camera's vertical field of view. It is a tangent trigonometric function.

[0067] Constructing ideal observation position coordinates: Based on the principle of orthogonal observation, the system uses spatial vector translation transformation to determine the optimal position of the camera's optical center.

[0068] The system reads the unit normal vector of the scale surface. , target point Translate by distance along the positive direction of the normal vector To obtain the ideal observation position : ; This formula defines the spatial geometric constraints of the ideal observation point. Physically, it ensures that the camera is located directly in front of the center of the liquid surface, and at a distance that meets the requirements for optimal image sharpness. From this position, the line of sight is parallel to the normal to the scale surface, geometrically eliminating horizontal and vertical parallax.

[0069] Constructing the ideal observation pose matrix: The system uses the look-at algorithm to construct an orthogonal rotation matrix, defining the three-axis orientation of the camera when it is in the ideal position.

[0070] The system first calculates the three orthogonal basis vectors of the camera coordinate system: Forward basis vector (optical axis direction): Meaning: The camera's principal optical axis should be aligned against the direction of the normal vector of the scale surface, i.e., looking directly at the scale surface; Right-hand basis vectors (horizontal axis): Meaning: The horizontal axis direction of the camera is determined by calculating the plane normal vector perpendicular to the axis and the optical axis using the cross product of vectors. Upward basis vectors (vertical axis): Meaning: By using the cross product operation again, the strict orthogonality of the coordinate system is ensured.

[0071] After normalization, the above vectors are used to construct the ideal rotation matrix of the system. : ; This matrix describes the rotational attitude of the rigid body. Physically, it defines the direction cosines of the camera's own coordinate axes (right, top, front) relative to the world coordinate axes under ideal conditions, which guides how the camera rotates to align with the target.

[0072] Generate spatial guidance deviation: The system calculates real-time pose based on the Lie group metric principle. with ideal position Spatial discrepancies between positions drive the UI to guide the user. This represents the Euclidean distance difference between the two coordinates. For attitude deviations, the system first calculates the relative rotation matrix. The angle error was then derived using trace calculation. : ; This formula is a conversion formula from rotation matrix to axis-angle representation. Physically, it quantifies the minimum rotation angle required between the current camera pose and the ideal pose. The smaller this value, the closer the user's shooting angle is to the orthogonal observation standard.

[0073] After calculating the ideal observation pose Then, the system executes a virtual guided interaction process, instructing the operator to start from the current position by overlaying virtual primitives locked to physical space coordinates on the mobile terminal screen. Move to the ideal pose. The specific implementation logic includes the following steps: Spatial construction and projection transformation of virtual guiding primitives: The system first constructs at the ideal observation position. A virtual 3D wireframe model is constructed, with its geometry corresponding to the physical dimensions of the region of interest (ROI), and its pose aligned to an ideal rotation matrix. .

[0074] Subsequently, based on the principles of the graphics rendering pipeline, the system maps virtual vertices defined in the world coordinate system to the imaging plane of the current camera. For any virtual vertex... The system calculates the screen projection coordinates based on the real-time updated camera pose parameters. : ; This formula describes the complete transformation process from three-dimensional world coordinates to two-dimensional screen pixel coordinates. Physically, it projects the spatial position of the virtual object relative to the current camera onto the image sensor plane, ensuring that the virtual object's display position on the screen undergoes correct perspective changes as the camera moves, thereby achieving a visual lock effect.

[0075] in: Let $\mathbf{ ... This is the transpose of the current camera rotation matrix, used to transform vectors from the world coordinate system to the current camera coordinate system; This is the camera intrinsic parameter matrix; Depth scale factor; Homogeneous coordinates of the projected pixels .

[0076] Field-outside guidance indicator generation: when the ideal observation position is... When the target is outside the current camera's field of view, the system initiates out-of-field guidance logic. The system calculates the relative relationship between the current line of sight and the target direction, mapping the target's orientation to a pointing vector on the two-dimensional screen plane. : ; This formula achieves dimensionality reduction mapping of three-dimensional direction vectors. Physically, it calculates the azimuth angle of the target point relative to the camera's current orientation and projects it onto the screen plane. The resulting vector direction is the direction in which the user should rotate or move the device.

[0077] in: This represents the displacement difference vector from the current position to the ideal position; This is a dimension reduction projection function, which extracts the x and y components of the transformed vector while ignoring the depth component.

[0078] Multi-level visualization feedback within the field of view: When the ideal position enters the field of view, the system switches to fine guidance mode. The system calculates position deviation in real time. With attitude deviation And adopt a tiered rendering strategy: Roughly approaching the stage (e.g.) The rendering color is a warning color (such as red), the guide box is displayed as a dashed line, and the spatial guide vector path is displayed; Fine-tuning stage (e.g.) ): The rendered color gradually changes to a transitional color (such as yellow), the guide box becomes a solid line, and the alignment mark is displayed in the center of the screen; Locking the ready phase (e.g.) and ): The rendering color abruptly changes to the confirmation color (such as green), the guide box is highlighted and triggers haptic feedback (such as vibration) on the terminal. At this time, the virtual guide box and the physical scale surface are visually geometrically superimposed.

[0079] Dynamic transparency blending control: To prevent virtual primitives from obscuring the actual liquid level scale and affecting the reading, the system establishes a mapping function between positional deviation and rendering transparency (Alpha channel). Transparency function The definition is as follows: ; The formula defines a piecewise linear function to control the visibility density of the virtual layer. Physically, this achieves a gradient effect: when the user is far from the target, the guide box is clearly visible (high opacity); as the distance approaches, the guide box gradually fades, eventually becoming semi-transparent, ensuring that the camera can clearly capture the physical scale details behind the virtual layer.

[0080] in: The distance threshold for initiating transparency decay (recommended 80mm to 120mm). and These represent the maximum and minimum opacity, respectively.

[0081] Stability assessment and automatic triggering: A sliding window for system maintenance is used to monitor pose stability. When the deviation... and In continuous When all frames (e.g., 10 to 30 frames) remain within the locked threshold range, the system determines that the orthogonal observation condition has been met, automatically locks the camera's focus and exposure parameters, and triggers high-resolution image acquisition, entering the secondary fine detection process.

[0082] During the virtual guided interaction, the system performs quantitative parallax analysis on the observation perspective of each frame of image. Based on spatial geometric relationships, it calculates in real time the angular deviation between the current observation line of sight and the ideal orthogonal observation line of sight, and decomposes this deviation into vertical parallax components and horizontal parallax components. The specific calculation and judgment process includes the following steps: Constructing the observation line-of-sight space vector: The system acquires the camera optical center coordinates output by the SLAM module in real time. Combined with the target point of the liquid level center Construct and normalize the observation line-of-sight vector : ; This formula uses the difference in coordinates between two points in space to define the three-dimensional line of sight direction, and eliminates the influence of distance factors through norm normalization to obtain a unit vector that only represents the direction, providing a standardized geometric reference for subsequent angle calculations.

[0083] Solving for the vertical parallax angle: Vertical parallax is the main source of reading error caused by the displacement of the meniscus projection inside the cylindrical glass tube. The system calculates the unit line-of-sight vector. With the normalized liquid level gauge axis direction vector Vertical deviation angle between : ; This formula calculates orthogonality deviation using the vector dot product and trigonometric functions. Physically, it measures the angle by which the line of sight deviates from the ideal horizontal observation plane (i.e., the plane perpendicular to the axis). When the line of sight is perpendicular to the axis, the dot product is 0. A value of 0 degrees indicates no vertical parallax, which eliminates the projection error of the meniscus.

[0084] Solving for horizontal parallax angle: Horizontal parallax refers to the degree to which the observation point deviates from its facing position in the horizontal direction. The system calculates the angle between the projection of the line-of-sight vector onto the horizontal plane and the normal vector of the scale surface.

[0085] The system first uses the Gram-Schmidt orthogonalization principle to subtract the component along the axis from the original view vector, thus constructing a projection vector perpendicular to the axis. : ; Subsequently, the projection vector and the ideal observation direction are calculated. The horizontal parallax angle between (i.e., the reverse of the normal vector of the scale surface) : ; The first step involves spatial vector decomposition to separate the projection component, which contains only horizontal orientation information. The second step calculates the angle between this projection component and the ideal facing direction. Physically, this precisely quantifies the degree of horizontal deviation of the camera relative to the front of the scale panel, avoiding perspective distortion of characters during OCR recognition due to excessively large horizontal angles.

[0086] Integrated disparity determination and threshold logic: The system sets a disparity tolerance threshold and determines whether the current frame meets the acquisition conditions based on a weighted evaluation of the vertical and horizontal components. Determination logic. Defined as: This formula is a logical AND condition, meaning that the system considers the current acquisition conditions valid only when both the vertical parallax and the horizontal parallax simultaneously meet their respective accuracy thresholds.

[0087] in: The vertical parallax threshold (preferably 0.5° to 1.5°) is determined based on measurement accuracy requirements. ,in Where is the radius of the glass tube. Allow for reading errors (e.g., 0.5 mm); The horizontal parallax threshold (preferably 5° to 10°) is set to improve ease of operation while ensuring OCR recognition rate.

[0088] Real-time status feedback output: The system updates the output status in real time based on the above calculation results. 1. If If true, the system generates an "orthogonal locking" signal, activating the image acquisition process; 2. If If false, the system generates an adjustment command based on the direction of the deviation. The vertical offset is determined using the sign of the dot product: if... . If the angle between the line of sight and the axis is less than 90 degrees (the camera is too low), a "move up" command is output; otherwise, a "move down" command is output. This mechanism eliminates parallax based on imaging principles, ensuring geometric consistency of observations at different distances.

[0089] During image acquisition, the system does not trigger the shutter at a single moment, but instead defines a "low parallax tolerance range" and continuously buffers images at a high frame rate within this range. The specific implementation logic includes the following steps: Mathematical definition of low disparity tolerance: The system defines a multidimensional sampling activation region in the pose state space, which is the low disparity tolerance. This range is based on the error tolerance setting of the image post-processing algorithm, ensuring that the geometric distortion of the acquired image is within the non-destructive repair threshold.

[0090] The system defines the low parallax tolerance range as the set of states that simultaneously satisfy both vertical and horizontal parallax constraints: ; This formula describes the set of state spaces constrained by parallax angles. Physically, it defines the geometric boundary conditions under which the system "allows image buffering," meaning that the system considers the current observation data valid and initiates buffering only when the camera's pose simultaneously satisfies specific angular constraints in both the vertical and horizontal directions.

[0091] in: Indicates time The system pose state at any given moment; and These are the vertical and horizontal parallax angles calculated in real time; A horizontal sampling boundary threshold (ranging from 8° to 12°) is set, which is wider than the final locking threshold, and the number of samples is increased to leverage the robustness of the OCR algorithm against trapezoidal distortion. This is the vertical sampling boundary function that dynamically changes with the observation distance.

[0092] Dynamic threshold adjustment based on observation distance: Due to the distance between the camera and the level gauge The absolute projection error caused by parallax is directly affected, and the system uses a dynamic function to adjust the vertical sampling boundary in real time. Based on the pinhole imaging principle, a stricter angle threshold is required for close-range observation to control pixel-level displacement.

[0093] Vertical sampling boundary function The structure is as follows: ; This formula establishes a positive correlation between observation distance and allowable parallax angle using a power function. Physically, it implements an adaptive constraint strategy of "stricter at close range, wider at distant range": the closer the observation distance, the smaller the allowable tilt angle, in order to offset the projection error caused by the magnification effect at close range and ensure the consistency of imaging projection error at different distances.

[0094] in: The baseline parallax tolerance angle (e.g., 2.0°); The distance from the optical center of the camera to the center of the level gauge in the current frame; To preset the optimal observation distance (e.g., 400mm); The distance sensitivity coefficient (0.8 to 1.2) is used to adjust the sensitivity of the threshold as distance changes.

[0095] The hysteresis comparator-based interval triggering logic: To avoid frequent switching (oscillation) of the sampling state at the threshold due to sensor noise or minor jitter, the system introduces hysteresis comparator logic to control the start and stop of the sampling state machine. The system sets an entry threshold. and exit threshold ( ), sampling activation signal The judgment logic is as follows: ; This formula describes a state machine with memory properties. Physically, this is achieved by setting an asymmetric threshold (hysteresis margin). (e.g., 0.3° to 0.5°), which enhances the system's resistance to interference in critical states. Even with slight hand tremors, as long as the sampling threshold is not exceeded, the system maintains the sampling state, ensuring the continuity of the sampling sequence.

[0096] in: Corresponding to dynamic boundary .

[0097] Variable frequency sampling density control: During sampling activation, the system employs a variable frequency sampling strategy based on pose change rate, dynamically allocating computational resources according to the current pose quality. Sampling frequency The calculation is as follows: ; This formula defines an exponential decay relationship between the sampling frequency and the absolute value of vertical parallax. Physically, it achieves adaptive resource scheduling: when the user is extremely close to a perfectly orthogonal position (parallax approaches 0), the system acquires samples at full speed to obtain more high-quality samples; when at the edge of the permissible range, it automatically reduces the frequency to reduce redundant data processing and power consumption.

[0098] in: and These are the maximum and minimum sustained frame rates, respectively. The frequency attenuation coefficient (0.2 to 0.5) controls the rate at which the frequency decreases as parallax increases.

[0099] To select the most suitable frames for reading recognition from the image sequence acquired within the dynamic sampling interval, the system is equipped with a multi-factor weighted evaluation unit. This unit establishes an evaluation function based on three dimensions: geometric orthogonality, optical interference, and image sharpness. Cache images for each frame Real-time scoring is performed. The specific evaluation logic includes the following steps: Constructing a comprehensive quality evaluation function: The system defines each frame of the image. Overall quality score This step, based on the principle of multi-objective decision-making, maps image feature parameters of different physical dimensions into dimensionless normalized values ​​to facilitate unified sorting.

[0100] The comprehensive evaluation formula is as follows: ; The formula uses a linear weighted summation method to calculate the total score. Physically, it comprehensively considers three key factors affecting reading accuracy and clarifies the optimal strategy through weight allocation: first, ensuring the correct geometric perspective; second, avoiding glare interference; and finally, ensuring image clarity.

[0101] in: The parallax factor weighting coefficient (preferably 0.5 to 0.6) is used because vertical parallax directly determines the liquid level height. Projection error is an irreversible geometric systematic error, so it is set as the dominant term. A reflectivity factor weighting coefficient (preferably 0.2 to 0.3) is used to reduce the priority of images severely affected by ambient light interference; The sharpness factor weighting coefficients (preferably 0.1 to 0.2) are used to assist in removing motion-blurred frames; the three weighting coefficients must satisfy normalization constraints: .

[0102] Calculate the disparity rating factor: disparity factor This step is used to quantify the quality of the geometric pose at the current frame acquisition time. It utilizes the decay property of a Gaussian function to convert the angle error into a score between 0 and 1; the larger the error, the faster the score decreases.

[0103] The calculation formula is as follows: ; This formula indicates that the disparity score has a negative exponential relationship with the square of the disparity angle. Physically, it simulates the probability density characteristics of a normal distribution, mapping angular deviation to confidence scores and achieving a non-linear quality penalty.

[0104] in: and For the first The vertical and horizontal parallax angles corresponding to the frame image; The standard deviation of vertical parallax attenuation is set to 0.5° to 1.0°. Smaller values ​​control the scoring curve to be steeper, reflecting the strict requirements for vertical orthogonality. The standard deviation of horizontal parallax attenuation is set to 0.5° to 8.0°, indicating that the system has a wide tolerance range for horizontal parallax.

[0105] Calculating the reflection interference factor: To address the issue of specular reflection in the glass tube of the level gauge causing obstruction of the scale lines, the system calculates the reflection factor by analyzing the brightness histogram within the region of interest (ROI). This step is based on statistical photometry principles, assessing the degree of illumination interference by detecting the proportion of bright, saturated pixels.

[0106] The calculation formula is as follows: ; This formula defines an inverse proportional deduction mechanism with a cutoff limit. Physically, it quantifies the "effective visible area" of the image: the larger the reflective area, the smaller the effective reading area, and the lower the score.

[0107] in: This represents the total number of pixels within the ROI region. The brightness value exceeds the saturation threshold. The number of pixels (e.g., 245 to 250); This is the reflectivity coefficient (preferably 5 to 10), used to amplify the influence of the reflectivity ratio. When the reflective area ratio reaches... When the percentage is between 10% and 20%, the deduction item saturates to 1, making... The value is reduced to 0, thus forcibly discarding images with obvious glare.

[0108] Calculating the sharpness factor: In order to eliminate blurry images caused by hand shakiness or focus failure, the system uses the Laplacian gradient operator to calculate the high-frequency component energy of the image as the basis for sharpness evaluation.

[0109] Sharpness factor The calculation formula is as follows: ; This formula uses the variance of the image gradient to characterize the richness of image detail and performs dynamic normalization based on the statistical characteristics of the current sampling queue. Physically, instead of calculating absolute sharpness, it compares the "relative sharpness" of the current frame within the entire acquisition sequence to select the sharpest image in the batch.

[0110] in: This indicates that a Laplacian second-order differential convolution operation is performed on the ROI region of the image; This indicates the calculation of the statistical variance. and Each of the current sampling queues contains all The minimum and maximum values ​​of the gradient variance corresponding to the frame image; For a minimal constant (e.g.) (), used to prevent the denominator from being zero.

[0111] Frame selection and output strategy: After each dynamic sampling period, the system optimizes the frame selection and output strategy for the buffer queue. Frame image calculation above The system selects the value and performs a sorting operation. The highest-scoring frame is used as the primary reference frame for subsequent processing, while the second-highest-scoring frame is retained. Frames (e.g.) This strategy constructs a redundant data pool based on quality grading. When the confidence level of the main reference frame is insufficient in the subsequent OCR recognition stage, the system can directly call the alternative frame for secondary recognition, avoiding the need for the user to re-acquire the image.

[0112] After completing the scoring based on multidimensional quality factors, the system does not directly output the raw data of the highest-scoring image, but instead enters the fine recognition stage. This stage is used to correct the image using known geometric prior information and establish a multi-level verification mechanism to ensure the accuracy of the final reading.

[0113] The specific processing procedure includes the following steps: Candidate frame queue construction based on quality ranking: The system selects all candidates from the dynamic buffer queue. Frame images according to overall quality score Sort the images in descending order. This step is based on the principle of redundancy backup, and constructs a candidate pool containing multiple high-quality frames to deal with possible local random noise in a single frame image (such as highlight points that happen to block digits).

[0114] Candidate Frame Set The definition is as follows: ; The formula defines a set of filtering criteria to ensure that only "high-quality" and "high-ranked" images can enter the subsequent recognition process.

[0115] in: Indicates the sorted number of... Frame image, in which The frame with the highest score; This is the overall quality score for the image frame. A quality truncation threshold (e.g., set to 0.6) is used to remove invalid frames that have too low a score and clearly do not meet the recognition requirements. The maximum candidate capacity (e.g., set to 3) indicates that the system will cache and attempt to identify the best 3 frames at most; This represents a logical AND relationship, meaning that an image entering the candidate set must simultaneously satisfy both of the above conditions.

[0116] Perspective distortion correction based on pose prior: for selected candidate frames (Initially) Although its parallax is within acceptable limits, slight trapezoidal distortion may still exist. The system utilizes the real-time camera pose information obtained during the pose calculation step to construct a homography matrix. This transforms an image projected from an oblique perspective into an ideal orthographic view.

[0117] homography matrix The calculation formula is as follows: ; This formula, based on the principle of plane-induced homography in multi-view geometry, describes the geometric mapping relationship from the current camera view to the orthogonal plane view. Physically, it guides the system to perform an "inverse perspective transformation" on the image: correcting the slanted scale lines in the image to be horizontal and equally spaced, and correcting deformed digits to be vertical, thereby fundamentally eliminating the geometric distortion caused by perspective projection and providing standardized input data for OCR recognition.

[0118] in: This is the camera's intrinsic parameter matrix (3×3), which includes parameters such as focal length and principal point coordinates; This is the rotation matrix (3×3) of the current frame relative to the world coordinate system. The translation vector (3×1) of the current frame relative to the world coordinate system. The unit normal vector (3×1) of the level gauge scale panel in the world coordinate system is typically defined as follows under standard orthogonal alignment: ; The vertical distance from the optical center of the camera to the plane of the scale panel; This represents the matrix transpose operation.

[0119] Enhancement and binarization of gas-liquid interface features: Targeting the unique gas-liquid interface (meniscus) features of glass tube level gauges, the system performs image correction before recognition. Image enhancement is performed. The system uses an adaptive contrast enhancement algorithm (CLAHE) to improve the local contrast of the region of interest (ROI), and then uses Otsu's method (maximum inter-class variance method) for adaptive threshold binarization.

[0120] Binarization threshold The determination of the value needs to satisfy the maximization of inter-class variance. : ; This formula is used to find an optimal segmentation threshold that maximizes the statistical difference in grayscale between the foreground (characters / scales) and the background (glass tube wall) segmented at that threshold.

[0121] in: The threshold value for the grayscale level to be traversed (0-255); At the threshold respectively The percentage (probability) of pixels that are identified as background and foreground pixels out of the total number of pixels. These are the average gray values ​​of the background and foreground regions, respectively. The term represents the square of the difference between the two means; the larger the term, the better the separation between the foreground and background. This step converts a continuous grayscale image into a binary image containing only black and white information, effectively separating the scale characters from the background and filtering out color interference from the liquid.

[0122] Character recognition based on sequence features (OCR): The system inputs the enhanced binary image into a pre-trained convolutional recurrent neural network (CRNN) model for reading recognition. The model structure includes convolutional layers (CNN) for extracting the morphological features of characters and recurrent layers (RNN / LSTM) for inferring sequence relationships based on context, making it particularly suitable for handling broken or stuck numbers on level gauges.

[0123] The recognition result is output as a numerical result. and confidence level Data tuples: ; in This represents the probability assessment value (range 0 to 1) of the model for the recognition result.

[0124] Multi-level verification and rollback mechanism: To ensure the reliability of output data, the system performs the following logical verifications: Confidence check: If (If a pre-set confidence threshold is set, e.g., 0.85), the recognition result of the current frame is determined to be unreliable, and the system automatically discards the result; Range verification: If the recognized value... Exceeding the preset physical measurement range of the level gauge This also triggers a rollback mechanism; rollback execution: Once any of the above rollback conditions are triggered, the system will move from the candidate set... The next frame is fetched from the middle (i.e.) (The second highest frame) Repeat steps S760 to S780. If all frames have been traversed... If a candidate image fails the verification, the system will send a "Recognition failed, please adjust the angle and try again" message to the user interface instead of outputting an incorrect reading. This backoff strategy based on multi-frame selection effectively solves the recognition errors caused by occasional blurring, reflection, or partial occlusion in traditional single-frame recognition. By utilizing redundant information in the time series, it improves the robustness of the system in complex industrial environments.

[0125] To address potential monitoring gaps or device failures associated with single data sources, the system employs a multi-source heterogeneous data verification model. This model utilizes high-precision AR readings acquired via mobile devices. Based on this, combined with continuous monitoring readings collected by fixed cameras on site and sensor electronic readings acquired by the DCS (Distributed Control System). It performs real-time consistency analysis and status determination. The specific processing includes the following steps: Time alignment and acquisition of heterogeneous data: The system receives data streams from three different channels. Since the sampling frequencies of each data source are different (e.g., DCS uses continuous sampling at the second level, while AR acquisition uses discrete sampling triggered by humans), and the liquid level may be in dynamic fluctuation, the comparison unit first performs a timestamp-based alignment operation.

[0126] This step is based on the principle of nearest neighbor retrieval within a time window, aiming to eliminate dynamic errors caused by asynchronous sampling times.

[0127] The system uses the AR data acquisition time Based on this, time windows are retrieved from the DCS historical database and the fixed camera cache queue. Get the nearest neighbor data record within the range. and .

[0128] for The numerical registers of level sensors (such as differential pressure transmitters or magnetostrictive level gauges) can be directly read via OPCUA or Modbus industrial protocols; for It calls a lightweight OCR algorithm deployed on edge computing nodes to recognize and obtain numerical values ​​from images from fixed camera positions.

[0129] Calculate the relative deviation index: The system uses high-confidence AR readings validated by previous steps. As a temporary ground truth, the relative deviation rates of the fixed camera and the DCS sensor are calculated separately. The AR reading is chosen as the benchmark here because it has been verified by "human-machine loopback" and has higher logical reliability than unattended sensors.

[0130] The formula for calculating deviation is defined as follows: ; This formula is used to calculate the percentage error of the observed value relative to full scale. Physically, the range normalization calculation eliminates the numerical magnitude differences between different specifications of level gauges, making the deviation index universal and comparable across different devices.

[0131] in: The relative deviation rate of the electronic sensor readings; The relative deviation rate of fixed camera readings; The physical range span of the level gauge (i.e. ); This represents absolute value operations.

[0132] Three-way consistency logic determination: The system uses a comparison logic matrix to classify and determine the above deviations in order to identify the equipment status. The system sets two verification thresholds: a warning threshold and a warning threshold. (e.g., 1%) and fault threshold (e.g., 5%).

[0133] The comparison logic is as follows: Sensor zero drift determination: if and In other words, the AR reading is consistent with the fixed camera reading, but both are significantly different from the DCS reading. The system determines that the DCS sensor has "zero drift" or "nonlinear error".

[0134] Visual recognition anomaly determination: If and This means that the AR reading is consistent with the DCS sensor reading, but the fixed camera reading is abnormal. The system determines that the fixed camera may have a dirty lens, be displaced, or have light obstruction.

[0135] Judgment of errors in manual data collection: If and ,and This means that the DCS is consistent with the fixed camera, but the AR readings do not match either of them. The system determines that the current mobile terminal may be pointing at the wrong instrument or there may be human reading interference.

[0136] S830, alarm generation and closed-loop feedback strategy: Based on the judgment result of S820, the system generates hierarchical alarm signals and triggers corresponding feedback control commands.

[0137] To address the issue of DCS sensor deviations, the system defines a comprehensive alarm index. The calculation formula is as follows: ; This formula constructs a hybrid evaluation model containing discrete step terms and continuous linear terms to quantify the severity of anomalies. Physically, it not only classifies fault levels (warning / fault) through a step function, but also introduces the continuous influence of deviation amplitude through linear terms. That is, under the same fault level, the larger the deviation value, the higher the final alarm index, reflecting subtle differences in fault severity.

[0138] in: This is an indicative function. The function value is 1 when the condition in parentheses is met, and 0 otherwise. It is used to implement step triggering for graded alarms. Basic weighting coefficient (preferably, set) ), which correspond to the basic penalty points for warning level and fault level, respectively; This is the bias gain coefficient (e.g., 0.5).

[0139] The system according to Implement the following strategy: When When the calibration trigger threshold is reached, the system automatically generates a "recommended calibration work order" and sets the current deviation value. Write the suggestion into the compensation register area of ​​the DCS system for instrument maintenance personnel to confirm before performing linear compensation operations.

[0140] When a data collection error is detected, the mobile app immediately displays a "data conflict warning," forcing the operator to scan the form's QR code again and perform a second data collection to prevent incorrect data from being entered into the system.

[0141] To overcome long-term measurement drift caused by minor deformations of the mounting bracket, thermal expansion and contraction, or initial calibration errors in fixed CCTV cameras, the system is equipped with a parameter adaptive correction module. This module uses high-precision AR measurement results as "truth feedback" to update the parallax correction coefficient of the fixed camera, achieving closed-loop optimization of monitoring accuracy. The specific processing includes the following steps: Validation of Corrected Data: Before performing parameter correction, the system must ensure that the reliability of the AR data used for calibration is significantly higher than the current state of the fixed camera. The correction module first executes a set of verification procedures: High confidence constraint: Confidence level of AR recognition results It must be greater than the correction threshold. (For example, 0.95, this threshold setting is usually higher than the usual user-displayed threshold to ensure the reliability of the benchmark data); Steady-state environment constraints: Verify the rate of change of DCS sensor data at the current moment. This ensures that the liquid level is in a relatively static state, preventing time alignment errors caused by dynamic fluctuations in the liquid level from being incorrectly included in the system deviation. Non-outlier constraint: Confirm that the deviation between the AR reading and the fixed camera reading is within the "correctable range", i.e. .in The maximum permissible calibration deviation is set to 10% of the range. If the deviation exceeds this value, it indicates a possible hardware-level fault or severe displacement. In this case, the system will prohibit parameter correction and instead trigger an alarm signal.

[0142] Calculating the disparity correction coefficient observations: Visual measurement models using fixed cameras typically include a disparity compensation factor. It is used to map the pixel height of an image from a tilted viewpoint to the vertical physical height.

[0143] Its basic measurement formula is: This formula describes the linear mapping relationship from the image pixel domain to the physical numerical domain.

[0144] In this step, when the detected liquid level position deviates from the zero-point scale line by more than a preset safe pixel distance (to avoid the singularity of zero in the denominator), the system uses the high-precision liquid level value (i.e., as the reference true value) to deduce the ideal parallax compensation factor observation value at the current moment. : ; This formula essentially uses known physical true values ​​(high-precision liquid level values) and current image feature observations to solve for the gain parameters required in the system's transfer function. It dynamically calculates the "perfect parallax factor" needed to ensure that the current camera reading matches the high-precision reading from the mobile device.

[0145] in: The vertical coordinates of the detected liquid surface in the image coordinate system (unit: pixels); The vertical coordinates (in pixels) of the preset zero-point scale line in the image coordinate system. A pixel-to-physical-length calibration scale (unit: millimeters / pixel); This is the current high-precision liquid level value.

[0146] Weighted smoothing update of parameters: To prevent random errors from a single measurement from causing drastic fluctuations in system parameters, the system does not directly use... Instead of replacing the original coefficients, the exponentially weighted moving average (EWMA) algorithm is used to smoothly update the parameters.

[0147] The updated formula is as follows: ; This formula utilizes a weighted fusion of historical memory and current observations to construct a digital low-pass filter, achieving gradual parameter convergence. Physically, this ensures that parameter correction is a smooth process, avoiding abrupt changes in monitoring parameters due to single AR recognition errors (such as occasional noise).

[0148] in: The updated parallax correction coefficient for the current moment; The coefficient is the one from the previous moment (or the one currently in effect); To update the gain factor (preferably set to 0.05 to 0.1). Smaller The value implies that the system gives higher weight to historical experience and only responds slowly to systematic deviation trends.

[0149] After the update is complete, the system will have a new... The data is written into the non-volatile memory of the fixed camera and takes effect immediately in subsequent acquisition cycles. This closed-loop calibration mechanism based on heterogeneous data feedback enables the fixed camera to adaptively compensate for system errors caused by loose installation or environmental changes, eliminating the need for frequent manual on-site recalibration.

[0150] Specific application examples: This embodiment constructs a complete intelligent identification and analysis system for graduated cylinder liquid levels. The system hardware includes: Mobile terminal (AR inspection terminal): A smart handheld device equipped with a high-precision IMU (inertial measurement unit) and a 12-megapixel camera is used to run the AR guidance and image acquisition APP of this invention.

[0151] Fixed monitoring terminal: A fixed industrial camera (5 megapixels) is deployed at the location of the measuring cylinder in the laboratory for long-term fixed-point monitoring.

[0152] Cloud server: Deploys image processing algorithms, multidimensional verification modules, and databases.

[0153] To verify the technical effectiveness of this invention in terms of parallax correction, anti-interference recognition, and long-term operational stability, this embodiment conducted the following three sets of comparative experiments and recorded the relevant data (see appendix). Figure 3 To be continued Figure 5 ).

[0154] Verification of the impact of observation angle deviation on measurement accuracy: In actual industrial inspections, the angle at which inspectors take photos while holding the camera is quite arbitrary. To verify the effectiveness of the "dynamic orthogonal viewpoint guidance" in step S400 of this system, we set up observation points at different vertical angles (0° to 40°) in front of the measuring cylinder and conducted liquid level reading tests using both "existing technology (direct shooting without guidance)" and "this invention (AR guidance + spatial inverse calculation)".

[0155] See attached document Figure 3 , Figure 3 This demonstrates the beneficial effects of the system in handling spatial parallax.

[0156] Coordinate definition: The horizontal axis represents the vertical angle deviation between the optical axis of the imaging device and the horizontal line of the liquid surface in the graduated cylinder (unit: degree), with 0 degrees representing the ideal level view; the vertical axis represents the absolute error between the final liquid level reading output by the system and the true value (unit: millimeter).

[0157] Effect Analysis: The dashed line in the graph (labeled data1) represents the test results of "existing technology (unguided)". The data shows that the data1 curve increases exponentially with increasing observation angle deviation. For example, when the deviation reaches 30°, the reading error displayed by data1 exceeds 15mm, indicating that ordinary shooting methods are severely affected by differences in glass refractive index and perspective projection distortion, and cannot meet the requirements for precision measurement.

[0158] The solid line in the figure (labeled data2) represents the test results of "This Invention (AR Guidance)". This system calculates the parallax angle in real time using the mobile terminal's attitude sensor and generates a virtual AR arrow on the screen to guide the user to adjust the angle to within 5°. Simultaneously, the algorithm automatically applies a spatial inverse matrix to geometrically compensate for any remaining small parallax. Experiments show that the data2 curve remains consistently stable; even with a large deviation in the user's initial shooting angle, the system can stably control the final measurement error within 1.5mm.

[0159] Conclusion: By comparing data1 and data2, it can be seen that the present invention has significantly solved the problem of uncontrollable accuracy caused by the high randomness of handheld shooting, and enabled non-professionals to take laboratory-level accurate readings.

[0160] Verification of image optimization process under complex lighting conditions: The industrial environment is complex, and the glass surface of the measuring cylinder is easily affected by reflections from lights, sunlight, etc. To verify the effectiveness of the "image quality evaluation function" in the S500 and S700 steps of the system, we simulated a continuous acquisition process, during which strong light source interference was artificially introduced.

[0161] See attached document Figure 4 , Figure 4 This demonstrates the process by which the system automatically selects the best recognition frame under complex lighting conditions.

[0162] Coordinate definition: The horizontal axis represents the time series frame number of the video stream acquisition (t=1 to 25 frames); the left vertical axis corresponds to the overall image quality score. The right vertical axis corresponds to the intensity of environmental reflective interference.

[0163] Effect Analysis: The bar chart (labeled data2) represents the environmental reflection interference factor. The data shows that during the data acquisition process (especially in the 10-15 frame interval), the data2 value increases significantly due to strong reflections on the graduated cylinder surface, indicating a serious risk of feature masking at this point.

[0164] The solid curve in the figure (labeled as data1) represents the overall image quality score. The score integrates sharpness, parallax angle, and reflectivity. Data shows a clear negative correlation between data1 and data2: when reflectivity (data2) increases, the quality score (data1) drops rapidly. The algorithm of this invention automatically avoids low-scoring areas by calculating the data1 value of each frame in real time, and locks the highest point of the score curve (marked with an asterisk in the figure, approximately frame 19) as the "optimal recognition frame".

[0165] Conclusion: The dynamic relationship between data1 and data2 proves that the system can effectively remove poor-quality images caused by reflection and jitter, ensuring that the subsequent OCR recognition algorithm only processes the highest quality data source, thus greatly improving the recognition success rate.

[0166] Long-term stability verification of mobile and fixed-end collaboration: To address the long-term drift issue of fixed cameras caused by slight deformation and vibration of the bracket, this embodiment simulates a 60-day long-term operation test.

[0167] See attached document Figure 5 , Figure 5 The closed-loop calibration effect achieved by working collaboratively between mobile and fixed terminals is demonstrated.

[0168] Coordinate definition: The horizontal axis represents the number of days the system has been running continuously in the industrial field; the vertical axis represents the measurement error of the fixed camera (unit: millimeters). The two thin dashed lines (y=±5mm) in the figure represent the upper limit of the error allowed by industrial standards.

[0169] Effect Analysis: The dashed line in the graph (labeled data1) represents the error evolution of the fixed camera when closed-loop calibration is not enabled. The data shows that, affected by environmental thermal expansion and contraction and minor mechanical vibrations, data1 exhibits a monotonic cumulative drift trend. Around day 35, the error exceeds the upper limit of +5mm, rendering subsequent data unreliable.

[0170] The solid line in the diagram (labeled data2) represents the error trajectory after enabling the "collaborative closed-loop calibration" of this invention. Whenever an inspection personnel use a high-precision AR device to conduct a routine inspection of this area (triggered approximately every 7 days), the system automatically uses the precise spatial coordinates obtained from the mobile device to calculate and update the homography matrix of the fixed camera, performing a "zeroing" calibration. The trajectory of data2 clearly demonstrates this periodic error elimination effect.

[0171] Conclusion: The comparison of the results of data1 and data2 proves that the present invention can ensure the continuous high-precision operation (high reliability) of fixed monitoring equipment for several months by utilizing low-frequency manual inspection (low cost), and significantly reduce the maintenance cost of manual professional calibration of fixed equipment.

[0172] In summary, this embodiment fully verifies the significant progress and beneficial effects of the present invention in the three core technical aspects of spatial parallax correction, anti-interference image optimization, and multi-terminal collaborative closed-loop calibration through three sets of quantitative experiments.

Claims

1. A computer vision-based intelligent identification and analysis system for graduated cylinder liquid levels, characterized in that, It includes a graduated cylinder level gauge (1), a fixed camera (2), a mobile augmented reality terminal (3), a visual anchor point (4), a data processing server (5), and a distributed control unit (6). The fixed camera (2) is used to collect the video stream of the graduated cylinder level gauge (1); the visual anchor point (4) is fixed to the outside of the graduated cylinder level gauge (1) as the absolute coordinate reference of the local space on site. The mobile augmented reality terminal (3) is used to acquire environmental images containing visual anchor points (4) to establish a local spatial coordinate system and to perform guided observation in response to instructions; The data processing server (5) is used to perform the following operations: Based on the visual anchor point (4), the spatial pose of the mobile augmented reality terminal (3) relative to the graduated cylinder level gauge (1) is calculated, the ideal observation position coordinates to eliminate parallax are calculated, and a guidance command is sent to the mobile augmented reality terminal (3) to instruct the mobile augmented reality terminal (3) to move to the ideal observation position coordinates. Identify the target image collected by the mobile augmented reality terminal (3) at the coordinates of the ideal observation position, and calculate the high-precision liquid level value; The high-precision liquid level value is used as the reference true value and compared with the first liquid level value identified by the fixed camera (2). The parallax correction parameters of the fixed camera (2) are updated according to the comparison result.

2. The intelligent identification and analysis system for graduated cylinder liquid level based on computer vision according to claim 1, characterized in that, The data processing server (5) pre-stores the physical geometric parameters of the graduated cylinder level gauge (1), which include the spatial position equation of the connecting pipe axis relative to the visual anchor point (4), the normal vector of the scale plate plane, and the physical height range of the scale plate. The data processing server (5) uses the PnP algorithm to calculate the extrinsic parameter matrix of the mobile augmented reality terminal (3) relative to the visual anchor point (4), thereby establishing a rigid geometric mapping relationship between the mobile augmented reality terminal (3) and the graduated cylinder level gauge (1).

3. The intelligent identification and analysis system for graduated cylinder liquid level based on computer vision according to claim 2, characterized in that, Before sending the guidance instruction, the data processing server (5) performs a rough estimation of the liquid level: The approximate liquid level value is calculated based on the position of the common perpendicular between the current camera line vector of the mobile augmented reality terminal (3) and the axis of the connecting pipe of the graduated cylinder level gauge (1). Based on the rough liquid level value, a target point is determined on the axis of the connecting pipe. The target point is then translated by a preset observation distance along the normal vector direction of the scale plate plane to obtain the coordinates of the ideal observation position.

4. The intelligent identification and analysis system for graduated cylinder liquid level based on computer vision according to claim 3, characterized in that, The mobile augmented reality terminal (3) is used to calculate the spatial deviation vector between its real-time position and the coordinates of the ideal observation position, and to render virtual guide graphics in the near-eye display module until its real-time position enters the low parallax tolerance range centered on the coordinates of the ideal observation position. The low parallax tolerance range refers to the spatial region where the observed optical axis is perpendicular to the plane of the scale and its height is level with the liquid level.

5. The intelligent identification and analysis system for graduated cylinder liquid level based on computer vision according to claim 1, characterized in that, Before identifying the target image, the data processing server (5) performs a multi-factor weighted quality evaluation on the image sequence uploaded by the mobile augmented reality terminal (3) based on an evaluation function. The evaluation function includes a disparity scoring factor, a reflection interference factor, and a sharpness factor, and selects the image with the highest comprehensive score as the target image.

6. The intelligent identification and analysis system for graduated cylinder liquid level based on computer vision according to claim 2, characterized in that, When calculating high-precision liquid level values, the data processing server (5) uses the external parameter matrix when acquiring the target image and the normal vector of the scale plate plane to construct a homography matrix, performs inverse perspective transformation on the target image to generate an orthogonal view, and identifies the scale values ​​in the orthogonal view based on a neural network model.

7. The intelligent identification and analysis system for graduated cylinder liquid level based on computer vision according to claim 1, characterized in that, The data processing server (5) is also used to obtain the remote instrument values ​​of the distributed control unit (6); When the deviation rate between the high-precision liquid level value and the first liquid level value is less than a preset threshold, and the deviation rate between the high-precision liquid level value and the remote instrument value is greater than a preset threshold, it is determined that the sensor corresponding to the distributed control unit (6) has a drift fault.

8. The intelligent identification and analysis system for graduated cylinder liquid level based on computer vision according to claim 1, characterized in that, When updating the parallax correction parameters of the fixed camera (2), the data processing server (5) is used for: The high-precision liquid level value is used to deduce the current parallax correction coefficient observation value of the fixed camera (2) in reverse; The historical parallax correction coefficients of the fixed camera (2) are updated using the observed values ​​of the parallax correction coefficients by employing an exponentially weighted moving average algorithm.

9. The intelligent identification and analysis system for graduated cylinder liquid level based on computer vision according to claim 4, characterized in that, The mobile augmented reality terminal (3) is an intelligent wearable device that integrates an image acquisition sensor and an inertial measurement unit; the low parallax tolerance range is dynamically adjusted according to the real-time Euclidean distance from the optical center of the camera to the graduated cylinder level gauge.

10. A computer vision-based intelligent identification and analysis method for graduated cylinder liquid levels, characterized in that, The computer vision-based graduated cylinder level intelligent identification and analysis system according to any one of claims 1-9 includes the following steps: S100, the fixed camera (2) continuously collects video streams and calculates the first liquid level value, while the distributed control unit (6) obtains the remote instrument value; the data processing server (5) compares the above values ​​in real time, and when the deviation between the two exceeds the preset safety threshold, it generates an on-site verification task and sends it to the mobile augmented reality terminal (3). S200, the mobile augmented reality terminal (3) responds to the on-site verification task and collects environmental images containing visual anchor points (4); the data processing server (5) uses the PnP algorithm to calculate the extrinsic parameter matrix of the mobile augmented reality terminal (3) relative to the visual anchor points (4) and establishes a local spatial coordinate system; S300, Data processing server (5) uses the external parameter matrix obtained in step S200 and the current camera line vector to calculate the position of the common perpendicular line between the line of sight and the axis of the graduated cylinder level gauge (1), and calculates the rough level value. S400, Data Processing Server (5) calculates the coordinates of the ideal observation position to eliminate parallax based on the rough liquid level value and the normal vector of the scale plate plane; Mobile Augmented Reality Terminal (3) displays a virtual guide graphic based on the deviation between its real-time position and the coordinates of the ideal observation position, instructing the operator to move to the orthogonal observation position; S500: When the mobile augmented reality terminal (3) enters the low parallax tolerance range, it automatically collects image sequences and uses multi-factor evaluation to select the target image; the data processing server (5) performs perspective distortion correction and neural network recognition on the target image to obtain a high-precision liquid level value. S600, the data processing server (5) uses the high-precision liquid level value as the reference true value to verify the sensor status of the distributed control unit (6); at the same time, it uses the high-precision liquid level value to reverse deduce the parallax correction coefficient observation value of the fixed camera (2) and update the parameter configuration of the fixed camera (2).

Images