Monocular camera-based real-time oil storage tank attitude monitoring method

Abstract

The invention relates to the field of artificial intelligence and computer vision, and provides a real-time oil storage tank posture monitoring scheme based on a monocular camera for solving the problems of high cost and low precision of a traditional oil storage tank monitoring technology. According to the technical scheme, accurate monitoring of deformation, settlement and early warning of the oil storage tank is achieved, the specific implementation mode comprises the steps that 360-degree full-coverage monitoring is achieved through three anti-explosion high-precision intelligent vision measurement cameras, image distance measurement and attitude estimation are conducted through a depth estimation algorithm framework, the monitoring precision and the real-time performance are improved, the monitoring cost is remarkably reduced, and the method is suitable for popularization and application. The safety production level is improved, innovative and efficient technical support is provided for posture monitoring of the large oil storage tank, and remarkable social and economic benefits are brought.

Description

Technical Field

[0001] The invention belongs to the field of artificial intelligence and computer vision, and specifically discloses a real-time oil storage tank posture monitoring method based on a monocular camera. Background Art

[0002] After the completion and commissioning of large storage tanks, the uneven geological structure of the foundation, different physical properties of the soil, changes in atmospheric temperature, plastic deformation of the foundation, seasonal and periodic changes in groundwater levels, and the load of the tank itself will cause the tank to sink, deform, tilt, etc. In daily inspections and maintenance, tiny settlements and deformations are difficult to observe with the naked eye. They will only be discovered when the instability and deformation accumulate to a certain extent and functional disorders occur. At this time, the deformation of the tank is irreversible, and even major accidents such as tank tilting and tearing may occur. Therefore, high-precision, all-weather, real-time tank posture monitoring methods have become an urgent need.

[0003] The existing measurement technologies include fiber grating strain measurement technology and Beidou measurement technology. Fiber grating strain measurement technology requires grating fiber strain sensors (multiple sets surround the entire tank body) to monitor the deformation of the tank body, and uses a static level sensor installed on the foundation wall to monitor the foundation settlement of the tank body. These sensors need to be pasted and fixed on the tank body, which will destroy the insulation layer of the oil storage tank, and a large number of cables need to be laid in the tank group area, which is complicated to install. Beidou measurement technology requires a Beidou monitoring terminal (generally 4 orthogonally distributed) fixed on the top of the tank body, a differential base station to monitor the settlement of the tank body + multiple sets of tilt sensors fixed on the tank body to monitor the tilt. Although the fixation is simple, the measurement accuracy is only at the centimeter level.

[0004] Existing measuring instruments and technical means are unable to simultaneously meet the needs of high-precision, all-weather, real-time monitoring and intelligent early warning of settlement, deformation, tilt, etc. of large storage tanks. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention proposes a real-time oil storage tank posture monitoring method based on a monocular camera. The data is acquired in real time by the monocular camera, and the camera data is used to construct a high-precision dense point cloud based on an artificial intelligence network to form depth data. It is combined with unsupervised learning to realize model iteration of long-term deformation, settlement and on-site early warning.

[0006] The present invention includes the following technical solutions:

[0007] A real-time oil tank posture monitoring method based on a monocular camera comprises the following steps:

[0008] a) Install visual measurement cameras: Install at least three visual measurement cameras around the oil storage tank, ensuring that each camera can independently cover part of the tank body and together achieve 360° full coverage monitoring;

[0009] b) Image acquisition and preprocessing: The image data of the oil storage tank is collected in real time through a monocular camera, and data preprocessing is performed, including noise removal, deduplication, and deblurring steps, to improve the quality and availability of image data;

[0010] c) Depth Estimation Algorithm: Using an unsupervised network architecture, combined with a single-view depth and multi-view pose network, nearby views are projected to the target image through the trained depth and pose, thereby generating a loss to guide training to obtain the depth information of each pixel in the oil tank image;

[0011] d) 3D reconstruction and deformation analysis: Based on the acquired depth information, combined with the camera internal parameters and camera posture, the oil storage tank is reconstructed in 3D, and the settlement, deformation and tilt of the tank are analyzed;

[0012] e) Real-time monitoring and early warning: The analyzed tank posture data is transmitted to the server in real time for data storage, analysis and monitoring. When abnormal deformation or settlement is found, an early warning signal is issued in time.

[0013] Furthermore, in the above-mentioned real-time oil storage tank posture monitoring method based on a monocular camera, the installation conditions of the visual measurement camera include laying anti-settling cement piles to ensure the stability of the camera.

[0014] Furthermore, in the above-mentioned real-time oil tank posture monitoring method based on a monocular camera, the depth estimation algorithm includes constructing a network model, which includes a part for predicting the image pixel depth and a part for predicting the camera motion posture between two adjacent frames of images; three frames of images are taken as target images and source images during the training process, the depth value of each pixel is obtained through model reasoning, and the source image is projected to the target image using the camera intrinsic parameters and the camera posture, and the model parameters are updated by calculating the loss between the projected image and the projected frame.

[0015] Furthermore, the above-mentioned real-time oil tank posture monitoring method based on a monocular camera,

[0016] 1) The network model includes two parts: Depth CNN and Pose CNN, which are used to predict the image pixel depth and the motion posture of the camera between two adjacent frames of images respectively;

[0017] 2) During the training process, three frames of images (I t ,I t-1 ,I t+1 ), called I t is the target image, and the other two are source images. First, I t The depth value of each pixel is obtained by inference through the Depth CNN model. t , It-1 )、(I t , I t+1 ) The two teams of images are sent to Pose CNN to obtain the motion posture of the camera between adjacent image pairs.

[0018] Take I t To I t-1 Take the projection of a frame as an example:

[0019]

[0020] Where K is the camera internal parameter, Indicates I t The depth value of this frame image, I t→t-1 Indicates I t To I t-1 The image is obtained by frame projection, and the camera internal parameter K is composed of the focal length f and the coordinates of the image center point (c x ,c y ) determines that the camera intrinsic parameter matrix is ​​in the following form. For the same camera, the intrinsic parameters are fixed;

[0021]

[0022] The camera's motion posture is also called the camera posture, which is used to describe the camera's spatial position. It includes the camera's rotation and translation. The rotation R can be described by a three-dimensional matrix, which can describe any rotation of the camera in three directions. Its expression is:

[0023]

[0024] The camera translation t is represented by a 3x1 matrix, where each element represents the translation in the xyz direction,

[0025]

[0026] Merge R and T to get the augmented matrix<R|T> , that is, the camera posture T is in the form of

[0027]

[0028] On the top I t To I t-1 In the process of frame projection, when T is known t and T t-1 In the case of frame image camera pose, from I t Frame I t-1 The camera pose of the frame is transformed into:

[0029]

[0030] For continuous video frames, the camera pose of each frame is changing. If the camera pose of each frame is available in the dataset, it can be used directly.

[0031] If this item does not exist, the camera pose estimation is required through the Pose CNN network. The specific process is to input two images, and the network gives the camera pose transformation matrix required from the first image to the second image through inference.

[0032] Furthermore, in the above-mentioned real-time oil tank posture monitoring method based on a monocular camera, the depth estimation algorithm includes the following steps:

[0033] 1) Obtain the initial image and camera intrinsic parameters: read the target image and the corresponding source image in each sequence, and determine the camera intrinsic parameters;

[0034] 2) Obtain the camera pose of each frame: If there is a camera pose matrix available for each frame, it can be read into the network for use; if not, the Pose CNN network is used to estimate the camera pose. The Pose CNN network takes the target image and all the source images, and after the network passes the feature encoding and decoding, the output is the relative pose between the current image and each source image; for each pair of images, a 6-dimensional vector is output, which respectively represents the 3 Euler angle parameters and 3 spatial translation parameters of each source view to the target image;

[0035] 3) Obtain the depth value of the target image: Use Depth CNN to infer the image, input a single frame of the target image, and the overall network architecture is a U-net structure, based on an encoder-decoder design with skip connections and multi-scale prediction. Depth-related features are obtained through feature encoding, and depth maps at various resolutions are obtained after decoding, where the pixel value of each point represents the depth information of the pixel point;

[0036] 4) Constructing the projection image of the adjacent frame: According to formula 1.1, the projection is performed using the previously obtained information to obtain the projection image of the adjacent frame;

[0037] 5) Construct the loss of the projected image and the projected frame: Taking formula 1.1 as an example, I t To I t-1 The frame is projected to obtain I t→t-1 , through I t→t-1 with I t-1 Construct similarity loss, the formula is L = ∑ p |I t-1 -I t→t-1 |, where p represents each pixel, and the loss L is used to measure the effect of depth estimation;

[0038] 6) Update the model based on the loss: Back propagate the calculated loss and use the Adam optimization algorithm to update each neuron in the model. The update operation is as follows:

[0039]

[0040] Among them, θ represents the parameters of the model, ω is the learning rate, It is the gradient of the loss function J(θ) with respect to the parameter θ. In this way, the depth estimation model can continuously update and optimize the model parameters through the calculation of the loss function and the back propagation of the gradient, thereby improving the model's ability to estimate depth.

[0041] Furthermore, in the above-mentioned real-time oil storage tank posture monitoring method based on a monocular camera, in the three-dimensional reconstruction and deformation analysis steps, multi-point data is combined to reconstruct the 3D data of the tank body, and the elevation difference is calculated by measuring the three-dimensional coordinates of the symmetrical marking points on the tank body, thereby obtaining the tilt settlement value and uneven settlement amount of the tank body.

[0042] Furthermore, the above-mentioned real-time oil storage tank posture monitoring method based on a monocular camera also includes a measurement algorithm and analysis software, which has image acquisition control, camera calibration, mark recognition, image processing and analysis and solution functional modules for real-time solution of tank deformation and settlement parameters.

[0043] Furthermore, the above-mentioned real-time oil storage tank posture monitoring method based on a monocular camera, the measurement algorithm and analysis software also have monitoring, alarm, message, equipment management, account and authority management, deployment management functions, and realize remote access and operation of the system through B / S architecture and mobile terminal.

[0044] The present invention discloses the application of the above-mentioned real-time oil storage tank posture monitoring method based on a monocular camera in the posture monitoring of a large oil storage tank.

[0045] The present invention also discloses an integrated real-time oil tank posture monitoring device based on a monocular camera, which is characterized by comprising:

[0046] 1) Camera installation assembly, used to stably install at least three visual measurement cameras around the oil storage tank, and to ensure the stability of the camera by laying anti-settling cement piles to achieve 360° full coverage monitoring;

[0047] 2) Image processing and analysis system, which is connected to the visual measurement camera and includes:

[0048] 21) Image processing unit, used to receive and pre-process the oil tank image data collected by the monocular camera in real time, and perform operations such as denoising, de-duplication, and deblurring to improve image quality;

[0049] 22) Depth estimation module, which adopts an unsupervised network architecture and integrates Depth CNN and Pose CNN network models. Depth CNN predicts the image pixel depth, and Pose CNN predicts the camera motion posture between adjacent frames. The training is guided by the loss generated by the trained depth and posture projection to obtain the depth information of each pixel in the oil tank image.

[0050] 23) 3D reconstruction and analysis unit, which reconstructs the oil tank in 3D based on depth information, camera internal parameters and camera posture, and analyzes the settlement, deformation, tilt and other states of the tank;

[0051] 3) Measurement algorithm and analysis software, which has image acquisition control, camera calibration, mark recognition, image processing and analysis and solution functions, is used to solve the deformation and settlement parameters of the tank in real time, and integrates monitoring, alarm, message notification, equipment management, account and authority management, deployment management functions, and supports B / S architecture and mobile terminal remote access and operation;

[0052] 4) Data server, used to store and analyze the tank posture data transmitted from the 3D reconstruction and analysis unit, and to perform historical data analysis to predict future deformation or settlement trends;

[0053] 5) The monitoring and early warning system is connected to the 3D reconstruction and analysis unit and the data server to receive and analyze the tank posture data in real time. When abnormal deformation or settlement is found, early warning signals are issued in time through measurement algorithms and analysis software. It can also be integrated into the remote monitoring center to achieve remote monitoring and emergency response through wired or wireless methods.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The present invention discloses a real-time oil storage tank posture monitoring method based on a monocular camera. It draws on existing experience and the advantages and disadvantages of various technical approaches, is based on a high-precision intelligent visual hardware platform and a high-precision photogrammetry algorithm, and combines artificial intelligence, edge computing, big data and other technologies. It breaks through the limitations of large visual data volume, wide bandwidth transmission, and high computing power requirements. At the same time, it gives full play to the advantages of high spatial resolution, high precision, high dynamics, real-time, large field of view array measurement, non-contact, and intelligent recognition of vision, and develops a set of multifunctional large-scale storage tank safety operation and maintenance software and hardware platform system with high-precision settlement, deformation, tilt and other real-time measurement, disaster warning, and intelligent recognition. This system mainly performs real-time measurement of the deformation, settlement and tilt state of the tank body. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Visual measurement camera layout renderings;

[0057] Figure 2Depth CNN structure diagram;

[0058] Figure 3 Schematic diagram of Depth CNN algorithm;

[0059] Figure 4 Software overall architecture diagram;

[0060] Figure 5 GB50128-2014 Permissible value of radial settlement difference of tank foundation;

[0061] Figure 6 Depth visualization legend, from left to right, represents the depth from near to far;

[0062] Figure 7 Test effect display 1;

[0063] Figure 8 Test effect display 2;

[0064] Fig. 9 Test effect display 3;

[0065] Fig.10 Test effect display 4. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Embodiment 1:

[0068] This embodiment takes the reasonable installation space of a single tank as an example. The size of the tank is 80m *21m (H), the distance between the tanks is 43m; there are 3 fixed installation points corresponding to a single tank, and each explosion-proof high-precision intelligent visual measurement camera can independently cover the envelope of the tank to 120°. Therefore, it is designed that 3 explosion-proof high-precision intelligent visual measurement cameras will cover a single tank 360° in total. Installation conditions of explosion-proof high-precision intelligent visual measurement cameras: To ensure the stability of the camera, anti-settling cement piles need to be laid. The layout effect diagram of the explosion-proof high-precision intelligent visual measurement camera of the visual monitoring system is as follows Figure 1 Shown:

[0069] This method proposes an unsupervised network architecture for image depth acquisition of monocular video sequences. The basic idea is to use single-view depth and multi-view pose networks, and use the calculated depth and pose to project nearby views to the target, thereby generating losses to guide training.

[0070] (1) Design of algorithm network framework for image ranging based on monocular camera

[0071] The basic structure of the network is as follows Figure 2 As shown in Figure 1, it includes two parts: Depth CNN and Pose CNN, which are used to predict the image pixel depth and the camera motion posture between two adjacent frames of images respectively. During the training process, three frames of images (I t ,I t-1 ,I t+1 ), called I t is the target image, and the other two are the source images. t The depth value of each pixel is obtained by inference through the Depth CNN model. t , I t-1 )、(I t , I t+1 ) The two teams of images are sent to Pose CNN to obtain the motion posture of the camera between adjacent image pairs. Take I t To I t-1 Take the projection of a frame as an example:

[0072]

[0073] Where K is the camera internal parameter, Indicates I t The depth value of this frame image, I t→t-1 Indicates I t To I t-1 The camera internal parameter K is composed of the focal length f and the coordinates of the image center point (c x ,c y ) determines that the camera intrinsic parameter matrix is ​​in the following form. For the same camera, the intrinsic parameters are generally fixed.

[0074]

[0075] The camera's motion posture is also called the camera posture, which is used to describe the camera's spatial position. It includes the camera's rotation and translation. The rotation R can be described by a three-dimensional matrix, which can describe any rotation of the camera in three directions. Its expression is:

[0076]

[0077] The camera translation t is represented by a 3x1 matrix, where each element represents the translation in the xyz direction,

[0078]

[0079] Merge R and T to get the augmented matrix<R|T> , that is, the camera posture T is in the form of

[0080]

[0081] On the top I t To I t-1 In the process of frame projection, when T is known t and T t-1 In the case of frame image camera pose, from I t Frame I t-1 The camera pose of the frame is transformed into:

[0082]

[0083] For continuous frames of video, the camera pose of each frame is changing. If the camera pose of each frame is available in the dataset, it can be used directly. If it does not exist, the camera pose estimation is required through the Pose CNN network. The specific process is to input two images, and the network gives the camera pose transformation matrix required from the first image to the second image through inference.

[0084] From the above formula, we can know that through the camera internal parameters, the camera posture of each frame and the depth information of the target image, we can construct a projection map from the target image to the source image. The quality of the projection map is determined by the depth map generated by Depth CNN and the camera posture generated by PoseCNN. The more accurate the depth map and camera posture inferred by these two networks, the more similar the projection map is to the source image. Based on the above theory, the general steps of the algorithm for image ranging of monocular cameras in this paper are as follows:

[0085] 1. Get the initial image and camera intrinsic parameters. Read the target image and the corresponding source image in each sequence, and determine the camera intrinsic parameters.

[0086] 2. Get the camera pose of each frame. If there is a camera pose matrix available for each frame, it can be read into the network for use; if not, the Pose CNN network is used to estimate the camera pose. The Pose CNN network takes the target image and all the source images, and after the network passes the feature encoding and decoding, the output is the relative pose between the current image and each source image. For each pair of images, a 6-dimensional vector is output, which respectively represents the 3 Euler angle parameters (representing spatial rotation) and 3 spatial translation parameters from each source view to the target image.

[0087] 3. Obtain the depth value of the target image. Use Depth CNN to infer the image, input a single frame of the target image, and the overall network architecture is similar to the U-net structure. Figure 2As shown, based on the encoder-decoder design with skip connections and multi-scale prediction, depth-related features are obtained through feature encoding, and depth maps at various resolutions are obtained after decoding, where the pixel value of each point represents the depth information of the pixel.

[0088] 4. Construct the projection image of the adjacent frame. According to formula 1.1, the projection is performed using the previously obtained information to obtain the projection image of the adjacent frame.

[0089] 5. Construct the loss of the projected image and the projected frame. Taking formula 1.1 as an example, I t To I t-1 The frame is projected to obtain I t→t-1 , through I t→t-1 with I t-1 Construct similarity loss, the formula is L = ∑ p |I t-1 -I t→t-1 |, where p represents each pixel. The depth estimation effect is measured by the loss L. The algorithm is shown as follows Figure 3 shown.

[0090] 6. Update the model based on the loss. Back propagate the calculated loss and use the Adam optimization algorithm to update each neuron in the model. The update operation is usually as follows:

[0091]

[0092] Among them, θ represents the parameters of the model, ω is the learning rate, is the gradient of the loss function J(θ) with respect to the parameter θ. In this way, the depth estimation model can continuously update and optimize the model parameters through the calculation of the loss function and the back propagation of the gradient, thereby improving the model's ability to estimate depth.

[0093] (2) Collection of algorithm training data and training of algorithm models

[0094] At the beginning of the project, web crawler technology will be used to collect relevant data of oil storage tanks. In order to ensure the accuracy and availability of the data, this method will perform meticulous data preprocessing. First, the data information of the image is checked to ensure that each piece of data has the correct label so that the deep learning model can correctly understand and learn. Each pixel value in the depth label usually represents the distance from the camera to a point in the scene. Checking the range of depth values ​​can help identify outliers (such as invalid depth or noise). For example, in indoor scenes, the effective depth range of the depth sensor is 0-10m, and data beyond this range is considered invalid depth data. Then data denoising is performed, which is one of the very important steps in data preprocessing. It is a process of recovering more accurate and reliable signals or patterns by removing noise or invalid information from the data. This step can be performed using Gaussian smoothing denoising. A Gaussian filter is applied to the data. It will weight the surrounding values ​​according to the normal distribution and is suitable for removing Gaussian noise. Data deduplication and deblurring are also key steps. Image data deduplication refers to identifying and removing duplicate image data to avoid storage redundancy and improve processing efficiency. By extracting image features (such as color histograms, local feature points, etc.), the feature vectors of similar images can be compared to find duplicate images. Image deblurring refers to the process of restoring or improving image quality degraded by blur effects (such as camera shake, motion blur, unclear focus, etc.) through algorithms. Blurring causes the details of the image to be lost and reduces the clarity of the image. Deblurring aims to reconstruct these lost details. This can be done by performing a frequency domain transform (such as Fourier transform) on the image to separate the blur and detail parts, thereby restoring a clear image. After completing data preprocessing, this method will organize and classify the dataset to better serve the machine learning model. Next, the dataset will be split into a training set, a validation set, and a test set. The training set is used for the initial training of the model, the validation set is used to adjust the model parameters during the training process, and the test set is used to evaluate the final performance of the model.

[0095] (3) Based on the data set collected in specific scenarios, by adding image data collected on site, including oil tank data maps of various scenarios, and combining the two steps 1) and 2), the scene migration of oil tank data is realized, making the model more adaptable to actual application scenarios, further improving the detection accuracy, and improving its performance in actual environments. In this way, the existing real oil tank data can be converted into virtual data of the oil tank area through the generative model, while providing more training data, and the detection accuracy of the model can be further improved through local data enhancement. In this way, a more robust and accurate oil tank detection system is built to meet the needs of high-precision detection in practical applications.

[0096] Example 2

[0097] Based on Example 1, the measurement algorithm and analysis software are disclosed.

[0098] The measurement and analysis software can calculate the deformation and settlement parameters of the tank in real time through functional modules such as image acquisition control, camera calibration, mark recognition, image processing and analysis and solution. Figure 4 shown.

[0099] The overall project is divided into four parts: client, server, browser, and mobile.

[0100] (1) Client

[0101] The camera software has the ability to collect and perform edge computing, and can perform simple analysis on the data collected by the camera and transmit it to the server. It also has the ability to report status, receive instructions (take photos to troubleshoot faults), and perform OTA self-upgrades.

[0102] (2) Server

[0103] The server backend has basic functions such as data reception, data storage, data analysis, spatial data services (combining multi-point data to reconstruct tank 3D data), OTA upgrades, etc. It also has monitoring, alarm, messaging, equipment management, account and permission management, deployment management and other functions, and is the core of the entire system.

[0104] (3) Browser

[0105] The system adopts B / S architecture and can be easily accessed using mainstream browsers. The browser front end can display various functions of the server, making it easy for system-related personnel to operate.

[0106] (4) Mobile:

[0107] It supports Android, and can also conveniently control various functions of the system. The deployment assistant can obtain the surrounding camera and monitoring point (target) data through the positioning of the mobile phone, automatically obtain surrounding tasks, and assist installers in performing related operations.

[0108] According to GB50128-2014 related standards ( Figure 5 ), calculate the foundation settlement (inclined settlement, uneven settlement) and tank ovality. The uneven settlement value of the tank should not exceed the requirements of the design documents. When there is no requirement in the design documents, the settlement difference in the diameter direction of the tank foundation shall not exceed the provisions of the following table, and the foundation part supporting the tank wall shall not have a sudden settlement change; the settlement difference within any 10m arc length along the circumference of the tank wall shall not be greater than 25mm.

[0109] S1 Tank Tilt Settlement Calculation

[0110] By measuring the three-dimensional coordinates of the symmetrical marking points on the tank body, the elevation difference can be obtained, thereby calculating the inclination and settlement value of the tank body.

[0111] S2 tank uneven settlement calculation

[0112] Uneven settlement of tank foundation U i The following two requirements should be met:

[0113]

[0114] In the formula, S i is the relative vertical deformation of observation point i, U i , U i-1 , U i+1 They are the local settlement of observation point i, observation point i-1, and observation point i+1, respectively. L is the arc length between observation points, H is the height of the tank (m), Y is the yield strength of the bottom tank wall steel, and E is the elastic modulus of the bottom tank wall steel (kPa). 8 cameras measure the 3D coordinates of 24 settlement observation points arranged on the foundation ring wall and calculate the elevation values. The relative vertical deformation of the observation points is calculated by the above two formulas, and then the tank foundation settlement curve is drawn.

[0115] S3 tank ovality calculation

[0116] The least squares fitting method is used to calculate the ellipticity of each measuring and monitoring layer of the tank (the ellipticity is calculated by measuring the three-dimensional coordinates of the marking points of each layer), and the deformation of the tank is analyzed at any time. The least squares method can easily obtain unknown data by minimizing the sum of squares of errors to find the best function match for the data, and minimize the sum of squares of errors between these data and the actual data.

[0117] The general equation of an ellipse is:

[0118] Ax 2 +Bxy+Cy 2 +Dx+Ey+F=0

[0119] A, B, C, D, E, and F are calculated using the monitoring data of each layer, and then the tank ovality is calculated according to the following formula:

[0120] σ=ab

[0121] in

[0122]

[0123] Where σ is the ellipticity, a is the major axis of the ellipse, and b is the minor axis of the ellipse.

[0124] Effect: For various oil tank data images, the depth monitoring model can quickly obtain the depth value of the corresponding actual scene, and has high accuracy, fast speed, and can well distinguish the main body of the oil tank from the background. By performing depth monitoring on oil tank data images from various angles, the depth visualization image, for example Figure 6 As shown, from left to right, the depth is from near to far; the test results are shown as follows Figure 7-10 As shown (the left side of each picture is the captured RGB image, and the right side is the depth image obtained by the model).

[0125] From the above examples, it can be seen that in the field of large oil storage tank posture monitoring, the visual monitoring technology based on monocular camera has significant practical application value and social and economic benefits, which are mainly reflected in the following aspects:

[0126] 1. Improve production safety: This technology can timely detect and warn of potential safety hazards through high-precision, all-weather, real-time monitoring of the settlement, deformation and tilt of storage tanks, thereby effectively avoiding major accidents such as tank tilting and tearing.

[0127] 2. Reduce monitoring costs: Compared with traditional fiber grating strain measurement technology and Beidou measurement technology, the visual monitoring technology based on a monocular camera reduces the damage to the tank insulation layer and the laying of a large number of cables, simplifies the installation process, and reduces monitoring costs.

[0128] 3. Social and economic benefits: By improving the safety and operating efficiency of storage tanks, reducing economic losses and environmental risks caused by storage tank accidents, and improving the management efficiency of tank areas, it brings significant social and economic benefits.

[0129] The above are limited to several preferred embodiments of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A real-time oil tank posture monitoring method based on a monocular camera, characterized in that: The following steps are involved: a) Install visual measurement cameras: Install at least three visual measurement cameras around the oil storage tank, ensuring that each camera can independently cover part of the tank body and together achieve 360° full coverage monitoring; b) Image acquisition and preprocessing: The image data of the oil storage tank is collected in real time through a monocular camera, and data preprocessing is performed, including noise removal, deduplication, and deblurring steps, to improve the quality and availability of image data; c) Depth Estimation Algorithm: Using an unsupervised network architecture, combined with a single-view depth and multi-view pose network, nearby views are projected to the target image through the trained depth and pose, thereby generating a loss to guide training to obtain the depth information of each pixel in the oil tank image; d) 3D reconstruction and deformation analysis: Based on the acquired depth information, combined with the camera internal parameters and camera posture, the oil storage tank is reconstructed in 3D, and the settlement, deformation and tilt of the tank body are analyzed; e) Real-time monitoring and early warning: The analyzed tank posture data is transmitted to the server in real time for data storage, analysis and monitoring. When abnormal deformation or settlement is found, an early warning signal is issued in time.

2. The method according to claim 1, characterized in that The installation conditions of the visual measurement camera include laying out anti-settling cement piles to ensure the stability of the camera.

3. The method according to claim 1, characterized in that The depth estimation algorithm includes constructing a network model, which includes a part for predicting the pixel depth of an image and a part for predicting the camera motion posture between two adjacent frames of images; taking three frames of images as target images and source images during the training process, obtaining the depth value of each pixel through model reasoning, and projecting the source image to the target image using the camera intrinsic parameters and the camera posture, and updating the model parameters by calculating the loss between the projected image and the projected frame.

4. The method according to claim 3, characterized in that 1) The network model includes two parts: Depth CNN and Pose CNN, which are used to predict the image pixel depth and the motion posture of the camera between two adjacent frames of images respectively; 2) During the training process, three frames of images (I t ,I t-1 ,I t+1 ), called I t is the target image, and the other two are source images. First, I t The depth value of each pixel is obtained by inference through the Depth CNN model. t , I t-1 )、(I t , I t+1 ) The two teams of images are sent to Pose CNN to obtain the motion posture of the camera between adjacent image pairs. Take I t To I t-1 Take the projection of a frame as an example: Where K is the camera internal parameter, Indicates I t The depth value of this frame image, I t→t-1 Indicates I t To I t-1 The image is obtained by frame projection, and the camera internal parameter K is composed of the focal length f and the coordinates of the image center point (c x ,c y ) determines that the camera intrinsic parameter matrix is ​​in the following form. For the same camera, the intrinsic parameters are fixed; The camera's motion posture is also called the camera posture, which is used to describe the camera's spatial position. It includes the camera's rotation and translation. The rotation R can be described by a three-dimensional matrix, which can describe any rotation of the camera in three directions. Its expression is: The camera translation t is represented by a 3x1 matrix, where each element represents the translation in the xyz direction, Merge R and T to get the augmented matrix<R|T> , that is, the camera posture T is in the form of On the top I t To I t-1 In the process of frame projection, when T is known t and T t-1 In the case of frame image camera pose, from I t Frame I t-1 The camera pose of the frame is transformed into: For continuous video frames, the camera pose of each frame is changing. If the camera pose of each frame is available in the dataset, it can be used directly. If this item does not exist, the camera pose estimation is required through the Pose CNN network. The specific process is to input two images, and the network gives the camera pose transformation matrix required from the first image to the second image through inference.

5. The method according to claim 4, characterized in that The depth estimation algorithm consists of the following steps: 1) Obtain the initial image and camera intrinsic parameters: read the target image and the corresponding source image in each sequence, and determine the camera intrinsic parameters; 2) Obtain the camera pose of each frame: If there is a camera pose matrix available for each frame, it can be read into the network for use; if not, the Pose CNN network is used to estimate the camera pose. The Pose CNN network takes the target image and all the source images, and after the network passes the feature encoding and decoding, the output is the relative pose between the current image and each source image; for each pair of images, a 6-dimensional vector is output, which respectively represents the 3 Euler angle parameters and 3 spatial translation parameters of each source view to the target image; 3) Obtain the depth value of the target image: Use Depth CNN to infer the image, input a single frame of the target image, and the overall network architecture is a U-net structure, based on an encoder-decoder design with skip connections and multi-scale prediction. Depth-related features are obtained through feature encoding, and depth maps at various resolutions are obtained after decoding, where the pixel value of each point represents the depth information of the pixel point; 4) Constructing the projection image of the adjacent frame: According to formula 1.1, the projection is performed using the previously obtained information to obtain the projection image of the adjacent frame; 5) Construct the loss of the projected image and the projected frame: Taking formula 1.1 as an example, I t To I t-1 The frame is projected to obtain I t→t-1 , through I t→t-1 with I t-1 Construct similarity loss, the formula is L = ∑ p |I t-1 -I t→t-1 |, where p represents each pixel, and the loss L is used to measure the effect of depth estimation; 6) Update the model based on the loss: Back propagate the calculated loss and use the Adam optimization algorithm to update each neuron in the model. The update operation is as follows: Among them, θ represents the parameters of the model, ω is the learning rate, It is the gradient of the loss function J(θ) with respect to the parameter θ. In this way, the depth estimation model can continuously update and optimize the model parameters through the calculation of the loss function and the back propagation of the gradient, thereby improving the model's ability to estimate depth.

6. The method according to claim 1, characterized in that In the three-dimensional reconstruction and deformation analysis step, the 3D data of the tank body is reconstructed by combining multi-point data, and the elevation difference is calculated by measuring the three-dimensional coordinates of the symmetrical marking points on the tank body, so as to obtain the tilt settlement value and uneven settlement amount of the tank body.

7. The method according to claim 1, characterized in that It also includes measurement algorithms and analysis software, which has image acquisition control, camera calibration, mark recognition, image processing and analysis and solution functional modules, and is used to solve the tank deformation and settlement parameters in real time.

8. The method according to claim 5, characterized in that The measurement algorithm and analysis software also have monitoring, alarm, message, equipment management, account and authority management, and deployment management functions, and can achieve remote access and operation of the system through B / S architecture and mobile terminals.

9. Application of the method according to any one of claims 1 to 8 in posture monitoring of large oil storage tanks.

10. An integrated real-time oil tank posture monitoring device based on a monocular camera, characterized in that: include: 1) Camera installation assembly, used to stably install at least three visual measurement cameras around the oil storage tank, and to ensure the stability of the camera by laying anti-settling cement piles to achieve 360° full coverage monitoring; 2) Image processing and analysis system, which is connected to the visual measurement camera and includes: 21) Image processing unit, used to receive and pre-process the oil tank image data collected by the monocular camera in real time, and perform operations such as denoising, de-duplication, and deblurring to improve image quality; 22) Depth estimation module, which adopts an unsupervised network architecture and integrates Depth CNN and Pose CNN network models. Depth CNN predicts the image pixel depth, and Pose CNN predicts the camera motion posture between adjacent frames. The training is guided by the training depth and posture projection loss to obtain the depth information of each pixel in the oil tank image. 23) 3D reconstruction and analysis unit, which reconstructs the oil tank in 3D based on depth information, camera internal parameters and camera posture, and analyzes the settlement, deformation, tilt and other states of the tank; 3) Measurement algorithm and analysis software, which has image acquisition control, camera calibration, mark recognition, image processing and analysis and solution functions, is used to solve the deformation and settlement parameters of the tank in real time, and integrates monitoring, alarm, message notification, equipment management, account and authority management, deployment management functions, and supports B / S architecture and mobile terminal remote access and operation; 4) Data server, used to store and analyze the tank posture data transmitted from the 3D reconstruction and analysis unit, and to perform historical data analysis to predict future deformation or settlement trends; 5) The monitoring and early warning system is connected to the 3D reconstruction and analysis unit and the data server to receive and analyze the tank posture data in real time. When abnormal deformation or settlement is found, early warning signals are issued in time through measurement algorithms and analysis software. It can also be integrated into the remote monitoring center to achieve remote monitoring and emergency response through wired or wireless methods.

Images