Smoke plume flow velocity inversion and verification and fusion imaging high-precision emission flux quantification method

Abstract

The invention discloses a high-precision emission flux quantification method for smoke plume flow velocity inversion and verification and fusion imaging, and the method comprises the steps: extracting plume features in an image through employing a clustering algorithm in the smoke plume flow velocity inversion measurement of an improved dense optical flow method Farneback, and screening a plume main body; calculating image texture intensity based on the plume main body and determining optical flow parameters corresponding to the image texture intensity, calculating forward optical flow based on the optical flow parameters, introducing a consistency error field based on bidirectional optical flow, and performing time consistency correction on the forward optical flow based on the consistency error field; a fusion weight factor is determined based on calculation of the gradient of the forward light flow, and the flow velocity is calculated after weighted total variation denoising is performed on the forward light flow, so that a smoke plume flow velocity inversion result is obtained; and verifying the smoke plume flow velocity inversion result, and combining the smoke plume flow velocity inversion result with hyper-spectral remote sensing imaging to realize high-precision emission flux quantification, so that the emission flux calculation precision of industrial organized emission can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of plume velocity inversion method and corresponding accuracy verification, and specifically relates to a high-precision emission flux quantification method of plume velocity inversion and verification and fusion imaging. Background Art

[0002] Currently, industrial emissions remain a major source of air pollutants. Accurately calculating the emission fluxes of organized factory emissions and establishing a dynamic emission inventory, particularly those at chimney outlets, are crucial for air pollution control. Obtaining information on plume concentration and velocity is crucial for calculating chimney outlet emission fluxes.

[0003] Currently, methods for calculating plume velocity information can be divided into: wind speed data approximation, traditional optical flow, and deep learning-based modeling. Wind speed data approximation approximates plume velocity using wind speed data, but ignores the plume's own rising velocity. Furthermore, methods for obtaining wind speed based on in-situ measurements and model simulations suffer from systematic errors due to matching issues in the observation space. Traditional optical flow methods suffer from plume misidentification, poor temporal consistency, and difficulty determining parameters. Although deep learning-based modeling offers certain advantages, it is prone to prediction bias when faced with scenarios not covered by the training set (such as new emission sources or extreme meteorological conditions). Furthermore, deep learning methods suffer from poor interpretability, and training sets in real-world scenarios are difficult to obtain.

[0004] Currently, machine vision-based gas velocity calculation methods (such as optical flow and deep learning-based modeling) are typically validated using CFD-based model verification or gas distribution device verification. However, both methods suffer from a common drawback: excessive reliance on the accuracy of the model or gas distribution device itself, and a lack of observational data independent of visual methods for verification. Machine vision-based gas velocity calculation methods can be used to quantify chimney emission fluxes, but they also require information on the gas concentration distribution.

[0005] Currently, a common method for acquiring concentration data in conjunction with gas flow rate calculation methods is to use a gas camera. However, using a gas camera to obtain concentration distribution presents several challenges. First, due to the limitations of optical filters, gas cameras can only capture very limited information about gas components, typically only measuring sulfur dioxide and nitrogen dioxide. Second, gas cameras cannot effectively eliminate cross-absorption interference between multiple gas components. Furthermore, due to the nonlinearity of the calibration factor, gas cameras may introduce conversion errors, further affecting the accuracy of concentration data.

[0006] Therefore, how to obtain a plume velocity calculation method with high spatial matching and good scene adaptability based on actual observation data, and how to design a corresponding verification method to achieve high-precision emission flux calculation, has become a technical problem that needs to be solved urgently. Summary of the Invention

[0007] In view of the above, the purpose of the present invention is to provide a high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging, which provides a high-precision, robust plume velocity calculation method and its accuracy verification method that are independent of training data for the field of environmental monitoring, and combines a ground-based hyperspectral remote sensing system with rapid high-resolution imaging of multi-component atmospheric trace components to achieve high-precision emission flux quantification, further improving the emission flux calculation accuracy of industrial organized emissions.

[0008] To achieve the above-mentioned purpose of the invention, the embodiment provides a high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging, comprising the following steps:

[0009] In the Farneback dense optical flow method for plume velocity inversion, a clustering algorithm is used to extract plume features from images and screen the plume body. The image texture intensity is calculated based on the plume body, and the optical flow parameters corresponding to the image texture intensity are determined. The forward optical flow is calculated based on the optical flow parameters. A consistency error field based on bidirectional optical flow is introduced, and the forward optical flow is corrected for temporal consistency based on the consistency error field. The fusion weight factor is determined based on the calculated gradient of the forward optical flow. The forward optical flow is then subjected to weighted total variation denoising based on the fusion weight factor, and the velocity is calculated to obtain the plume velocity inversion result.

[0010] The plume velocity inversion results were verified and combined with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification.

[0011] Preferably, a clustering algorithm is used to extract plume features in the image and screen the plume body, including:

[0012] A clustering algorithm is used to cluster pixels to extract plume features. The plume is considered to be the area with the highest brightness, and the cluster with the largest channel value is selected as the plume body.

[0013] Preferably, calculating the image texture intensity based on the plume body and determining the optical flow parameters corresponding to the image texture intensity include:

[0014] Based on the plume body, the gradient fields in the horizontal and vertical directions are calculated, and the gradient amplitude statistics of each pixel are counted based on the gradient fields in the two directions. The standard deviation of the gradient amplitude statistics is used to represent the image texture intensity. The optical flow parameters corresponding to the image texture intensity are determined according to the corresponding relationship between the image texture intensity and the optical flow parameters.

[0015] Preferably, the corresponding relationship is:

[0016] When the image texture intensity σ is in [0,5), the texture level is very low texture, corresponding to a pyramid scaling factor of 0.2, a pyramid layer number of 2, a window size of 7, and a number of iterations of 2;

[0017] When the image texture intensity σ is in [5,10), the texture level is low texture, corresponding to the pyramid scaling factor of 0.25, the number of pyramid layers of 3, the window size of 11, and the number of iterations of 3;

[0018] When the image texture intensity σ is in [10, 20), the texture level is medium texture, corresponding to the pyramid scaling factor of 0.4, the number of pyramid layers of 4, the window size of 15, and the number of iterations of 4;

[0019] When the image texture intensity σ is in [20,30), the texture level is high texture, corresponding to the pyramid scaling factor of 0.5, the number of pyramid layers is 5, the window size is 21, and the number of iterations is 5;

[0020] When the image texture intensity σ is in [30, +∞), the texture level is extremely high texture, corresponding to a pyramid scaling factor of 0.6, 6 pyramid levels, 25 window size, and 6 iterations.

[0021] Preferably, the forward optical flow is calculated based on the optical flow parameters, a consistency error field based on the bidirectional optical flow is introduced, and temporal consistency correction is performed on the forward optical flow based on the consistency error field, including:

[0022] The dense optical flow method Farneback is used to calculate the forward optical flow F based on the optical flow parameters f , the consistency error field E based on bidirectional optical flow is introduced and expressed as:

[0023] E=|F f +F b |

[0024] Among them, F b is the reverse optical flow obtained by reverse calculation,

[0025] The temporal consistency correction of the forward optical flow is performed based on the consistency error field, which is expressed as:

[0026] F=F f -λ·E

[0027] Among them, λ is the corrected weight factor, and F represents the corrected forward optical flow.

[0028] Preferably, a fusion weight factor is determined based on the gradient of the forward optical flow, and the forward optical flow is subjected to weighted total variation denoising based on the fusion weight factor to calculate the flow velocity, thereby obtaining a plume velocity inversion result, including:

[0029] Calculate the magnitude gradient G of the forward optical flow flow , using the linear combination gradient method based on the amplitude gradient G of the forward optical flow flow and the gradient magnitude statistic G image Get the fusion weight factor w:

[0030] w=α·G flow +(1-α)·G image

[0031] Among them, α represents the weight of the optical flow amplitude gradient;

[0032] Based on the fusion weight factor, weighted total variation denoising is performed on the forward optical flow. The specific optimization goal of weighted total variation denoising is to make the optical flow smoother while maintaining clear edges. The energy function is constructed in combination with the weight factor, and the energy functional E(F) is minimized:

[0033] E(F)=∫w·G flow (F)dxdy

[0034] G flow (F) represents the gradient of the forward optical flow. After varying the energy functional E(F), we get the divergence operator. The divergence is the embodiment of the energy gradient and is used as a tool to optimize the energy functional. The discretized gradient divergence is calculated as:

[0035]

[0036] ∈ is a very small quantity, is the weight update term, (i, j) represents the coordinate position, ΔF x (i,j) and ΔF y (i, j) represents the change in gradient divergence in the horizontal x direction and the vertical y direction. Gradient descent optimizes the energy functional through divergence. The formula for gradient descent iteratively updating the optical flow field is as follows:

[0037]

[0038] η is the learning rate of gradient descent, the update direction is the negative energy gradient, and k represents the update round;

[0039] The forward optical flow after iterative optimization is recorded as F′, which includes the gradient divergence F in the horizontal x direction and the vertical y direction. x ′ (i,j) and F y ′ (i,j), then:

[0040]

[0041] m(i,j) refers to the amplitude of the displacement vector of the (i,j)th pixel in a frame, that is, the pixel frame speed. The actual speed v(i,j) of each pixel is:

[0042] v(i,j)=m(i,j)×pm×FPS

[0043] The direction of the true velocity θ(x,y) is:

[0044] θ(x,y)=tan -1 (F x ′ (i,j),F y ′ (i,j))

[0045] Among them, pm is the conversion coefficient between pixels and actual physical relationship, and FPS refers to the frame rate of the video.

[0046] Preferably, the plume velocity inversion results are verified, including:

[0047] Based on the particle tracking model provided by COMSOL software, the plume velocity data was exported as model data. The model data was filtered out of low velocities and the velocity data at different heights in the same frame were averaged to obtain the true value data.

[0048] The plume velocity inversion result data is also filtered out with low speed and the velocity data at different heights in the same frame are averaged, and then the correlation with the true value data is verified.

[0049] Preferably, the plume velocity inversion results are verified, including:

[0050] The pulsed smoke generator and wind speed detector are used to verify the inversion results of the smoke plume velocity, including:

[0051] A pulsed smoke generator is used in conjunction with a wind speed detector to collect anemometer data. At the same time, the smoke plume velocity inversion results are matched with the pulse interval and time resolution of the pulsed smoke generator. The error and variance between the matched plume velocity inversion result data and the anemometer data are calculated to achieve verification.

[0052] Preferably, the plume velocity inversion results are combined with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification, including:

[0053] Before observation, the acquisition angle of the image acquisition camera in the improved dense optical flow method Farneback is ensured to be consistent with the observation angle of the hyperspectral remote sensing system. The spatial resolution of the camera is usually set higher than that of the ground-based hyperspectral remote sensing system to match the spatial resolution. The temporal resolution is also matched between the acquired image and the integration time of the ground-based hyperspectral remote sensing system. On this basis, the plume velocity inversion results are combined with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification.

[0054] Preferably, the plume velocity inversion results are combined with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification, including:

[0055]

[0056] Where M is the molecular mass of the component whose flux is required, N A is Avogadro's constant, SCD i,j is the column concentration value of the jth column and the ith row, v i,j is the velocity value in the jth column and the ith row, N is the number of pixels in the jth column that contain plumes, and h i,j It refers to the actual height represented by the height of each pixel in the results of the ground-based hyperspectral remote sensing system.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention improves the dense optical flow method Farneback and performs plume velocity inversion measurement based on the improved dense optical flow method Farneback, which can improve the accuracy and robustness of plume velocity inversion and is independent of training data. On this basis, a verification scheme is proposed to verify the effect of the dense optical flow method Farneback in plume velocity inversion measurement, and the plume velocity inversion results are combined with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification, further improving the emission flux calculation accuracy of industrial organized emissions. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0060] Figure 1 This is a flow chart of a high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging provided in an embodiment;

[0061] Figure 2This is a flowchart of the improved dense optical flow method Farneback provided in the embodiment;

[0062] Figure 3 1 is a schematic diagram of the plume velocity inversion results of Farneback using the improved dense optical flow method provided in the embodiment;

[0063] Figure 4 This is the verification effect of the original optical flow method and model data provided by the embodiment;

[0064] Figure 5 This is the verification effect of the improved dense optical flow method Farneback provided by the embodiment and the model data. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0066] The inventive concept of the present invention is: an embodiment of the present invention provides a high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging, which includes three improvements: first, the improved dense optical flow method Farneback is used to perform plume velocity inversion measurement; second, verification of the plume velocity inversion measurement results of the improved dense optical flow method Farneback is proposed; third, high-precision emission flux quantification is achieved by combining a ground-based hyperspectral remote sensing system for rapid high-resolution imaging of multi-component atmospheric trace components and the plume velocity inversion measurement results.

[0067] like Figure 1 As shown, the embodiment provides a high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging, including the following steps:

[0068] S1, using the improved dense optical flow method Farneback to perform plume velocity inversion measurement and obtain the plume velocity inversion result.

[0069] In the embodiment, the improvement of the dense optical flow method Farneback includes: using a clustering algorithm to extract plume features, adaptive parameter determination based on image texture, time consistency correction, and weighted total variation denoising based on fusion weight factors. Figure 2 As shown in Figure 1, the improved dense optical flow method Farneback is used to perform plume velocity inversion measurement, including the following steps:

[0070] S1-1, clustering algorithm is used to extract plume features in the image and filter the plume body.

[0071] In plume velocity measurement based on the Farneback optical flow method, traditional methods directly calculate the dense optical flow field for the entire image. However, due to the complex interaction between the plume and the background (such as turbulent diffusion, illumination changes, and background interference), it will lead to serious noise interference, blurred target areas, and high computational redundancy. To solve this problem, the present invention separates plume-related vectors (clusters) and background noise (outliers) through clustering analysis of the spatial distribution and motion consistency of optical flow vectors, significantly improving the signal-to-noise ratio.

[0072] First, the cluster aggregation degree is expressed by the sum of square errors to determine the optimal number of clusters k:

[0073]

[0074] Among them, C i is the i-th cluster, x j represents the jth sample, μ i is the i-th cluster center.

[0075] A clustering algorithm is used to cluster pixels to extract plume features. Since the plume is usually the area with the highest brightness, the cluster with the largest channel value is selected as the plume body.

[0076] S1-2, calculating the image texture intensity based on the plume body and determining the optical flow parameters corresponding to the image texture intensity.

[0077] Traditional fixed-parameter optical flow methods for inverting smoke plume velocity have technical drawbacks. For example, setting a window that is too small can cause optical flow discontinuity, while setting a window that is too large can lead to motion blur. The same problem exists with the number of pyramid levels and iterations. Increasing them blindly can lead to redundant computational resources and the risk of overfitting, while decreasing them blindly can prevent convergence in high-gradient regions. Therefore, for smoke plumes in different scenes or at different times, adaptive optical flow parameter adjustments are required based on the texture complexity of the plume in the current frame.

[0078] First, the plume body obtained in S1-1 is recorded as M, and its horizontal and vertical gradient fields M are calculated. x and M y :

[0079] M x =K x *M

[0080] M y =K y *M

[0081] Among them, K x With K y is the Sobel operator kernel:

[0082]

[0083] Then for the (i, j)th pixel, based on the gradient field in two directions and Compute gradient magnitude statistics for:

[0084]

[0085] The gradient magnitude statistics G of all pixels image The standard deviation std is used to represent the image texture intensity σ:

[0086] σ=std(G image )

[0087] The optical flow parameters corresponding to the image texture intensity are determined based on the corresponding relationship between the image texture intensity and the optical flow parameters. The corresponding relationship is shown in Table 1:

[0088] Table 1 Correspondence between image texture intensity and optical flow parameters

[0089]

[0090] S1-3, calculate the forward optical flow based on the optical flow parameters, introduce the consistency error field based on the bidirectional optical flow, and perform temporal consistency correction on the forward optical flow based on the consistency error field.

[0091] In the embodiment, the bidirectional optical flow is first calculated and then the consistency error field E for temporal consistency correction is calculated:

[0092] E=|F f +F b |

[0093] Among them, F f is the forward optical flow obtained by normal calculation (i.e., the pixel motion vector field from time t to t+1), F b is the reverse optical flow obtained by reverse calculation (i.e., the pixel motion vector field from time t+1 to time t). Ideally, the forward and reverse optical flows should be opposite matrices to each other, with E=0. In reality, complex motions such as turbulence and rotation of the smoke plume may cause motion model mismatch, and the reflection / absorption of light during the diffusion of the smoke plume may cause the brightness to not meet the constant assumption, so the consistency error field is not 0. The physical meaning of this inconsistency is a measure of the local unreliability of the optical flow estimation. In order to avoid errors caused by abnormal optical flow vectors, the forward optical flow F should be calculated. f Correction for temporal consistency:

[0094] F=F f -λ·E

[0095] Among them, λ is the correction weight factor, which is set to λ = 0.2 according to the empirical value, and F represents the corrected forward optical flow.

[0096] S1-4, determine the fusion weight factor based on the calculated gradient of the forward optical flow, and perform weighted total variation denoising optimization on the forward optical flow based on the fusion weight factor.

[0097] The significance of total variation denoising lies in preserving edges while smoothing optical flow by minimizing the total variation energy function. By fusing the optical flow and image gradient information and calculating the fusion weight factor, the smoothing strength can be adaptively controlled, balancing the requirements of denoising and edge preservation in motion estimation.

[0098] First, calculate the magnitude gradient G of the forward optical flow flow , perform discretization difference in horizontal and vertical directions on the corrected forward optical flow obtained by S1-3:

[0099]

[0100] Then the amplitude gradient G of the forward optical flow is flow for:

[0101]

[0102] The linear combination gradient is used based on the amplitude gradient G of the forward optical flow flow and the gradient magnitude statistic G image Get the fusion weight factor w:

[0103] w=α·G flow +(1-α)·G image

[0104] Among them, α represents the weight of the optical flow amplitude gradient, which is generally set to 0.7.

[0105] After obtaining the fusion weight factor, the forward optical flow is optimized by weighted total variation denoising. The optimization goal of weighted total variation denoising is to make the optical flow smoother while maintaining edge clarity. The energy function is constructed by combining the weight factor and minimizing the energy functional E(F):

[0106] E(F)=∫w·G flow (F)dxdy

[0107] G flow (F) represents the gradient of the forward optical flow, which minimizes the energy functional E(F). At this time, the overall weighted gradient is minimized, and the image reaches the optimal smooth state while maintaining clear edges.

[0108] After varying the energy functional E(F), we get the divergence operator. The divergence is the embodiment of the energy gradient and is used as a tool to optimize the energy functional. The discretized gradient divergence is calculated as:

[0109]

[0110] ∈ is a very small amount to avoid division by zero error, is the weight update term, (i, j) represents the coordinate position, ΔF x (i,j) and ΔF y (i, j) represents the change in gradient divergence in the horizontal x direction and the vertical y direction. Gradient descent optimizes the energy functional through divergence. The formula for gradient descent iteratively updating the optical flow field is as follows:

[0111]

[0112] η is the learning rate of gradient descent, which is set to 1. The update direction is the negative energy gradient. The optical flow after each update will make the energy functional smaller. k represents the update round.

[0113] S1-5, calculate the velocity after weighted total variation denoising optimization and obtain the plume velocity inversion result.

[0114] The calculation result of the forward optical flow actually records the displacement vector of each pixel in each optical flow body. The forward optical flow after iterative optimization is recorded as F′, which includes the gradient divergence F in the horizontal x direction and the vertical y direction. x ′ (i,j) and F y ′ (i,j), then:

[0115]

[0116] m(i,j) refers to the amplitude of the displacement vector of the (i,j)th pixel in a frame, that is, the pixel frame speed, the unit is pixel / frame, and the actual speed v(i,j) of each pixel is:

[0117] v(i,j)=m(i,j)×pm×FPS

[0118] The direction of the true velocity θ(x,y) is:

[0119] θ(x,y)=tan -1 (F x ′ (i,j),F y ′ (i,j))

[0120] Among them, pm is the conversion coefficient between pixels and actual physical relationship, which is obtained by the rangefinder combined with geometric calculations, and FPS refers to the frame rate of the video.

[0121] Among them, the true speed of each pixel and the corresponding direction constitute the plume velocity inversion result, such as Figure 3 shown.

[0122] S2, verify the plume velocity inversion results.

[0123] The traditional machine vision-based plume velocity calculation method is usually limited to the verification means of the gas distribution device and simple model verification. The existing verification method usually obtains a continuous velocity field (such as a velocity vector diagram or a streamline diagram) by directly solving the Navier-Stokes equations (CFD module). However, the optical flow method relies on the brightness changes in the image, and the direct use of the CFD velocity field cannot effectively simulate the noise, particle occlusion, image sampling rate and other problems in the actual imaging process, which may cause the verification results to be too idealized. To solve this problem, the present invention adopts two methods to test the adaptability of the improved dense optical flow method Farneback of the present invention when facing situations such as uneven particle concentration, image noise, and dynamic range changes.

[0124] Verification method 1

[0125] COMSOL software provides a smoke plume simulation example based on a particle tracking model. Using this example, a 10-second visualization model with a frame rate of 50 was generated as the basic data source for verification. After exporting the model data, the resulting smoke plume velocity data contained 120,000 rows of height information (each height was divided 484 times horizontally) and 500 columns of time information. To verify the accuracy of the improved dense optical flow method, Farneback, the model data was processed as follows:

[0126] (1) Perform low-speed filtering on the model data, i.e., filter out data points with a speed less than 0.1 m / s, in order to filter out areas in the model where there is no particle motion or the particle motion is sparse;

[0127] (2) The speed data at different heights in the same frame are averaged, that is, the data at different heights in the same frame are averaged to obtain 500 data points for verification as the true value data;

[0128] The same processing is done on the plume velocity inversion result data obtained using the optical flow method and the improved dense optical flow method Farneback, and the correlation is verified with the true value data. The results are as follows Figure 4 、 5 Analysis Figure 4 and Figure 5The correlation between the improved Farneback dense optical flow method and the true data increased from 0.76 to 0.91, and the error decreased from 15% to 9%. This shows that the accuracy of the improved Farneback dense optical flow method has been significantly improved compared to the original method.

[0129] Verification method 2

[0130] Existing field verification methods use a gas distribution device. This method relies heavily on the device's readings and fails to capture real-time data on the gases within the diffused plume. This invention utilizes a pulsed smoke generator in conjunction with an anemometer to accurately verify the Farneback method using the modified dense optical flow method. The specific steps are as follows:

[0131] Turn on the pulse smoke generator and record data for 10 seconds at a frame rate of 60. Place the anemometer horizontally at a location where the smoke plume diffusion radius is equivalent to the anemometer rotor radius. Turn on the pulse smoke generator and record the anemometer reading for 10 seconds. Repeat this process multiple times to collect anemometer data.

[0132] At the same time, the inversion results of the plume velocity are matched with the pulse interval and time resolution of the pulse smoke generator. Specifically, for the pulse interval matching, the plume velocity inversion result data obtained by the improved dense optical flow method Farneback needs to be processed as follows:

[0133] V k =max(V k·n ,V k·n+1 ,…,V k·(n+1) )

[0134] Among them, V k represents the kth data point used for validation by the improved optical flow method, and n is the number of frames captured during a pulse. Because the pulsed smoke generator does not produce smoke uniformly, and the anemometer's temporal resolution is limited, the smoke plume velocity captured approximates the velocity at each pulse. Therefore, the maximum value of the plume velocity per n frames is used as the plume velocity captured by the improved dense optical flow method Farneback at each pulse.

[0135] In order to match the time resolution, the time resolution of the improved dense optical flow method Farneback needs to be matched with the time resolution of the anemometer, that is, the plume velocity inversion result data after matching pulse interval processing is resampled according to the time resolution of the anemometer;

[0136] Then the error and variance between the matched plume velocity inversion result data and the anemometer data are calculated to achieve verification.

[0137] S3, combines the plume velocity inversion results with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification.

[0138] The ground-based hyperspectral remote sensing system for rapid high-resolution imaging of multi-component atmospheric trace components can realize push-broom hyperspectral remote sensing imaging of multiple components and can obtain concentration information of a column in the plume at the same time. Compared with the camera snapshot imaging used by the improved dense optical flow method Farneback, the ground-based hyperspectral remote sensing system for rapid high-resolution imaging of multi-component atmospheric trace components has the advantages of eliminating cross-interference effects, high accuracy, and no need for calibration. The steps for achieving high-precision emission flux quantification using the improved dense optical flow method Farneback and the ground-based hyperspectral remote sensing system for rapid high-resolution imaging of multi-component atmospheric trace components are as follows:

[0139] Before observation, ensure that the image acquisition camera's viewing angle in the improved dense optical flow method Farneback is consistent with the observation angle of the hyperspectral remote sensing system, and start observation at the same time;

[0140] Perform spatial resolution matching, that is, set the spatial resolution of the camera to be higher than that of the ground-based hyperspectral remote sensing system, so the number of pixel merging α required by the camera is calculated:

[0141]

[0142] Among them, N c represents the number of pixels occupied by the plume in the camera's field of view, N h,L represents the number of pixels occupied by the plume in the ground-based hyperspectral remote sensing system at L nm. It should be noted that this calculation needs to be performed separately in the horizontal and vertical directions.

[0143] Temporal resolution matching is also performed. Specifically, the temporal resolution of the collected image is matched with the integration time of the ground-based hyperspectral remote sensing system. The typical value of the integration time of the ground-based hyperspectral remote sensing system is 1500ms, so the camera needs to average the 90 frames of data collected within 1500ms.

[0144] Based on this, the plume velocity inversion results are combined with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification. Due to the thickness of the plume near the chimney mouth, light sometimes has difficulty penetrating the plume, leading to underestimated concentrations. The plume's terminal boundary is difficult to define, leading to underestimation. Therefore, the column used to calculate the flux is the maximum value from the column-by-column calculation.

[0145]

[0146] Where M is the molecular mass of the component whose flux is required, N A is Avogadro's constant, SCD i,jis the column concentration value of the jth column and the ith row, v i,j is the velocity value in the jth column and the ith row, N is the number of pixels in the jth column that contain plumes, and h i,j It refers to the actual height represented by the height of each pixel in the results of the ground-based hyperspectral remote sensing system.

[0147] The above method can provide a high-precision, strong robustness and training data-independent plume velocity calculation method and its accuracy verification method, and combined with a ground-based hyperspectral remote sensing system with rapid high-resolution imaging of multi-component atmospheric trace components to achieve high-precision emission flux quantification, further improving the emission flux calculation accuracy of industrial organized emissions.

[0148] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging, characterized by: The following steps are involved: In the improved dense optical flow method Farneback's plume velocity inversion measurement, a clustering algorithm is used to extract plume features in the image and screen the plume body. The image texture intensity is calculated based on the plume body and the optical flow parameters corresponding to the image texture intensity are determined. The forward optical flow is calculated based on the optical flow parameters. A consistency error field based on bidirectional optical flow is introduced and the forward optical flow is corrected for temporal consistency based on the consistency error field. The fusion weight factor is determined based on the calculated forward optical flow gradient. The forward optical flow is weighted and denoised with the fusion weight factor, and the velocity is calculated after the velocity is calculated. The plume velocity inversion result is obtained. The plume velocity inversion results were verified and combined with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification.

2. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 1 is characterized in that: Clustering algorithms are used to extract plume features from images and screen plume entities, including: A clustering algorithm is used to cluster pixels to extract plume features. The plume is considered to be the area with the highest brightness, and the cluster with the largest channel value is selected as the plume body.

3. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 1 is characterized in that: Calculate the image texture intensity based on the plume body and determine the optical flow parameters corresponding to the image texture intensity, including: Based on the plume body, the gradient fields in the horizontal and vertical directions are calculated, and the gradient amplitude statistics of each pixel are counted based on the gradient fields in the two directions. The standard deviation of the gradient amplitude statistics is used to represent the image texture intensity. The optical flow parameters corresponding to the image texture intensity are determined according to the corresponding relationship between the image texture intensity and the optical flow parameters.

4. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 3 is characterized in that: The corresponding relationship is: When the image texture intensity σ is in [0,5), the texture level is very low texture, corresponding to a pyramid scaling factor of 0.2, a pyramid layer number of 2, a window size of 7, and a number of iterations of 2; When the image texture intensity σ is in [5,10), the texture level is low texture, corresponding to the pyramid scaling factor of 0.25, the number of pyramid layers of 3, the window size of 11, and the number of iterations of 3; When the image texture intensity σ is in [10, 20), the texture level is medium texture, corresponding to the pyramid scaling factor of 0.4, the number of pyramid layers of 4, the window size of 15, and the number of iterations of 4; When the image texture intensity σ is in [20,30), the texture level is high texture, corresponding to the pyramid scaling factor of 0.5, the number of pyramid layers is 5, the window size is 21, and the number of iterations is 5; When the image texture intensity σ is in [30, +∞), the texture level is extremely high texture, corresponding to a pyramid scaling factor of 0.6, 6 pyramid levels, 25 window size, and 6 iterations.

5. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 1 is characterized in that: The forward optical flow is calculated based on the optical flow parameters, a consistency error field based on the bidirectional optical flow is introduced, and the temporal consistency correction of the forward optical flow is performed based on the consistency error field, including: The dense optical flow method Farneback is used to calculate the forward optical flow F based on the optical flow parameters f , the consistency error field E based on bidirectional optical flow is introduced and expressed as: E=|F f +F b | Among them, F b is the reverse optical flow obtained by reverse calculation, The temporal consistency correction of the forward optical flow is performed based on the consistency error field, which is expressed as: F=F f -λ·E Among them, λ is the corrected weight factor, and F represents the corrected forward optical flow.

6. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 1 is characterized in that: The fusion weight factor is determined based on the gradient of the forward optical flow. The forward optical flow is weighted and denoised using the fusion weight factor. The velocity is then calculated to obtain the plume velocity inversion result, including: Calculate the magnitude gradient G of the forward optical flow flow , using the linear combination gradient method based on the amplitude gradient G of the forward optical flow flow and the gradient magnitude statistic G image Get the fusion weight factor w: w=α·G flow +(1-a)·G image Among them, α represents the weight of the optical flow amplitude gradient; Based on the fusion weight factor, weighted total variation denoising is performed on the forward optical flow. The specific optimization goal of weighted total variation denoising is to make the optical flow smoother while maintaining clear edges. The energy function is constructed in combination with the weight factor, and the energy functional E(F) is minimized: E(F)=∫w·G flow (F)dxdy G flow (F) represents the gradient of the forward optical flow. After varying the energy functional E(F), we get the divergence operator. The divergence is the embodiment of the energy gradient and is used as a tool to optimize the energy functional. The discretized gradient divergence is calculated as: ∈ is a very small quantity, is the weight update term, (i, j) represents the coordinate position, ΔF x (i,j) and ΔF y (i, j) represents the change in gradient divergence in the horizontal x direction and the vertical y direction. Gradient descent optimizes the energy functional through divergence. The formula for gradient descent iteratively updating the optical flow field is as follows: η is the learning rate of gradient descent, the update direction is the negative energy gradient, and k represents the update round; The forward optical flow after iterative optimization is recorded as F′, which includes the gradient divergence F in the horizontal x direction and the vertical y direction. x ′ (i,j) and F y ′ (i,j), then: m(i,j) refers to the amplitude of the displacement vector of the (i,j)th pixel in a frame, that is, the pixel frame speed. The actual speed v(i,j) of each pixel is: v(i,j)=m(i,j)×pm×FPS The direction of the true velocity θ(x,y) is: θ(x,y)=time -1 (F x ′ (i,j),F y ′ (i,j)) Among them, pm is the conversion coefficient between pixels and actual physical relationship, and FPS refers to the frame rate of the video.

7. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 1 is characterized in that: Verify the plume velocity inversion results, including: Based on the particle tracking model provided by COMSOL software, the plume velocity data was exported as model data. The model data was filtered out of low velocities and the velocity data at different heights in the same frame were averaged to obtain the true value data. The plume velocity inversion result data is also filtered out with low speed and the velocity data at different heights in the same frame are averaged, and then the correlation with the true value data is verified.

8. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 1 is characterized in that: Verify the plume velocity inversion results, including: The pulsed smoke generator and wind speed detector are used to verify the inversion results of the smoke plume velocity, including: A pulsed smoke generator is used in conjunction with a wind speed detector to collect anemometer data. At the same time, the smoke plume velocity inversion results are matched with the pulse interval and time resolution of the pulsed smoke generator. The error and variance between the matched plume velocity inversion result data and the anemometer data are calculated to achieve verification.

9. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 1 is characterized in that: Combining the plume velocity inversion results with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification, including: Before observation, the acquisition angle of the image acquisition camera in the improved dense optical flow method Farneback is ensured to be consistent with the observation angle of the hyperspectral remote sensing system. The spatial resolution of the camera is usually set higher than that of the ground-based hyperspectral remote sensing system to match the spatial resolution. The temporal resolution is also matched between the acquired image and the integration time of the ground-based hyperspectral remote sensing system. On this basis, the plume velocity inversion results are combined with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification.

10. The high-precision emission flux quantification method for plume velocity inversion and verification and fusion imaging according to claim 1 or 9, characterized in that: Combining the plume velocity inversion results with hyperspectral remote sensing imaging to achieve high-precision emission flux quantification, including: Where M is the molecular mass of the component whose flux is required, N A is Avogadro's constant, SCD i,j is the column concentration value of the jth column and the ith row, v i,j is the velocity value in the jth column and the ith row, N is the number of pixels in the jth column that contain plumes, and h i,j It refers to the actual height represented by the height of each pixel in the results of the ground-based hyperspectral remote sensing system.

Images