A wind field measurement method based on leaf motion characteristics

Abstract

The present application relates to a kind of wind field measurement method based on leaf movement characteristics, belong to wind speed measurement technical field, solve the problem that cannot perceive wind speed distribution in scene in prior art. Specifically include: keep camera pixel equivalent consistent and obtain the single-channel video of the leaf shape under different wind speeds, according to time sequence, form image data set;While measuring the wind speed corresponding to each frame image in image data set, obtain wind speed true value;Using mixed Gaussian model obtains the foreground point in each frame image;According to the obtained foreground point, the outline of the region in the foreground point set in each frame image is obtained, and the outline with wind speed true value is obtained;The feature of each outline and the corresponding wind speed true value are used as a sample, and the machine learning model is trained;The wind speed of the image to be predicted is predicted using the trained machine learning model, and the wind speed value corresponding to the image is obtained. It is realized that only optical camera can perceive wind speed distribution, and the calculation is simple and the cost is low.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind speed measurement, and in particular to a wind field measurement method based on leaf movement characteristics. BACKGROUND

[0002] Meteorological conditions have an important influence on human activities such as transportation and security. Wind speed is an important meteorological condition. Traditional wind speed measurement methods mainly include mechanical, ultrasonic, and laser radar. Mechanical anemometers not only have mechanical wear but also can only measure single-point wind speed. Ultrasonic speed measurement has no wear and can measure three-dimensional wind speed vectors, but it is still a single-point wind speed measurement and is easily affected by air temperature and noise. Laser radar wind measurement can obtain wind speed and direction, but it measures the projection of wind speed and is expensive. SUMMARY

[0003] In view of the above analysis, the embodiments of the present application aim to provide a wind field measurement method based on leaf movement characteristics to solve the problem that the existing method cannot perceive the wind speed level distribution in the scene.

[0004] In one aspect, the embodiments of the present application provide a wind field measurement method based on leaf movement characteristics, which specifically includes:

[0005] Maintain consistent camera pixel equivalence and acquire single-channel video of leaf shape under different wind speeds at different times, expand the video according to time sequence to form an image data set;

[0006] Measure the wind speed corresponding to each frame of image in the image data set at the same time as the image is collected to obtain the true value of the wind speed of each frame of image;

[0007] Obtain foreground points in each frame of image using a mixture Gaussian model;

[0008] Obtain the contour of the region in the foreground point set in each frame of image according to the foreground points in each frame of image, and screen the contour according to the true value of the wind speed to obtain a contour with a true value of the wind speed;

[0009] Extract the features of the contour, and use the features of each contour and the true value of the wind speed corresponding to the contour as a sample to train a machine learning model to obtain a trained machine learning model;

[0010] Use the trained machine learning model to predict the wind speed of an image to be predicted to obtain the wind speed value corresponding to the image.

[0011] Optionally, the mixture Gaussian model is used to obtain the foreground points in each frame of image, which includes:

[0012] Each pixel point of each image in the image dataset is matched with the mixed Gaussian model, if the matching is successful, the pixel point is a background point, if the matching is unsuccessful, the pixel point is a foreground point.

[0013] Optionally, the acquiring the foreground points in each image by using the mixed Gaussian model comprises:

[0014] Initializing the mixed Gaussian model parameters;

[0015] Matching the images in the image dataset by using the initialized mixed Gaussian model, and updating the mixed Gaussian model parameters after matching each pixel point, matching the pixel points corresponding to the next image in the image dataset by using the updated mixed Gaussian model, until the matching of all the images in the image dataset is completed, and the foreground points in each image are found.

[0016] Optionally, after the foreground points are found, the pixel value of the foreground points is set to 255, and the pixel value of the background points is set to 0.

[0017] Optionally, the mixed Gaussian model is:

[0018]

[0019] wherein ω i is the weight of the i th Gaussian distribution, is the probability density function of the i th Gaussian distribution, μ i is the mean of the i th Gaussian distribution, is the variance of the i th Gaussian distribution.

[0020] Optionally, the mixed Gaussian model parameters are initialized in the following manner:

[0021] Obtaining the images in a period of time before the image dataset is taken, calculating the mean and variance of each pixel point in the period of time, and initializing the parameters of the mixed Gaussian model based on the mean and variance of each pixel point. Specifically as follows:

[0022]

[0023]

[0024] wherein the image sequence length is N, I m,t represents the gray value of the m th pixel point on the image at t time, μ m represents the mean gray value of the m th pixel point, μ m,k represents the mean gray value of the k th Gaussian distribution of the m th pixel point, represents the variance of the k th Gaussian distribution of the m th pixel point, and K represents the number of Gaussian distributions in each mixed Gaussian model.

[0025] Optionally, the I m,t is updated according to the following formula:

[0026] ω m,k,t = (1 - a) ω m,k,t-1 + a

[0027] μ m,k,t = (1 - β) μ m,k,t-1 + β I m,t

[0028]

[0029]

[0030] wherein I m,t represents the gray value of the mth pixel point at time t; ω m,k,t represents the weight of the kth Gaussian distribution of the mth pixel point at time t; a represents the learning rate; β represents the parameter learning rate; ω m,k,t-1 represents the weight of the kth Gaussian distribution of the mth pixel point at time t-1; represents the variance of the kth Gaussian distribution of the mth pixel point at time t; represents the variance of the kth Gaussian distribution of the mth pixel point at time t-1; μ m,k,t represents the average gray value of the kth Gaussian distribution of the mth pixel point at time t.

[0031] Optionally, the features of the contour include:

[0032] Entropy in the contour region, the number of foreground points in the contour, the average distance between foreground points in the contour, the contour perimeter, the contour area, the contour density, and the distance from the contour center point to the camera.

[0033] Optionally, the wind speed prediction on the image to be predicted is performed by using the trained machine learning model to obtain the wind speed value corresponding to the image, comprising:

[0034] The foreground points in the image to be predicted are extracted by using the mixed Gaussian model, and the contours of the foreground point concentrated region are obtained according to the foreground points. The features of each contour are extracted and input into the trained machine learning model to obtain the predicted wind speed of each contour in the image. The predicted wind speed is the wind speed value corresponding to the image.

[0035] Optionally, the contours of the foreground point concentrated region in each frame of image are obtained according to the foreground points in each frame of image, comprising: performing erosion and expansion processing on the foreground points in each frame of image to obtain the contours of the foreground point concentrated region.

[0036] On the other hand, compared with the prior art, the present application can at least achieve one of the following beneficial effects:

[0037] 1、 The present application uses an optical camera to obtain a scene image, uses image processing technology to obtain different motion states of leaf shapes caused by different wind speed levels, extracts effective features, and uses a trained machine learning model to classify the obtained feature combinations, so that the wind speed level distribution in the scene can be perceived. Compared with the traditional method, the present application can perceive the wind speed level distribution only by using an optical camera, and the calculation is simple and the cost is low.

[0038] The above technical solutions in the present application can also be combined with each other to realize more preferred combination solutions. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification, or will be understood by implementing the present application. The purpose and other advantages of the present application can be realized and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0039] The accompanying drawings are included to provide a further understanding of the present application, and are incorporated herein and constitute a part of the specification. The drawings illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application.

[0040] Figure 1 A flowchart of a wind field measurement method based on leaf motion features in an embodiment of the present application;

[0041] Figure 2 An image obtained by a camera in an embodiment of the present application;

[0042] Figure 3 An image obtained by a camera in an embodiment of the present application;

[0043] Figure 4 An image obtained by a camera in an embodiment of the present application;

[0044] Figure 5 A wind speed level prediction graph in an embodiment of the present application. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present application will be specifically described below with reference to the accompanying drawings, wherein the drawings constitute a part of the present application and are used to illustrate the principles of the embodiments of the present application, and are not used to limit the scope of the present application.

[0046] One specific embodiment of the present application discloses a wind field measurement method based on leaf motion features, as shown in Figure 1

[0047] ​Step S1: keep the camera pixel equivalent consistent and acquire the single-channel video of the leaf shape under different wind speeds at different times, expand the video according to the time sequence to form an image data set;

[0048] Specifically, in order to train the prediction model of the machine learning model, first, the camera is used to shoot the leaf shape under different wind speeds and different distances, a single-channel video of the leaf shape under different wind speeds is established, and the video is expanded according to the time sequence to form an image data set, as shown in Figure 2

[0049] Specifically, in order to make the actual size of each pixel point in the image the same, the pixel equivalent needs to be set. The pixel equivalent represents the actual physical size of a pixel point in the image. When acquiring the image data set, the pixel equivalent needs to be kept consistent. By adjusting the zoom of the camera, the pixel equivalent of each image can be kept consistent. Since the movement of the tree needs to be shot and the leaf shape under different distances needs to be shot, the focal length of the camera needs to be adjusted during image shooting to keep the pixel equivalent of each image consistent. The three-dimensional model of the environment in the current area is obtained by using a drone in the selected area where the leaf shape is collected. According to the process of combining virtual and real in the three-dimensional model, the virtual image in the three-dimensional model is one-to-one corresponding to the real image shot by the camera. Therefore, not only the position and size of the tree can be determined, but also the pixel equivalent consistency can be ensured when acquiring the image each time.

[0050] Step S2: measure the wind speed corresponding to each frame of image in the image data set to obtain the wind speed true value of each frame of image;

[0051] While collecting the image, the wind speed corresponding to each frame of image in the image data set is measured. A plurality of anemometers are used to record the wind speed true value at different positions at the real-time shooting time of the image. The wind speed true value at the shooting time corresponding to each frame of image is recorded.

[0052] It can be understood that the image collection time is the shooting time corresponding to the image in the single-channel video.

[0053] The sample for training the machine learning model needs the true value. The wind speed at different positions at each time of the camera shooting image is recorded by the anemometer.

[0054] Step S3: acquire the foreground point in each frame of image by using the mixed Gaussian model;

[0055] Each pixel point of each frame of image in the image data set is matched with the mixed Gaussian model. If the matching is successful, the pixel point is a background point. If the matching is unsuccessful, the pixel point is a foreground point.

[0056] ​The mixed Gaussian model establishes a Gaussian distribution model for the change in time length of each pixel in each image, that is:

[0057]

[0058] where ω i is the weight of the i-th Gaussian distribution, is the probability density function of the i-th Gaussian distribution, μ i is the mean of the i-th Gaussian distribution, is the variance of the i-th Gaussian distribution, and x is a sequence of images.

[0059] In the mixed Gaussian model, each pixel is represented by K (usually 3 to 5) Gaussian distributions. The probability of a pixel having a value X t at time t is:

[0060]

[0061] where ω i is the weight of the i-th Gaussian distribution at time t, η(X t | θ i ) is the probability density function of the i-th Gaussian distribution at time t, and can be expressed as:

[0062]

[0063] where μ i is the mean and σ i is the standard deviation of the i-th Gaussian distribution. After the mixed Gaussian model is established, it can be used to match whether each pixel in each image is background or not. If it does not conform to the Gaussian distribution, the point is a foreground point. After obtaining the foreground point, the foreground point pixel value is set to 255 and the background point is set to 0.

[0064] In this application, to obtain the foreground points in each image, the method of matching with the mixed Gaussian model is used, that is, all pixel points in each image in the image data set are matched with the mixed Gaussian model of the corresponding pixel points in the previous image. If the matching is successful, the pixel point is determined to be a background point, otherwise it is a foreground point. After matching each pixel point, the mixed Gaussian model is updated. This method can well remove the background, and when the wind speed is greater, the amplitude of the leaf swing is greater, and the foreground points generated are more. The size of the foreground points in the image obtained by the Gaussian mixed model and the true value of the wind speed have a direct relationship. As shown in Figure 3 .

[0065] S31: Initialize the mixed Gaussian model parameters:

[0066] In order to initialize the parameters in the M Gaussian mixture models, the images in a period of time before the image dataset is taken are obtained, the average gray value and variance of each pixel point in the images in the period of time are calculated, and the parameters of the Gaussian mixture model are initialized based on the average value and variance of each pixel point, as follows:

[0067]

[0068]

[0069] wherein, the image sequence length is N, I m,t represents the gray value of the mth pixel point on the image at t time, μ m,k represents the average value of the kth Gaussian distribution of the gray value of the mth pixel point, σ 2 m,k represents the variance of the kth Gaussian distribution of the gray value of the mth pixel point, and K represents the number of Gaussian distributions in each Gaussian mixture model.

[0070] It can be understood that the parameters of the Gaussian mixture model corresponding to each pixel point in the image are obtained according to the average value and variance of the corresponding pixel point in the N images, and therefore the parameters of the Gaussian mixture model corresponding to each pixel point can be different.

[0071] S32: The images in the image dataset are matched by using the Gaussian mixture model initialized with the parameters, and after each pixel point is matched, the Gaussian mixture model parameters are updated, the pixel point corresponding to the next image in the image dataset is matched by using the updated Gaussian mixture model, and the matching of all images in the image dataset is completed until the foreground points in each image are found.

[0072] The first image in the image dataset is matched by using the Gaussian mixture model initialized with the parameters, and each pixel point in the first image in the image dataset is matched with the K Gaussian distributions in the Gaussian mixture model of the corresponding pixel point obtained in step S31, and if the pixel point satisfies the following formula with any Gaussian distribution model, it is considered that the pixel point is matched with the Gaussian mixture model; wherein, I m,t represents the gray value of the mth pixel point on the image at t time, μ m,k,t-1 and σ m,k,t-1 represents the average gray value and variance of the kth Gaussian distribution of the mth pixel point at t-1 time, and λ is a self-defined parameter, which is generally taken as 2.5 in practice.

[0073] |I m,t -μ m,k,t-1 |≤λσ m,k,t-1

[0074] After all the pixels of the first image in the image set are matched with the K Gaussian distributions in the corresponding Gaussian mixture model, the foreground pixels in the image are obtained;

[0075] Then, the Gaussian distribution matched with I m,t is updated according to the following formula:

[0076] ω m,k,t = (1 - a) ω m,k,t-1 + a I

[0077] μ m,k,t = (1 - β) μ m,k,t-1 + β I m,t

[0078]

[0079]

[0080] wherein I m,t represents the gray value of the mth pixel point at time t; ω m,k,t represents the weight of the kth Gaussian distribution of the mth pixel point at time t; a represents the learning rate; β represents the parameter learning rate; ω m,k,t-1 represents the weight of the kth Gaussian distribution of the mth pixel point at time t-1; represents the variance of the kth Gaussian distribution of the mth pixel point at time t; represents the variance of the kth Gaussian distribution of the mth pixel point at time t-1; μ m,k,t represents the average gray value of the kth Gaussian distribution of the mth pixel point at time t.

[0081] The updated Gaussian mixture model obtained after each pixel point is matched is used to continue matching the corresponding pixel points in the next image in the image set, and the foreground pixels in the next image are found out; the Gaussian mixture model is updated after each pixel point is matched, and the foreground pixels in the next image are obtained by using the updated Gaussian mixture model, and the process is repeated until the foreground pixels of all images in the image set are found out.

[0082] It can be understood that in the matching process, only one Gaussian mixture model will be matched with each pixel point, and when a Gaussian mixture model is successfully matched, the remaining Gaussian mixture models that are not matched will not be matched, and the sum of the weights of the K Gaussian distributions will be kept as 1, therefore, the remaining Gaussian mixture models keep the original mean value and standard deviation, and the weight will decay.

[0083] If a pixel point does not match with all the updated Gaussian distribution models of the corresponding pixel point in the previous image, it is considered that the pixel point does not match with the Gaussian mixture model. The Gaussian mixture model is updated in the following manner:

[0084] The Gaussian distribution with the minimum weight is updated to a newly established Gaussian distribution, and the remaining Gaussian distributions with higher weights maintain the original mean value and standard deviation, but their weights are decayed, and the decayed weights are:

[0085] ω k,t =(1-α)ω k,t-1

[0086] wherein a is a learning rate.

[0087] The newly established Gaussian distribution is obtained in the following manner:

[0088] The gray value of the pixel point at the current moment is set as the mean gray value of the new Gaussian distribution, and a larger standard deviation and a smaller weight are initialized, wherein the larger standard deviation is an empirical value, and the sum of the smaller weight and the weights of the remaining Gaussian distributions after decay is 1.

[0089] Step S4: obtaining the contour of the region with concentrated foreground points in each frame of image according to the foreground points in each frame of image, and screening the contour according to the wind speed true value to obtain the contour with the wind speed true value.

[0090] Specifically, the contour of the region with concentrated foreground points is obtained by performing erosion and expansion processing on the foreground points in each frame of image.

[0091] When measuring the wind speed true value, the anemometer is used to measure at multiple different positions, and the anemometer corresponding to the position of each contour in the image is found. If the anemometer corresponding to the contour cannot be found, it indicates that the anemometer is not set at the actual position of the contour, and the contour is discarded. Finally, the remaining contours all have the wind speed true value measured by the corresponding anemometer.

[0092] S41: extracting the features of the contour.

[0093] Specifically, for each contour in the image obtained in the previous step, seven features corresponding to each contour are extracted, i.e., the entropy in the contour region, the number of foreground points in the contour, the average distance between the foreground points in the contour, the contour perimeter, the contour area, the contour density, and the distance from the center point of the contour to the camera. The above seven features are used to describe the different morphologies of the foreground points in each contour. As shown in the following table. Figure 4

[0094] 1. Information entropy:

[0095]

[0096] wherein P i represents the distribution probability of the image at the i gray level. ​

[0097] 2. Number of foreground points

[0098] The number of foreground points in the bounding rectangle of the contour is calculated using the following formula:

[0099] In the calculation of the number of foreground points, a bounding rectangle is selected for the contour, and the number of foreground points can be obtained from the foreground point image obtained by the Gaussian mixture model:

[0100]

[0101]

[0102] where i represents the row, j represents the column, f(i,j) represents the pixel value of the image, H and W represent the width and height of the bounding rectangle, respectively.

[0103] 3. Average distance between foreground points

[0104] The distance between any two foreground points in the contour is calculated, and the average distance between all pairs of foreground points is taken as the average distance between foreground points:

[0105]

[0106] dis(x m ,x n )=|i m -i n |+|j m -j n |

[0107] where x m ,x n represent the mth foreground point and the nth foreground point, respectively, i m ,j m represent the row and column of the foreground point x m ; i n ,j n represent the row and column of the foreground point x n .

[0108] 4. Contour perimeter

[0109]

[0110]

[0111] where x m represents the mth foreground point.

[0112] 5. Contour area

[0113]

[0114]

[0115] where i, j represent image row and column respectively.

[0116] 6. Contour density

[0117]

[0118] where L represents the number of foreground points in the contour, and A represents the total number of pixel points in the contour.

[0119] 7. Camera shooting distance

[0120] The shooting distance represents the distance from the object of the contour center point to the camera.

[0121] Specifically, the distance from the object of the contour center point to the camera can be calculated according to the three-dimensional model.

[0122] Step S5: Extracting the features of the contour, taking the features of each contour and the wind speed true value corresponding to the contour as a sample, and training the machine learning model to obtain a trained machine learning model;

[0123] The machine learning model is trained by the calculated features and the wind speed true value. The features of each contour extracted in each image and the wind speed true value corresponding to the contour are taken as a sample, and the machine learning model is trained. Through the machine learning model, the relationship between the wind speed features and the true value is established, so that the wind speed is obtained according to the extracted features of the obtained image. For a newly acquired image, find the foreground points in the image, and obtain multiple contours by eroding and expanding the foreground points. Calculate the 7 features of the contour, input the 7 features into the trained machine learning model, and obtain the wind speed grade corresponding to the current contour.

[0124] Step S6: Using the trained machine learning model to predict the wind speed of the image to be predicted, and obtaining the wind speed value corresponding to the image;

[0125] The foreground points in the image to be predicted are extracted by using the mixed Gaussian model, and the contours of the foreground point cluster region are obtained according to the foreground points. The features of each contour are extracted and input into the trained machine learning model to obtain the predicted wind speed of each contour in the image. The predicted wind speed is the wind speed value corresponding to the image.

[0126] Specifically, as shown in Figure 5 is the result of the discrimination. In the broken line graph, the horizontal coordinate is time, the unit is minute, and the vertical coordinate is wind speed, the unit is grade. The difference between the wind speed grade prediction and the actual anemometer prediction is basically within one level.

[0127] The present application acquires images by using an optical camera, acquires different motion states of leaf shapes caused by different wind speeds by using image processing technology, extracts effective features, and classifies the obtained feature combinations by using a trained machine learning model, so as to measure the wind speed grade distribution in a scene. Based on the proposed leaf shape feature combination and the trained machine learning model, the present application can realize real-time sensing of the wind speed grade distribution at different positions in a scene. In the embodiment of the present application, the machine learning model includes a support vector machine model, a random forest, a decision tree, and the like.

[0128] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. The computer readable storage medium includes a magnetic disk, an optical disk, a read-only memory, a random access memory, and the like.

[0129] The above description is only a preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any changes or replacements within the technical scope disclosed by the present application can be easily thought of by those skilled in the art, and should be covered within the protection scope of the present application.

Claims

1. A wind field measurement method based on leaf motion characteristics, characterized by, Specifically comprising: Keeping the camera pixel equivalent consistent and obtaining single-channel videos of leaf shapes under different wind speeds at different times, expanding the videos according to the time sequence to form an image dataset; Measuring the wind speed corresponding to each frame of image in the image dataset while collecting the image to obtain the wind speed true value of each frame of image; Obtaining foreground points in each frame of image by using a Gaussian mixture model; Obtaining the contour of the region in the foreground point set in each frame of image according to the foreground points in the frame of image, and screening the contour according to the wind speed true value to obtain a contour with a wind speed true value; Extracting the features of the contour, taking the features of each contour and the wind speed true value corresponding to the contour as a sample, and training a machine learning model to obtain a trained machine learning model; Using the trained machine learning model to predict the wind speed of an image to be predicted and obtaining the wind speed value corresponding to the image.

2. The wind field measurement method based on tree leaf motion characteristics according to claim 1, characterized in that, The method comprises the following steps: Matching each pixel point of each frame of image in the image dataset with the Gaussian mixture model, and if the matching is successful, the pixel point is a background point, and if the matching is unsuccessful, the pixel point is a foreground point.

3. The wind field measurement method based on tree leaf motion characteristics according to claim 2, characterized in that, The method comprises the following steps: Initializing the parameters of the Gaussian mixture model; Matching the images in the image dataset with the Gaussian mixture model after the parameter initialization, updating the parameters of the Gaussian mixture model after matching each pixel point, matching the pixel points corresponding to the next image in the image dataset with the updated Gaussian mixture model, and finding the foreground points in each image until the matching of all images in the image dataset is completed.

4. The wind field measurement method based on tree leaf motion characteristics according to claim 3, characterized in that, After finding the foreground points, setting the pixel value of the foreground points to 255 and setting the pixel value of the background points to 0.

5. The wind field measurement method based on tree leaf motion characteristics according to claim 3, characterized in that, The Gaussian mixture model comprises: where ω i is the weight of the i-th Gaussian distribution, is the probability density function of the i-th Gaussian distribution, μ i is the mean of the i-th Gaussian distribution, is the variance of the i-th Gaussian distribution, and x is a sequence of images.

6. The wind field measurement method based on tree leaf motion characteristics according to claim 3, characterized in that, The parameters of the Gaussian mixture model are initialized in the following manner: Obtaining images within a period of time before the image dataset is taken, calculating the average value and variance of each pixel point within the period of time, and initializing the parameters of the Gaussian mixture model based on the average value and variance of each pixel point, specifically as follows: where the image sequence length is N, I m,t denotes the gray value of the mth pixel point on the image at time t, μ m denotes the average gray value of the mth pixel point, μ m,k denotes the average gray value of the kth Gaussian distribution of the mth pixel point, denotes the variance of the kth Gaussian distribution of the mth pixel point, and K denotes the number of Gaussian distributions in each mixed Gaussian model.

7. The wind field measurement method based on tree leaf motion characteristics according to claim 3, characterized in that, With I m,t The Gaussian distribution that matches I is updated according to: ω m,k,t = (1 - a) ω m,k,t-1 + a μ m,k,t = (1 - β)μ m,k,t-1 + βI m,t wherein I m,t denotes the gray value of the mth pixel point at time t; ω m,k,t denotes the weight of the kth Gaussian distribution of the mth pixel point at time t; α denotes the learning rate; β denotes the parameter learning rate; ω m,k,t-1 denotes the weight of the kth Gaussian distribution of the mth pixel point at time t-1; denotes the variance of the kth Gaussian distribution of the mth pixel point at time t; denotes the variance of the kth Gaussian distribution of the mth pixel point at time t-1; μ m,k,t denotes the average gray value of the kth Gaussian distribution of the mth pixel point at time t.

8. The wind field measurement method based on tree leaf motion characteristics according to claim 1, characterized in that, The features of the contour comprise: The entropy in the contour region, the number of foreground points in the contour, the average distance between the foreground points in the contour, the contour length, the contour area, the contour density, and the distance between the contour center point and the camera.

9. The wind field measurement method based on tree leaf motion characteristics according to claim 1, characterized in that, The method comprises the following steps: Extracting the foreground points in the image to be predicted by using the Gaussian mixture model, obtaining the contour of the region in the foreground point set according to the foreground points, extracting the features of each contour and inputting the features into the trained machine learning model to obtain the predicted wind speed of each contour in the image, and the predicted wind speed is the wind speed value corresponding to the image.

10. The wind field measurement method based on tree motion characteristics according to any one of claims 1-3, characterized in that, The method of obtaining the contour of the region in the foreground point set according to the foreground points in each frame of image comprises: performing erosion and expansion processing on the foreground points in each frame of image to obtain the contour of the region in the foreground point set.

Images