A storm motion vector estimation method, system and computer storage medium

Abstract

The present invention provides a storm motion vector estimation method, system and computer storage medium, belonging to the field of weather radar monitoring technology. The estimation method can estimate the 3D storm motion field vector based on a deep learning network, with high robustness and accuracy, and the output motion field estimation result has high resolution. The estimation system has the function of implementing the above-mentioned storm motion vector estimation method. The computer storage medium stores computer instructions, which can implement the above-mentioned storm motion vector estimation method when executed by a processor.

Description

Technical Field

[0001] The present invention relates to the technical field of weather radar monitoring, and in particular to a storm motion vector estimation method, system and computer storage medium. Background Art

[0002] Doppler weather radar volume scan data is one of the important data used in various types of weather forecasts, especially in monitoring and warning of severe convective weather. In many prediction methods for severe convective and thunderstorm weather, volume scan data at different volume scan time nodes are often used as input data. The strong echo area in the input data is identified and extrapolated to obtain the warning result. The storm motion vector field is a commonly used intermediate product of this type of method.

[0003] In related technologies, there are two main methods to obtain storm motion information. One is to use the correlation coefficient method or the optical flow method to calculate the motion vector of each pixel in the radar image by inputting two consecutive 2D image frames. The other is to first identify the storm clouds in the radar image, and then perform similarity matching on the clouds in the previous and next frames, and connect the center points of the most suitable matching results to obtain the motion vector of the clouds.

[0004] With the improvement of radar resolution and the development of applications, such 2D motion vector estimation results have gradually failed to meet the needs of applications. It is impossible to directly obtain 3D storm motion field information from the volume scan data obtained by weather radar, and it has the disadvantages of low accuracy and resolution. Summary of the invention

[0005] The embodiments of the present invention provide a storm motion vector estimation method, system and computer storage medium, which can estimate 3D storm motion field vectors based on a deep learning network, have high robustness and accuracy, and output motion field estimation results with high resolution. The technical solution is as follows:

[0006] In a first aspect, an embodiment of the present invention provides a storm motion vector estimation method, comprising:

[0007] Acquire historical volume scan data of multiple different target areas, the historical volume scan data including position data and time data, the position data including first position data of a detected pixel point at the beginning of the volume scan, and the time data including a first time node corresponding to the first position data;

[0008] Generate a three-dimensional grid that matches the target area, and generate a random virtual motion field on the three-dimensional grid using a Perlin noise method so that each grid point in the three-dimensional grid has a corresponding motion vector;

[0009] Performing three-dimensional grid interpolation on the historical volume scan data based on the three-dimensional grid to obtain first point cloud data of a grid point corresponding to the first position data in the three-dimensional grid, then extrapolating the detection pixel point corresponding to the first position data based on the motion vector to obtain second position data corresponding to the detection pixel point at a second time node after a preset time, and performing three-dimensional grid interpolation based on the three-dimensional grid to obtain second point cloud data of a grid point corresponding to the second position data in the three-dimensional grid, recording the first point cloud data and the second point cloud data as a data pair, and collecting the data pairs corresponding to the historical volume scan data of the multiple different target areas to form a training data set;

[0010] Establishing a point cloud motion field estimation network model based on the FlowNet3D network structure, and inputting the training data set for training;

[0011] Acquire real-time volume scan data of the target area, the real-time volume scan data including third position data of the detection pixel point at a third time node, and fourth position data of the detection pixel point at a fourth time node, the third time node and the fourth time node being two time nodes separated by the preset time;

[0012] Based on the three-dimensional grid, three-dimensional grid interpolation is performed on the real-time volume scan data and the trained point cloud motion field estimation network model is input to output the estimated motion vector of the detection pixel point between the third time node and the fourth time node.

[0013] Optionally, the preset time interval ranges from 5 to 6 minutes.

[0014] Optionally, the three-dimensional grid interpolation is performed on the historical volume scan data based on the three-dimensional grid to obtain first point cloud data of a grid point corresponding to the first position data in the three-dimensional grid, and then the detection pixel point corresponding to the first position data is extrapolated based on the motion vector to obtain second position data corresponding to the detection pixel point at a second time node after a preset time, and the three-dimensional grid interpolation is performed based on the three-dimensional grid to obtain second point cloud data of a grid point corresponding to the second position data in the three-dimensional grid, the first point cloud data and the second point cloud data are recorded as a data pair, and the data pairs corresponding to the historical volume scan data of the multiple different target areas are collected to form a training data set, and further includes:

[0015] Set the echo strength threshold;

[0016] Filtering and resampling the first point cloud data based on the echo intensity threshold to obtain first high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold, and filtering and resampling the second point cloud data based on the echo intensity threshold to obtain second high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold;

[0017] The first high echo intensity point cloud data and the second high echo intensity point cloud data are recorded as the data pair, and the data pairs corresponding to the historical volume scan data of the multiple different target areas are collected to form the training data set.

[0018] Optionally, three-dimensional grid interpolation is performed on the real-time volume scan data based on the three-dimensional grid to obtain third point cloud data of grid points corresponding to the third position data in the three-dimensional grid, and fourth point cloud data of grid points corresponding to the fourth position data in the three-dimensional grid;

[0019] filtering and resampling the third point cloud data based on the echo intensity threshold to obtain third point cloud data with the highest echo intensity greater than the echo intensity threshold, and filtering and resampling the fourth point cloud data based on the echo intensity threshold to obtain fourth point cloud data with the highest echo intensity greater than the echo intensity threshold;

[0020] The third high echo intensity point cloud data and the fourth high echo intensity point cloud data are input into the trained point cloud motion field estimation network model, and the estimated motion vector of the detection pixel point between the third time node and the fourth time node is output.

[0021] Optionally, before establishing the point cloud motion field estimation network model based on the FlowNet3D network structure and inputting the training data set for training, the Storm motion vector estimation method further includes:

[0022] Obtain first radar reflectivity data corresponding to the detection pixel point at the first time node, and second radar reflectivity data corresponding to the detection pixel point at the second time node, and add the first radar reflectivity data and the second radar reflectivity data to the data pair for recording.

[0023] In a second aspect, the present invention provides an estimation system, comprising:

[0024] A first data acquisition module is used to acquire historical volume scan data of multiple different target areas, wherein the historical volume scan data includes position data and time data, wherein the position data includes first position data of a detected pixel point at the beginning of the volume scan, and the time data includes a first time node corresponding to the first position data;

[0025] A virtual sports field generation module is used to generate a three-dimensional grid that matches the target area, and generate a random virtual sports field on the three-dimensional grid using a Perlin noise method, so that each grid point in the three-dimensional grid has a corresponding motion vector;

[0026] a data set establishment module, configured to perform three-dimensional grid interpolation on the historical volume scan data based on the three-dimensional grid, obtain first point cloud data of a grid point corresponding to the first position data in the three-dimensional grid, then extrapolate the detection pixel point corresponding to the first position data based on the motion vector, obtain second position data corresponding to the detection pixel point at a second time node after a preset time, perform three-dimensional grid interpolation based on the three-dimensional grid, so as to obtain second point cloud data of a grid point corresponding to the second position data in the three-dimensional grid, record the first point cloud data and the second point cloud data as a data pair, and collect the data pairs corresponding to the historical volume scan data of the multiple different target areas to form a training data set;

[0027] A network model building module is used to build a point cloud motion field estimation network model based on the FlowNet3D network structure, and input the training data set for training;

[0028] a second data acquisition module, configured to acquire real-time volume scan data of the target area, wherein the real-time volume scan data includes third position data of the detection pixel point at a third time node and fourth position data of the detection pixel point at a fourth time node, wherein the third time node and the fourth time node are two time nodes spaced apart from each other;

[0029] A motion vector estimation module is used to perform three-dimensional grid interpolation on the real-time volume scan data based on the three-dimensional grid and input the trained point cloud motion field estimation network model, and output the estimated motion vector of the detection pixel point between the third time node and the fourth time node.

[0030] Optionally, the data set establishment module is further used to:

[0031] Set the echo strength threshold;

[0032] Filtering and resampling the first point cloud data based on the echo intensity threshold to obtain first high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold, and filtering and resampling the second point cloud data based on the echo intensity threshold to obtain second high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold;

[0033] The first high echo intensity point cloud data and the second high echo intensity point cloud data are recorded as the data pair, and the data pairs corresponding to the historical volume scan data of the multiple different target areas are collected to form the training data set.

[0034] Optionally, the data set establishment module is also used to obtain first radar reflectivity data corresponding to the detection pixel point at the first time node, and second radar reflectivity data corresponding to the detection pixel point at the second time node, and add the first radar reflectivity data and the second radar reflectivity data to the data pair for recording.

[0035] In a third aspect, the present invention further provides an estimation system, the estimation system comprising:

[0036] processor;

[0037] a memory for storing instructions executable by a processor;

[0038] Wherein, the processor is configured to execute the storm motion vector estimation method described in the first aspect above.

[0039] In a fourth aspect, the present invention further provides a computer storage medium having computer instructions stored thereon, wherein the computer instructions, when executed by a processor, implement the storm motion vector estimation method described in the first aspect.

[0040] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0041] First, historical volume scan data of multiple different target areas are obtained from the volume scan basic data recorded by the weather radar. The historical volume scan data can be the volume scan basic data recorded by different regions and different weather radars. Then, a three-dimensional grid matching the weather radar scanning range of the target area is generated by computer software, and a random virtual sports field is generated on the three-dimensional grid using the Perlin noise method, so that each grid point in the three-dimensional grid has a corresponding motion vector. Then, three-dimensional grid interpolation is performed on the historical volume scan data based on the three-dimensional grid to obtain the first point cloud data of the first position data corresponding to the grid point in the three-dimensional grid. Then, based on the motion vector of the grid point, the detection pixel point corresponding to the first position data is extrapolated to obtain the second position data corresponding to the detection pixel point at the second time node after a preset time. The second position data is also interpolated by three-dimensional grid to obtain the second point cloud data of the second position data corresponding to the grid point in the three-dimensional grid, and the first point cloud data and the second point cloud data are recorded as a set of data pairs. Then, the above operation is repeatedly performed on the historical volume scan data of multiple different target areas, and the data pairs corresponding to the historical volume scan data of multiple different target areas are collected to form a training data set.

[0042] Then use the training data set to train the point cloud motion field estimation network model established based on the FlowNet3D network structure. When the storm motion vector needs to be estimated later, the real-time scanning data in the target area is obtained, that is, the third position data and the fourth position data corresponding to the third time node and the fourth time node of the detection pixel point at the interval of the aforementioned preset time. Based on the three-dimensional grid, the real-time volume scanning data is interpolated into a three-dimensional grid and input into the trained point cloud motion field estimation network model, and the estimated motion vector of the detection pixel point between the third time node and the fourth time node can be output.

[0043] This storm motion vector estimation method uses a virtual motion field to estimate the motion trajectory of the detected pixel points and generate a virtual training data set, overcoming the difficulty that the basic scanning data of the weather radar does not include motion field information, and realizes the estimation of 3D storm motion field vectors based on a deep learning network. Compared with the traditional 2D motion vector estimation method, it has high robustness and accuracy, and the output motion field estimation result has high resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. 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.

[0045] Figure 1 is a flow chart of a storm motion vector estimation method provided by an embodiment of the present invention;

[0046] Figure 2 is a flowchart of step S13 provided by an embodiment of the present invention;

[0047] Figure 3 It is a display image of weather radar volume scan data provided by an embodiment of the present invention;

[0048] Figure 4 yes Figure 3 The point cloud data after interpolation of the mid-body scan data is displayed as an image;

[0049] Figure 5 yes Figure 4 Display image after filtering and resampling of point cloud data;

[0050] Figure 6 is a structural schematic diagram of a point cloud motion field estimation network model provided by an embodiment of the present invention;

[0051] Figure 7 is a flowchart of step S16 provided by an embodiment of the present invention;

[0052] Figure 8 is a structural schematic diagram of an estimation system provided by an embodiment of the present invention;

[0053] Fig. 9 It is a schematic diagram of a control structure of another estimation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0055] Figure 1 FIG. 1 is a flow chart of a method for estimating a storm motion vector provided by an embodiment of the present invention. Figure 1 As shown, through practice, the inventor provides a storm motion vector estimation method, system and computer storage medium, the storm motion vector estimation method comprising:

[0056] S11, acquiring historical volume scanning data of multiple different target areas, the historical volume scanning data including position data and time data, the position data including first position data of the detected pixel point when the volume scanning starts, and the time data including a first time node corresponding to the first position data.

[0057] S12, generating a three-dimensional grid matching the target area, and generating a random virtual motion field on the three-dimensional grid using the Perlin noise method, so that each grid point in the three-dimensional grid has a corresponding motion vector.

[0058] S13, performing three-dimensional grid interpolation on the historical volume scan data based on the three-dimensional grid to obtain first point cloud data of a grid point corresponding to the first position data in the three-dimensional grid, then extrapolating the detection pixel point corresponding to the first position data based on the motion vector, obtaining second position data corresponding to the detection pixel point at a second time node after a preset time, and performing three-dimensional grid interpolation based on the three-dimensional grid to obtain second point cloud data of a grid point corresponding to the second position data in the three-dimensional grid, recording the first point cloud data and the second point cloud data as a data pair, and collecting the data pairs corresponding to the historical volume scan data of multiple different target areas to form a training data set.

[0059] S14, establishing a point cloud motion field estimation network model based on the FlowNet3D network structure, and inputting the training data set for training.

[0060] S15, acquiring real-time volume scanning data of the target area, the real-time volume scanning data including third position data of the detection pixel point located at a third time node, and fourth position data of the detection pixel point located at a fourth time node, the third time node and the fourth time node are two time nodes separated by a preset time.

[0061] S16, performing three-dimensional grid interpolation on the real-time volume scan data based on the three-dimensional grid and inputting the trained point cloud motion field estimation network model, and outputting the estimated motion vector of the detection pixel point between the third time node and the fourth time node.

[0062] The storm motion vector estimation method provided by the embodiment of the present invention first obtains historical volume scan data of multiple different target areas from the volume scan basic data recorded by the weather radar. The historical volume scan data can be the volume scan basic data recorded by different regions and different weather radars. Then, a three-dimensional grid matching the weather radar scanning range of the target area is generated by computer software, and a random virtual motion field is generated on the three-dimensional grid using the Perlin noise method, so that each grid point in the three-dimensional grid has a corresponding motion vector. Then, three-dimensional grid interpolation is performed on the historical volume scan data based on the three-dimensional grid to obtain first point cloud data of the grid point corresponding to the first position data in the three-dimensional grid. Then, based on the motion vector of the grid point, the detection pixel point corresponding to the first position data is extrapolated to obtain the second position data corresponding to the detection pixel point at the second time node after a preset time. The second position data is also interpolated by three-dimensional grid to obtain second point cloud data of the grid point corresponding to the second position data in the three-dimensional grid, and the first point cloud data and the second point cloud data are recorded as a set of data pairs. Afterwards, the above operation is repeated for the historical volume scan data of multiple different target areas, and the data pairs corresponding to the historical volume scan data of multiple different target areas are collected to form a training data set. Then, the point cloud motion field estimation network model established based on the FlowNet3D network structure is trained with the training data set. When it is necessary to estimate the storm motion vector later, the real-time scan data in the target area is obtained, that is, the third position data and the fourth position data corresponding to the third time node and the fourth time node of the detection pixel point at the interval of the aforementioned preset time are obtained. Based on the three-dimensional grid, the real-time volume scan data is interpolated into a three-dimensional grid and input into the trained point cloud motion field estimation network model, and the estimated motion vector of the detection pixel point between the third time node and the fourth time node can be output.

[0063] This storm motion vector estimation method uses a virtual motion field to estimate the motion trajectory of the detected pixel points and generate a virtual training data set, overcoming the difficulty that the basic scanning data of the weather radar does not include motion field information, and realizes the estimation of 3D storm motion field vectors based on a deep learning network. Compared with the traditional 2D motion vector estimation method, it has high robustness and accuracy, and the output motion field estimation result has high resolution.

[0064] Optionally, the interval range of the preset time is 5 to 6 minutes. Exemplarily, due to the volume scanning method of the weather radar, a certain amount of time is required for a circular volume scan at each elevation angle. There is a time interval between the data at different elevation angles and azimuth angles. In each volume scanning data, the data at different azimuth angles and elevation angles are also recorded as the scanning results of the time node at the beginning of the volume scanning. The time required for a sweep at each elevation angle is usually 5 to 6 minutes. In an embodiment of the present invention, by setting the interval range of the preset time to 5 to 6 minutes, it is matched with the time taken for a weather radar to sweep once. The point cloud motion field estimation network model can be trained based on the data of the time interval to form a training data set. At the same time, when predicting and estimating the storm motion vector, the real-time volume scan data of two consecutive volume scans can be directly obtained from the volume scan basic data recorded by the weather radar, and can be directly input into the point cloud motion field estimation network model after three-dimensional grid interpolation. Data acquisition is convenient, no additional volume scan data correction is required, and the accuracy is higher.

[0065] Figure 2 It is a flowchart of step S13 provided by an embodiment of the present invention. Figure 3 It is a display image of weather radar volume scan data provided by an embodiment of the present invention. Figure 4 yes Figure 3 The point cloud data after interpolation of the mid-body scan data is displayed as an image. Figure 5 yes Figure 4 The image is displayed after the point cloud data is filtered and resampled. Figures 2 to 5 As shown, step S13 may specifically include the following steps:

[0066] S131, performing three-dimensional grid interpolation on the historical volume scan data based on the three-dimensional grid to obtain first point cloud data of a grid point corresponding to the first position data in the three-dimensional grid, then extrapolating the detection pixel point corresponding to the first position data based on the motion vector to obtain second position data corresponding to the detection pixel point at a second time node after a preset time, and performing three-dimensional grid interpolation based on the three-dimensional grid to obtain second point cloud data of a grid point corresponding to the second position data in the three-dimensional grid.

[0067] It should be noted that this step is the same as the operation in the aforementioned S13 and will not be described in detail here.

[0068] S132, setting the echo intensity threshold.

[0069] Exemplarily, in the embodiment of the present invention, the echo intensity threshold ranges from 35 dBZ to 40 dBZ.

[0070] S133, filtering and resampling the first point cloud data based on the echo intensity threshold to obtain first high echo intensity point cloud data with echo intensity greater than the echo intensity threshold, filtering and resampling the second point cloud data based on the echo intensity threshold to obtain second high echo intensity point cloud data with echo intensity greater than the echo intensity threshold.

[0071] For example, in an embodiment of the present invention, since the resolution of the radar volume scan data in the elevation direction is relatively low, the original point clouds corresponding to the first point cloud data and the second point cloud data are distributed in sheets. A thundercloud may have three pieces scanned in the previous frame (in an embodiment of the present invention, one frame represents a volume scan), but only two pieces are scanned in the next frame. This difference caused by the resolution makes it difficult to train the subsequent point cloud motion field estimation network model, or the learned features are inaccurate. By filtering and resampling the first point cloud data and the second point cloud data after three-dimensional interpolation based on the echo intensity threshold, and selecting the first high echo intensity point cloud data and the second high echo intensity point cloud data with higher echo intensity as data pairs for recording, the corresponding point clouds can be distributed in clusters in both frames, which is conducive to making the point cloud motion field estimation network model focus on learning the displacement features required by the present invention and improving the estimation accuracy.

[0072] S134, recording the first high echo intensity point cloud data and the second high echo intensity point cloud data as a data pair, and collecting the data pairs corresponding to the historical volume scan data of multiple different target areas to form a training data set.

[0073] It should be noted that this step is the same as the operation in the aforementioned S13, except that the first point cloud data and the second point cloud data are replaced with the first high echo intensity point cloud data and the second high echo intensity point cloud data that have been filtered and resampled, which will not be elaborated here.

[0074] Figure 6 is a schematic diagram of the structure of a point cloud motion field estimation network model provided by an embodiment of the present invention, such as Figure 6 Optionally, before establishing a point cloud motion field estimation network model based on the FlowNet3D network structure and inputting a training data set for training, the Storm motion vector estimation method further includes:

[0075] S135, obtaining first radar reflectivity data corresponding to the detection pixel point at the first time node, and second radar reflectivity data corresponding to the detection pixel point at the second time node, and adding the first radar reflectivity data and the second radar reflectivity data to a data pair for recording.

[0076] Exemplarily, in an embodiment of the present invention, the point cloud motion field estimation network model is based on the FlowNet 3D network structure, which can be generally divided into three layers, including a set conv layer for extracting point cloud features in a single frame of volume scan data, a flow embedding layer for mixing the features of two frames of volume scan data and learning motion field information, and a set upconv layer for converting feature information into a vector field output. Since the original network is for LiDAR, that is, laser radar data, for weather radar, by adding the first radar reflectivity data corresponding to the detection pixel point at the first time node and the second radar reflectivity data corresponding to the detection pixel point at the second time node, together with the first position data and the second position data, to the data pair input into the point cloud motion field estimation network model, the data dimension of the training data set is increased, so that the training of the point cloud motion field estimation network model is more adapted to the volume scan data of the weather radar. This further improves the accuracy of subsequent storm motion vector estimation.

[0077] For example, in the embodiment of the present invention, based on step S135, in order to further improve the training completion of the point cloud motion field estimation network model, the following adjustments and optimizations may be performed:

[0078] (1) Since the observation size of weather radar is large and the motion scale of radar point cloud data is at the kilometer level, the "search radius" of the point cloud network needs to match the motion scale of weather radar data, otherwise it may not be possible to extract useful information, thereby reducing the accuracy. The network hyperparameters are adjusted as shown in Table 1:

[0079] Point Cloud Network Hyperparameters

[0080]

[0081]

[0082] Table 1

[0083] (2) Redesign the loss function using the following formula:

[0084]

[0085]

[0086]

[0087] L=L1+λ cos L cos +λ d L d (4)

[0088] Among them, L1 constrains the distance from the predicted point to the true value, L cos Constrain the angle between the predicted vector and the true vector, L d Constrain the smoothness of the motion vector, n1 is the number of points in the previous frame of volume scan data, v′ i is the predicted motion vector value, v i is the true value, r i is the number of vectors within the search radius of the adjacent area, S i Indicates the discrete degree of a prediction vector and its adjacent prediction vectors, λ cos and λ d For L cos and L d The weight parameter of . L is the final loss function.

[0089] Lcos is used to increase the angle constraint between the estimated motion vector and the actual motion angle, thereby improving the accuracy of the estimation result; Ld adds a smoothness constraint to the motion vector field, which can make the output result smoother.

[0090] Figure 7 is a flow chart of step S16 provided by an embodiment of the present invention. Figure 4 As shown, step S16 may specifically include the following steps:

[0091] S161, performing three-dimensional grid interpolation on the real-time volume scan data based on the three-dimensional grid, obtaining third point cloud data of grid points corresponding to the third position data in the three-dimensional grid, and fourth point cloud data of grid points corresponding to the fourth position data in the three-dimensional grid.

[0092] It should be noted that this step is the same as the operation in the aforementioned S16 and will not be described in detail here.

[0093] S162, filter and resample the third point cloud data based on the echo intensity threshold to obtain the third highest echo intensity point cloud data with echo intensity greater than the echo intensity threshold, filter and resample the fourth point cloud data based on the echo intensity threshold to obtain the fourth highest echo intensity point cloud data with echo intensity greater than the echo intensity threshold.

[0094] Exemplarily, in an embodiment of the present invention, the data to be input into the point cloud motion field estimation network model is subjected to the same three-dimensional grid interpolation, filtering and resampling operations as step S133, and the third highest echo intensity point cloud data and the fourth highest echo intensity point cloud data with higher echo intensity are selected as real-time acquired data for input. This can make the point cloud data generated in the two frames distributed in clusters, which is more conducive to single-frame feature extraction of the point cloud motion field estimation network model, eliminate the system influence of weather radar volume scanning data, and further improve the estimation accuracy.

[0095] S163, input the third highest echo intensity point cloud data and the fourth highest echo intensity point cloud data into the trained point cloud motion field estimation network model, and output the estimated motion vector of the detection pixel point between the third time node and the fourth time node.

[0096] It should be noted that this step is the same as the operation in the aforementioned S16, except that the data input to the point cloud motion field estimation network model is replaced with the filtered and resampled third highest echo intensity point cloud data and fourth highest echo intensity point cloud data, which will not be elaborated here.

[0097] Figure 8 FIG. 1 is a schematic diagram of a structure of an estimation system provided by an embodiment of the present invention. Figure 8As shown, the estimation system has the function of implementing the above-mentioned storm motion vector estimation method. The estimation system includes a first data acquisition module 1, a virtual sports field generation module 2, a data set establishment module 3, a network model establishment module 4, a second data acquisition module 5 and a motion vector estimation module 6. Among them, the first data acquisition module 1 is used to obtain historical volume scan data of multiple different target areas, and the historical volume scan data includes position data and time data. The position data includes the first position data of the detected pixel point at the beginning of the volume scan, and the time data includes the first time node corresponding to the first position data. The virtual sports field generation module 2 is used to generate a three-dimensional grid that matches the target area, and uses the Perlin noise method to generate a random virtual sports field on the three-dimensional grid, so that each grid point in the three-dimensional grid has a corresponding motion vector. The data set establishment module 3 is used to perform three-dimensional grid interpolation on the historical volume scan data based on the three-dimensional grid, obtain the first point cloud data of the first position data corresponding to the grid point in the three-dimensional grid, and then extrapolate the detection pixel point corresponding to the first position data based on the motion vector, obtain the second position data corresponding to the detection pixel point at the second time node after the preset time, and perform three-dimensional grid interpolation based on the three-dimensional grid to obtain the second point cloud data of the second position data corresponding to the grid point in the three-dimensional grid, record the first point cloud data and the second point cloud data as a data pair, and collect the data pairs corresponding to the historical volume scan data of multiple different target areas to form a training data set. The network model establishment module 4 is used to establish a point cloud motion field estimation network model based on the FlowNet3D network structure, and input the training data set for training. The second data acquisition module 5 is used to obtain real-time volume scan data of the target area, and the real-time volume scan data includes the third position data of the detection pixel point at the third time node, and the fourth position data of the detection pixel point at the fourth time node, and the third time node and the fourth time node are two time nodes separated by a distance. The motion vector estimation module 6 is used to perform 3D grid interpolation on the real-time volume scan data based on the 3D grid and input the trained point cloud motion field estimation network model, and output the estimated motion vector of the detection pixel point between the third time node and the fourth time node.

[0098] Optionally, the data set establishment module 3 can also be used to set an echo intensity threshold. Then, the first point cloud data is filtered and resampled based on the echo intensity threshold to obtain the first high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold, and the second point cloud data is filtered and resampled based on the echo intensity threshold to obtain the second high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold. Finally, the first high echo intensity point cloud data and the second high echo intensity point cloud data are recorded as a data pair, and the data pairs corresponding to the historical volume scan data of multiple different target areas are collected to form a training data set.

[0099] Optionally, the data set establishment module 3 can also be used to obtain first radar reflectivity data corresponding to the detection pixel point at the first time node, and second radar reflectivity data corresponding to the detection pixel point at the second time node, and add the first radar reflectivity data and the second radar reflectivity data to the data pair for recording.

[0100] Fig. 9 FIG. 1 is a schematic diagram of a control structure of another estimation system provided by an embodiment of the present invention. Fig. 9 As shown, the estimation system also includes a processor and a memory for storing instructions executable by the processor. The estimation system 9100 may be a computer device, and the estimation system 9100 may include one or more of the following components: a processor 9101 , a memory 9102 , a communication interface 9103 , and a bus 9104 .

[0101] The processor 9101 includes one or more processing cores, and the processor 9101 executes various functional applications and information processing by running software programs and modules. The memory 9102 and the communication interface 9103 are connected to the processor 9101 through the bus 9104. The memory 9102 can be used to store at least one instruction, and the processor 9101 is used to execute the at least one instruction to implement each step in the above method embodiment.

[0102] In addition, memory 9102 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, and volatile or non-volatile storage devices include but are not limited to: magnetic disks or optical disks, electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), read-only memory (ROM), magnetic memory, flash memory, and programmable read-only memory (PROM).

[0103] Exemplarily, in an embodiment of the present invention, a non-transitory computer storage medium including instructions is also provided, such as a memory including instructions, and the above instructions can be used by a processor to execute the aforementioned storm motion vector estimation method. For example, the non-transitory computer storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0104] Unless otherwise defined, the technical terms or scientific terms used herein shall have the usual meanings understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar words used in the patent application specification and claims of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not indicate a quantity limitation, but indicate the existence of at least one. Words such as "include" or "comprise" and similar words mean that the elements or objects appearing before "include" or "comprise" include the elements or objects listed after "include" or "comprise" and their equivalents, and do not exclude other elements or objects. Words such as "connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0105] The above descriptions are merely optional embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A storm motion vector estimation method, characterized in that: include: Acquire historical volume scan data of multiple different target areas, the historical volume scan data including position data and time data, the position data including first position data of a detected pixel point at the beginning of the volume scan, and the time data including a first time node corresponding to the first position data; Generate a three-dimensional grid that matches the target area, and generate a random virtual motion field on the three-dimensional grid using a Perlin noise method so that each grid point in the three-dimensional grid has a corresponding motion vector; Performing three-dimensional grid interpolation on the historical volume scan data based on the three-dimensional grid to obtain first point cloud data of a grid point corresponding to the first position data in the three-dimensional grid, then extrapolating the detection pixel point corresponding to the first position data based on the motion vector to obtain second position data corresponding to the detection pixel point at a second time node after a preset time, and performing three-dimensional grid interpolation based on the three-dimensional grid to obtain second point cloud data of a grid point corresponding to the second position data in the three-dimensional grid, recording the first point cloud data and the second point cloud data as a data pair, and collecting the data pairs corresponding to the historical volume scan data of the multiple different target areas to form a training data set; Establishing a point cloud motion field estimation network model based on the FlowNet3D network structure, and inputting the training data set for training; Acquire real-time volume scan data of the target area, the real-time volume scan data including third position data of the detection pixel point at a third time node, and fourth position data of the detection pixel point at a fourth time node, the third time node and the fourth time node being two time nodes separated by the preset time; Based on the three-dimensional grid, three-dimensional grid interpolation is performed on the real-time volume scan data and the trained point cloud motion field estimation network model is input to output the estimated motion vector of the detection pixel point between the third time node and the fourth time node.

2. The storm motion vector estimation method according to claim 1, characterized in that: The preset time interval ranges from 5 to 6 minutes.

3. The storm motion vector estimation method according to claim 1, characterized in that: The method further comprises: performing three-dimensional grid interpolation on the historical volume scan data based on the three-dimensional grid to obtain first point cloud data of a grid point corresponding to the first position data in the three-dimensional grid, then extrapolating the detection pixel point corresponding to the first position data based on the motion vector to obtain second position data corresponding to the detection pixel point at a second time node after a preset time, and performing three-dimensional grid interpolation based on the three-dimensional grid to obtain second point cloud data of a grid point corresponding to the second position data in the three-dimensional grid, recording the first point cloud data and the second point cloud data as a data pair, and collecting the data pairs corresponding to the historical volume scan data of the multiple different target areas to form a training data set, and further comprising: Set the echo strength threshold; Filtering and resampling the first point cloud data based on the echo intensity threshold to obtain first high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold, and filtering and resampling the second point cloud data based on the echo intensity threshold to obtain second high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold; The first high echo intensity point cloud data and the second high echo intensity point cloud data are recorded as the data pair, and the data pairs corresponding to the historical volume scan data of the multiple different target areas are collected to form the training data set.

4. The storm motion vector estimation method according to claim 3, characterized in that: Performing three-dimensional grid interpolation on the real-time volume scan data based on the three-dimensional grid and inputting the trained point cloud motion field estimation network model, and outputting an estimated motion vector of the detection pixel point between the third time node and the fourth time node, including: Performing three-dimensional grid interpolation on the real-time volume scan data based on the three-dimensional grid to obtain third point cloud data of grid points corresponding to the third position data in the three-dimensional grid, and fourth point cloud data of grid points corresponding to the fourth position data in the three-dimensional grid; filtering and resampling the third point cloud data based on the echo intensity threshold to obtain third point cloud data with the highest echo intensity greater than the echo intensity threshold, and filtering and resampling the fourth point cloud data based on the echo intensity threshold to obtain fourth point cloud data with the highest echo intensity greater than the echo intensity threshold; The third high echo intensity point cloud data and the fourth high echo intensity point cloud data are input into the trained point cloud motion field estimation network model, and the estimated motion vector of the detection pixel point between the third time node and the fourth time node is output.

5. The storm motion vector estimation method according to claim 1, characterized in that: Before establishing a point cloud motion field estimation network model based on the FlowNet3D network structure and inputting the training data set for training, the Storm motion vector estimation method further includes: Obtain first radar reflectivity data corresponding to the detection pixel point at the first time node, and second radar reflectivity data corresponding to the detection pixel point at the second time node, and add the first radar reflectivity data and the second radar reflectivity data to the data pair for recording.

6. An estimation system, characterized in that The estimation system comprises: A first data acquisition module is used to acquire historical volume scan data of multiple different target areas, wherein the historical volume scan data includes position data and time data, wherein the position data includes first position data of a detected pixel point at the beginning of the volume scan, and the time data includes a first time node corresponding to the first position data; A virtual sports field generation module is used to generate a three-dimensional grid that matches the target area, and generate a random virtual sports field on the three-dimensional grid using a Perlin noise method, so that each grid point in the three-dimensional grid has a corresponding motion vector; a data set establishment module, configured to perform three-dimensional grid interpolation on the historical volume scan data based on the three-dimensional grid, obtain first point cloud data of a grid point corresponding to the first position data in the three-dimensional grid, then extrapolate the detection pixel point corresponding to the first position data based on the motion vector, obtain second position data corresponding to the detection pixel point at a second time node after a preset time, perform three-dimensional grid interpolation based on the three-dimensional grid, so as to obtain second point cloud data of a grid point corresponding to the second position data in the three-dimensional grid, record the first point cloud data and the second point cloud data as a data pair, and collect the data pairs corresponding to the historical volume scan data of the multiple different target areas to form a training data set; A network model building module is used to build a point cloud motion field estimation network model based on the FlowNet3D network structure, and input the training data set for training; a second data acquisition module, configured to acquire real-time volume scan data of the target area, wherein the real-time volume scan data includes third position data of the detection pixel point at a third time node and fourth position data of the detection pixel point at a fourth time node, wherein the third time node and the fourth time node are two time nodes spaced apart from each other; A motion vector estimation module is used to perform three-dimensional grid interpolation on the real-time volume scan data based on the three-dimensional grid and input the trained point cloud motion field estimation network model, and output the estimated motion vector of the detection pixel point between the third time node and the fourth time node.

7. The estimation system according to claim 6, characterized in that The data set building module is also used for: Set the echo strength threshold; Filtering and resampling the first point cloud data based on the echo intensity threshold to obtain first high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold, and filtering and resampling the second point cloud data based on the echo intensity threshold to obtain second high echo intensity point cloud data with an echo intensity greater than the echo intensity threshold; The first high echo intensity point cloud data and the second high echo intensity point cloud data are recorded as the data pair, and the data pairs corresponding to the historical volume scan data of the multiple different target areas are collected to form the training data set.

8. The estimation system according to claim 6, characterized in that The data set establishment module is also used to obtain first radar reflectivity data corresponding to the detection pixel point at the first time node, and second radar reflectivity data corresponding to the detection pixel point at the second time node, and add the first radar reflectivity data and the second radar reflectivity data to the data pair for recording.

9. An estimation system, characterized in that The estimation system comprises: processor; a memory for storing instructions executable by a processor; Wherein, the processor is configured to execute the storm motion vector estimation method described in any one of claims 1 to 5.

10. A computer storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the storm motion vector estimation method according to any one of claims 1 to 5 is implemented.

Images