A flow field feature tracking method based on graph optimization
By constructing a feature matching binary graph for flow field feature tracking, the problem of small size and fast moving feature tracking in the prior art is solved, and higher accuracy and versatility are achieved, threshold setting is avoided, and flow field is improved. Feature tracking effect.
Patent Information
- Application Number
- CN202210251043.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-15
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-03-15
AI Technical Summary
Existing flow field feature tracking methods have tracking interruptions when dealing with small and fast moving features, and relying on manual threshold setting results in inaccurate analysis results.
A graph-based optimization method is adopted to construct a feature matching binary graph and use the principle of maximum similarity matching to perform feature tracking, avoiding threshold setting and improving the continuity and accuracy of feature areas.
It effectively solves the problem of tracking interruption of small size and fast moving features, improves the accuracy of flow field feature tracking, reduces manual intervention, and enhances the universality of the method.
Smart Images

Figure CN114722735B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of flow field feature tracking, and in particular to a flow field feature tracking method based on graph optimization. Background Art
[0002] In order to assist scientists in conducting in-depth exploration and analysis of numerical simulation data, it is necessary to visualize the numerical simulation data. If standard visualization methods, such as isosurface and volume rendering, are used, important information will be omitted due to the huge amount of data, and the visualization results will be difficult to analyze. Usually, scientists only pay more attention to the evolution process of specific areas within the data. In this regard, feature-based flow field visualization can effectively reduce storage space and improve analysis efficiency by combining the definition of features by domain experts, extracting the feature area for visualization and subsequent analysis.
[0003] Based on the feature extraction results, scientists can more clearly understand the attribute information of the characteristic areas in the flow field. In addition, since the characteristic areas in the flow field are constantly changing in the time domain, they may split, merge or decay. Exploring the evolution of characteristic areas is an effective method for analyzing physical phenomena in the flow field in scientific simulation.
[0004] Feature tracking is to solve the correspondence problem between features of continuous time steps, and track the dynamic growth, reduction, birth, decay and interaction of feature regions in the time domain with other feature regions. At present, the feature tracking problem for feature-based time-varying flow field visualization research is mainly divided into feature tracking algorithms based on region correspondence and feature tracking algorithms based on attribute correspondence. Among them, the feature tracking algorithm based on region correspondence can find the feature correspondence problem through region overlap, distance, maximum cross-correlation standard or minimum affine transformation matrix. The feature tracking algorithm based on attribute correspondence is to achieve correspondence between continuous time steps through the basic attributes of the features, such as the position, size, volume, etc. of the features. Feature tracking is to find the corresponding feature objects in continuous time steps and combine them into spatiotemporal features. The above feature tracking method needs to compare any combination of features one by one when performing feature correspondence of adjacent time steps, which is time-consuming and costly. In this regard, Muelder et al. used frame consistency when performing feature extraction, that is, with the help of the feature extraction results of the previous time step, first predicted the position of a feature in the previous time step in the current time step, and then adjusted the position through the region growth or reduction algorithm. At the same time, the feature tracking was completed in the prediction process. Previous feature matching studies have mostly used overlapping or similar attribute criteria. The volume overlap criterion works well when the temporal sampling rate is high enough and the corresponding features have a large degree of overlap. However, this criterion cannot track small / thin and fast-moving features because the overlap of such features may not meet the set threshold after moving. In response to this, Ji et al. proposed using the Earth Mover's Distance (EMD) metric to measure the matching cost between arbitrary source and target feature components. A search tree is constructed, and finally the final feature tracking result is determined by the minimum path; however, the above tracking method relies on local tracking decisions between two consecutive time steps. When the data is very noisy and the number of features is too large, the tracking trajectory is prone to be broken.
[0005] It can be seen that for traditional feature tracking methods, corresponding measurement standards need to be set in advance when judging various feature events. For example, when judging continuity, splitting and merging events based on the overlap and volume difference between two feature areas, the corresponding overlap and volume difference thresholds need to be set. The corresponding feature event is judged to occur only when and only when the threshold conditions are met. Therefore, the setting of the threshold is an important factor affecting the tracking results. Due to the complexity of flow field data, simple threshold judgment cannot explain various flow field motion phenomena, and tracking errors are prone to occur. Summary of the invention
[0006] The purpose of the present invention is to provide a method for tracking flow field characteristics based on graph optimization for achieving effective tracking of flow field characteristics.
[0007] To this end, the technical solution of the present invention is as follows:
[0008] A method for tracking flow field features based on graph optimization, the steps are as follows:
[0009] S1, extracting vortex regions in the flow field by determining characteristic data points that meet the vortex characteristics from sampling points evenly distributed in the flow field at intervals of time steps; at the same time, drawing a distribution map of vortex regions in the flow field, and obtaining characteristic attributes of each vortex region, including the size of the characteristic region, the center of the characteristic region, and a numerical matrix of the characteristic region;
[0010] S2. According to the vortex area distribution map in the flow field, a feature matching bipartite graph G = (U ∪ V, E, w) is constructed to realize the construction of tracking relationship between feature areas at different time steps; wherein each feature area is represented by a feature node, based on which U represents the set of all feature nodes in time step t, and V represents the set of all feature nodes in time step t+1; the similarity value w of the correlation between two feature nodes;
[0011] The calculation formula of w is:
[0012]
[0013] In the formula, u i ∩v j Indicates the number of feature data points that overlap between two feature nodes, Max(u i ,v j ) represents the maximum number of feature data points contained in two feature nodes; when w>0, it indicates that there is a correlation between the two feature nodes, then an edge E is inserted between the two nodes obtained by the above calculation, and the similarity value w is marked on it; when w<0, w is reassigned to 0;
[0014] S3. Based on the optimal matching of the feature matching bipartite graph, complete the tracking of the largest components in the continuous events and split-merge events in the flow field feature tracking; the specific steps are:
[0015] S301, based on the maximum similarity value w matching principle, the feature nodes in U are matched with the feature nodes in V in a one-to-one correspondence;
[0016] S302, determining the final tracking result based on the characteristic events divided by the conflict domain;
[0017] 1) Define candidate points and find the candidate point set of each feature node in the previous time step in the adjacent time step; when each fully matched feature node u in U i ∈U enumerates all points v in V that currently have no complete match j ∈V, and satisfies w(u,v)>0, then this type of point vj ∈V is called a candidate point; correspondingly, this type of point v j The set of nodes is called feature node u i The candidate point set of
[0018] 2) Based on each complete match and its corresponding candidate point set, construct all possible split events and possible merge events;
[0019] 3) According to the similarity calculation formula: Calculate the probability of all consecutive events, possible split events, and possible merge events;
[0020] 4) Define conflict domain: Only one event can be selected from the possible events containing the feature node corresponding to any feature node at the same time step as the final tracking result. Therefore, the possible events containing the feature node are called a conflict domain; all possible events obtained according to the above step 2) are divided into several conflict domains, and the maximum value of the possibility calculation based on several possible events in the same conflict domain is selected as the tracking result, and the possible events corresponding to the same feature node as the tracking result in different conflict domains are deleted, and the possible events in each conflict domain are determined in turn; finally, the feature node of the unmatched event is determined as a new event or an extinction event according to its appearance time.
[0021] Furthermore, in step S1, the method for extracting the vortex region in the flow field is:
[0022] The vector at each sampling point in the flow field is usually represented by the following three-variable function:
[0023] F(x,y,z)=(F1(x,y,z),F2(x,y,z),F3(x,y,z)),
[0024] Where F1, F2 and F3 are the three-dimensional coordinate directions of F(x, y, z) in the Cartesian coordinate system. and The three elements on , then the gradient value is expressed as:
[0025]
[0026] The eddy area is defined as S 2 +Ω 2 The two eigenvalues of are negative in the region, where S and Ω are the symmetric and antisymmetric parts of the Jacobian matrix vector field, that is:
[0027]
[0028]
[0029] Calculate the vorticity value Q at each sampling point in the flow field to locate the vortex in the flow field:
[0030] Q=0.5(||Ω|| 2 -||S|| 2 ),
[0031] Among them, the larger the Q value is, the higher the vorticity value of the sampling point is, that is, the higher the probability of being detected as a vortex area in the flow field;
[0032] Define the threshold Tho, and extract the sampling points that satisfy Q>Tho as feature data points to extract the vortices in the flow field.
[0033] Furthermore, in step S1, the method for drawing the vortex area distribution map in the flow field is: arbitrarily select a number of data points from all the feature data points as seed points, and maintain a queue based on each seed point; then, search the area adjacent to the seed point, push the feature data points that meet the threshold conditions into the corresponding queue, and at the same time mark the feature data points belonging to the same connected area with the same numerical value until the search stops when the elements in the queue are empty; finally, the area corresponding to each queue is a feature area extracted from the flow field, that is, the vortex area in the flow field.
[0034] Compared with traditional feature tracking methods, this flow field feature tracking method based on graph optimization avoids the artificial threshold setting in existing feature tracking methods through optimal matching of bipartite graphs, and combines the similarity measurement results of adjacent time steps to achieve optimal feature area tracking, making the method more versatile; and then through the optimal matching between features, it effectively solves the problem of tracking interruption of features with small volume and fast moving speed in the flow field, so this method effectively improves the accuracy of flow field feature tracking. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A flow chart of a method for tracking flow field characteristics based on graph optimization according to the present invention;
[0036] Figure 2 It is a schematic diagram of the distribution of vortex regions in the flow field obtained by the region growing method in step S1 in the method for tracking flow field characteristics based on graph optimization of the present invention;
[0037] Figure 3 is a feature matching bipartite graph correspondingly constructed based on the vortex area distribution map in the flow field obtained in step S1 in the embodiment of the present invention;
[0038] Figure 4 It is a schematic diagram of a complete matching result obtained based on a feature matching bipartite graph in an embodiment of the present invention;
[0039] Figure 5In the embodiments of the present invention, based on Figure 4 A partial schematic diagram of the conflict detection graph of possible events obtained;
[0040] Figure 6 In the embodiments of the present invention, based on Figure 4 The schematic diagram of the result further represents the evolution of the characteristic area;
[0041] Figure 7 A schematic diagram of the drawing result of the flow field data around a semi-cylinder in an embodiment of the present invention;
[0042] Figure 8 In the embodiment of the present invention, the method for extracting vortices in the process of step S1 of the present application is used to process the original data of the numerical simulation data around the semi-cylinder and then draw the vortex area distribution map in the flow field
[0043] Fig. 9 It is a schematic diagram of the time evolution of a vortex in the 560th time step of numerical simulation data around a semi-cylinder in an embodiment of the present invention;
[0044] Fig.10 It is a schematic diagram of the drawing result of the simulation data of the water flow around the tail of the research vessel Tangaroa in an embodiment of the present invention;
[0045] Fig.11 In the embodiment of the present invention, the method for extracting vortices in the process of step S1 of the present application is used to process the original data of the numerical simulation data of the stern water flow around the research ship Tangaroa, and then draw the vortex area distribution map in the flow field;
[0046] Fig.12 A schematic diagram of the evolution of three vortices over time in the 40th time step of numerical simulation data of the tail water flow around the research vessel Tangaroa in an embodiment of the present invention;
[0047] Fig.13 It is a DGA comparison diagram of tracking results based on the tracking method of the present application and based on the traditional tracking method in the embodiment of the present invention. DETAILED DESCRIPTION
[0048] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention in any way.
[0049] This embodiment takes the vortex region as the research object, and tracks the movement process of the vortex in the entire time domain based on the extraction results of the vortex in the fluid, so as to study the evolution of the vortex over time in a graphical and intuitive way.
[0050] like Figure 1As shown, the specific implementation steps of the flow field feature tracking method based on graph optimization are as follows:
[0051] S1, extracting the vortex area in the flow field at intervals of time steps and drawing a distribution map of the vortex area in the flow field, and obtaining characteristic attributes of each vortex area;
[0052] S101, extracting features of the vortex region in the flow field;
[0053] Specifically, the steps of feature extraction are as follows:
[0054] The vector at each sampling point in the flow field is usually represented by the following three-variable function:
[0055] F(x,y,z)=(F1(x,y,z),F2(x,y,z),F3(x,y,z)),
[0056] Where F1, F2 and F3 are the three-dimensional coordinate directions of F(x, y, z) in the Cartesian coordinate system. and The three elements on , then the gradient value is expressed as:
[0057]
[0058] The eddy area is defined as S 2 +Ω 2 The two eigenvalues of are negative in the region, where S and Ω are the symmetric and antisymmetric parts of the Jacobian matrix vector field, that is:
[0059]
[0060]
[0061] Calculate the vorticity value Q at each sampling point in the flow field to locate the vortex in the flow field:
[0062] Q=0.5(||Ω|| 2 -||S|| 2 ),
[0063] Among them, the larger the Q value is, the higher the vorticity value of the sampling point is, that is, the higher the probability of being detected as a vortex area in the flow field;
[0064] Define a threshold value Tho, and extract sampling points that satisfy Q>Tho as feature data points, so as to extract vortices in the flow field; wherein the specific value of the threshold value Tho is determined based on actual experience;
[0065] S102, marking characteristic areas in the flow field;
[0066] Based on the characteristic data points extracted in step S101, the regional connectivity method is used to divide and mark each characteristic region in the flow field; specifically, the method is implemented based on regional growth, that is: first, a number of data points are randomly selected from all characteristic data points as seed points, and a queue is maintained based on each seed point; then, the region adjacent to the seed point is searched, and the characteristic data points that meet the threshold conditions are pressed into the corresponding queue, and the characteristic data points belonging to the same connected region are marked with the same value until the search stops when the elements in the queue are empty; finally, the region corresponding to each queue is a characteristic region extracted from the flow field, that is, the vortex region in the flow field;
[0067] like Figure 2 The figure shows the schematic diagram of the region growth gradually obtained through the above marking steps; Figure 2 In the figure, the chessboard corresponds to the flow field area, and the chessboard grids evenly divided on it correspond to the sampling points evenly distributed in the flow field. Figure 2 The left figure is a schematic diagram after the initial seed point is selected. Figure 2 The middle figure is a schematic diagram of searching for feature data points that meet the threshold conditions and pressing them into the corresponding queue. Figure 2 The figure on the right is a schematic diagram of the search being completed and all characteristic data points being pushed into the corresponding queue, i.e., the vortex area distribution diagram in the flow field obtained by the above-mentioned regional growth method;
[0068] In practical applications, due to the large number of sampling points in the flow field, Figure 2 The number of sampling points demonstrated is greater, and therefore the shape of the characteristic region (ie, the vortex region) actually finally obtained will be closer to the actual region contour.
[0069] S103, obtaining characteristic attributes of each characteristic region;
[0070] In order to facilitate the subsequent implementation of adjacent time step feature tracking based on feature attributes, it is also necessary to obtain the feature attributes of each feature area, including the size of the feature area, the center of the feature area and the numerical matrix of the feature area; specifically,
[0071] ① The size of the feature area: It is calculated by the number of feature data points contained in the feature area; the size of the feature is used to analyze the evolution of the feature area in the flow field;
[0072] ②The center point of the feature area: It is represented by the positions and unweighted mean of all feature data points contained in the feature area: In the formula, is the center point of the feature, N is the total number of feature data points contained in the feature area, P i is the position coordinate corresponding to any feature data point; the center point of the feature can reflect the position of the feature area in the flow field to a great extent;
[0073] ③ Numerical matrix of feature area: It is represented by the combination of spatial position information of all feature data points in the feature area; the numerical matrix of features records the specific position of the feature area in the flow field, which is important information for subsequent calculation of overlap and analysis of various properties of the feature area;
[0074] As shown in Table 1 below, the storage content of the feature attributes of one of the feature regions is summarized. The storage content uses the time step as the basic storage unit, and stores the feature number, feature size, feature center point and original numerical matrix of the feature of each feature region in the single-step data.
[0075] Table 1:
[0076] Feature attributes Storage Information Time step 1 Feature Number 1 Feature size 327 Feature Center (12.0,14.4,13.2) The original numerical matrix of features [(12.0,14.0,13.0),…] Feature Number 2
[0077] It should be noted that although the characteristic attributes of the above-mentioned characteristic regions can enable researchers to understand the characteristic attributes of characteristic regions in the flow field. However, considering the importance of time-varying information of fluid data, tracking the evolution of the characteristics of researchers' interest in the time domain is an important basis for analyzing various physical phenomena in the flow field. Therefore, it is further necessary to solve the correspondence problem between characteristic regions of continuous time steps to achieve tracking of the more important and obvious phenomena that occur in the evolution process of characteristic regions, which are called characteristic events, including: new events, continuous events, split events, merger events, and decay events.
[0078] S2. According to the vortex area distribution map in the flow field, a feature matching bipartite graph is constructed to achieve modeling of the tracking relationship;
[0079] The feature regions and feature region attributes extracted corresponding to each time step are regarded as a feature object set, and the feature matching bipartite graph G = (U ∪ V, E, w) is used to show the feature matching between two adjacent time steps; wherein each feature region is represented by a feature node, based on which U represents the set of all feature region nodes in time step t, and V represents the set of all feature nodes in time step t+1; edges E are set between nodes with correlation in different time steps to represent the potential correspondence between feature nodes in consecutive time steps; each edge E is marked with a similarity value w representing the correlation between two feature nodes;
[0080] Among them, the similarity value w between each pair of nodes (u, v) ∈ U×V is calculated using the standardized volume overlap between the two features, and its calculation formula is:
[0081]
[0082] In the formula, u i ∩ v jIndicates the number of feature data points that overlap between two feature nodes, Max(u i ,v j ) represents the maximum number of feature data points contained in two feature nodes; the specific value can be calculated by tracing back the feature region attributes obtained in step S103;
[0083] For the calculation result of the similarity value w, when w>0, an edge E is inserted between the two nodes obtained by the above calculation, and the similarity value w is marked on it; when w<0, w is reassigned to 0; for the weight w, a higher edge weight corresponds to more similar features, that is, a higher probability of corresponding relationship.
[0084] like Figure 3 The figure shows a feature matching bipartite graph corresponding to the vortex area distribution map in the flow field obtained in step S1; wherein, Figure 3 The left part is the vortex area distribution diagram in the flow field corresponding to time step t and time t+1. It can be seen from this part of the figure that the flow field at time step t includes five characteristic areas (characteristic area U1, characteristic area U2, characteristic area U3, characteristic area U4 and characteristic area U5), and the flow field at time step t+1 includes five characteristic areas (characteristic area V1, characteristic area V2, characteristic area V3, characteristic area V4 and characteristic area V5); Figure 3 The right part is a feature matching bipartite graph constructed through step S201, including five feature nodes at time step t (feature node u1, feature node u2, feature node u3, feature node u4 and feature node u5) and five feature nodes at time step t+1 (feature node v1, feature node v2, feature node v3, feature node v4 and feature node v5). At the same time, edges E are set between nodes with volume overlapping associations at different time steps to connect the nodes, and similarity values w are marked above the connecting lines.
[0085] S3, based on the optimal matching of the feature matching bipartite graph, complete the tracking of the largest components in the continuous events and split-merge events in the flow field feature tracking;
[0086] Based on the feature matching bipartite graph constructed above, and taking into account the characteristic event types in the unsteady flow field, they can be classified into two categories according to the feature correspondence:
[0087] (1) Continuous events: one-to-one correspondence;
[0088] (2) Split events: one-to-many, merge events: many-to-one correspondence;
[0089] Therefore, when tracking, we first perform the global optimal one-to-one matching based on the feature similarity weight to complete the tracking of the largest component in the continuous events and the split-merge events. In addition, considering the existence of birth and decay events in the flow field data, we only need to require the maximum weight matching, without finding a corresponding feature for each feature area.
[0090] Based on this, the specific implementation steps of step S3 are as follows:
[0091] S301, based on the maximum similarity value w matching principle, the feature nodes in U are matched with the feature nodes in V in a one-to-one correspondence;
[0092] Specifically, find out the specified feature node in U and all the feature nodes associated with V, and then select a pair with the largest similarity value w to form a complete match, and so on, until the matching of each feature node in U is completed; in this process, when two feature nodes in U correspond to the same feature node in V, select the one with the largest similarity value w to form a complete match;
[0093] like Figure 4 As shown, based on Figure 4 The feature matching bipartite graph shown in the left part is obtained after being processed in step S301. Figure 4 The middle part shows a schematic diagram of the one-to-one feature tracking result; for feature node u1, the feature node with the maximum similarity value w with feature node u1 is V1, so feature node u1 forms a complete match with feature node V1; the feature node with the maximum similarity value w with feature node u2 is V4, so feature node u2 forms a complete match with feature node V4; the feature node with the maximum similarity value w with feature node u3 is V5, the feature node with the maximum similarity value w with feature node u4 is V5, and the feature node with the maximum similarity value w with feature node u5 is V5, but based on the maximum similarity value principle, feature node u4 forms a complete match with feature node V5;
[0094] S302, determining the final tracking result based on the characteristic events divided by the conflict domain;
[0095] 1) Define candidate points and find the candidate point set of each feature node in the previous time step in the adjacent time step;
[0096] The definition of candidate points is: when each fully matched feature node u in U i ∈U enumerates all points v in V that currently have no complete match j ∈V, and satisfies w(u,v)>0, then this type of point v j ∈V is called a candidate point; correspondingly, this type of point v j The set of nodes is called feature node ui The candidate point set of
[0097] like Figure 4 As shown, for feature node u1, after step S301, feature node u1 forms a complete match with feature node v1, and feature node v2 and feature node v3 that also have an association relationship with feature node u1 are called candidate points, forming a candidate point set of {v2, v3}; for feature node u2, after step S301, feature node u2 forms a complete match with feature node v4, and feature node v3 that also has an association relationship with feature node u2 is called a candidate point, forming a candidate point set of {v3}; and for feature node u4, after step S301, feature node u4 forms a complete match with feature node v5, and feature node v4 that also has an association relationship with feature node u4 is called a candidate point, forming a candidate point set of {v4};
[0098] 2) Based on each complete match and its corresponding candidate point set, construct all possible split events and merge events; specifically,
[0099] (1) Construct all possible splitting events:
[0100] For example Figure 4 For the feature node u1 shown in the figure, each possible event consists of u1 and its matched node v1 plus its candidate point set {v2, v3}, and the formula is expressed as follows:
[0101] u1→{v1}+p({v2,v3}),
[0102] Then, for feature node u1, several possible events can be obtained, including:
[0103] u1→{v1}, that is, u1 evolves into v1, which is a continuous event;
[0104] u1→{v1, v2}, that is, u1 evolves into {v1, v2}, which is a split event;
[0105] u1→{v1, v3}, that is, u1 evolves into {v1, v3}, which is a split event;
[0106] u1→{v1, v2, v3}, that is, u1 evolves into {v1, v2, v3}, which is a split event;
[0107] Similarly, the possible events corresponding to feature node u2 include: u2→{v4}, u2→{v3, v4}; the possible events corresponding to feature node u3 include: u4→{v5};
[0108] (2) Construct all possible merge events:
[0109] The split event and the merge event are inverse events of each other; therefore, when constructing the merge event, trace back from time step t+1 to time step t, and similarly for each complete and matched point v in V j ∈V enumerates all points u in U that currently have no complete match i ∈U, where w(u, v)>0;
[0110] like Figure 4 As shown, according to the feature matching bipartite graph, for the feature node v i For example, the characteristic node {u3, u5} at time step t is v i Based on this, for feature node v1, it is not associated with the candidate point set at time step t, so there is no new possible merging event; for feature node v4, it is associated with {u3} in the candidate point set at time step t, so a new possible merging event is constructed: {u2, u3}→{v4}; for feature node v5, it is associated with {u3, u5} in the candidate point set at time step t, so three new possible merging events are constructed: {u3, u4}→{v5}, {u4, u5}→{v5}, {u3, u4, u5}→{v5};
[0111] 3) According to the similarity calculation formula: Calculate the probabilities of all consecutive events, possible split events, and possible merge events;
[0112] Take u1→{v1, v2} as an example,
[0113] 4) All possible events obtained according to the above step 2) are divided into several conflict domains, and the maximum value of the possibility calculation is selected as the tracking result based on several possible events in the same conflict domain, and possible events corresponding to the same feature nodes as the tracking result in different conflict domains are deleted, and the possible events in each conflict domain are determined in turn; finally, the feature nodes of the unmatched events are determined as new events or extinct events according to their appearance time;
[0114] Since it is often assumed that any feature can only participate in one feature event in the same time period [t, t+1] when tracking the features of fluid data, all feature events that share at least one node from U or V are conflicting. Based on this, this application constructs several conflict domains according to possible events to represent possible conflicts between possible events.
[0115] Specifically, first, define the conflict domain: only one event can be selected from the possible events containing the feature node corresponding to any feature node at the same time step as the final tracking result. Therefore, the possible events containing the feature node are called a conflict domain; the possible events specifically include continuous events, merged events, or split events; then, the maximum value of the possibility calculation of several possible events in the same conflict domain is selected as the tracking result, and the possible events corresponding to the same feature nodes as the tracking result in different conflict domains are deleted based on the tracking result; and so on until the tracking results of all conflict domains are determined; finally, detect whether there are unmatched feature nodes. If the unmatched node appears in the previous time step, it is determined as an extinction event. If the unmatched node appears in the later time step, it is determined as a new event;
[0116] The specific classification process of step 4) and the final classification results are shown in Table 2 below.
[0117] Table 2:
[0118]
[0119] From the tracking results shown in Table 2, it can be seen that the final evolution relationship between the five feature nodes at time step t and the five feature nodes at time step t+1 is: split event: {u1}→{v1, v2, v3}, continuous event: {u2}→{v4}, merge event: {u3, u4}→{v5}; all other unconnected nodes indicate feature objects whose corresponding features cannot be found in adjacent time steps, and such features are interpreted as birth and extinction events; u5 does not find the corresponding feature in V, so it is judged as a decay event.
[0120] In order to facilitate more intuitive understanding, Figure 5 Shown is based on Figure 4 The obtained partial schematic diagram of the conflict detection graph of possible events is marked with the tracking results. Figure 6 Shown is based on Figure 4 The evolution of the characteristic area is further represented, that is, the vortex u1 at t time step in the flow field evolves into vortex v1, vortex v2 and vortex v3 at t+1 time step, the vortex u2 at t time step in the flow field evolves into vortex v4 at t+1 time step, the vortex u3 and vortex u4 at t time step in the flow field evolve into vortex v5 at t+1 time step, and the vortex u5 at t time step in the flow field disappears at t+1 time step.
[0121] In this embodiment, the above process specifically demonstrates that the data at two adjacent time steps are successively subjected to feature extraction, optimal matching of feature matching bipartite graph and feature tracking to determine the specific change process of the feature area between two adjacent time steps, and then visualization is achieved based on the determined change process.
[0122] For traditional feature tracking methods, it is necessary to pre-set corresponding measurement standards when judging various feature events. For example, when judging continuous, split and merge events based on the overlap and volume difference between two feature areas, it is necessary to set corresponding overlap and volume difference thresholds. The corresponding feature event is judged to occur only when and only when the threshold conditions are met. Therefore, the setting of the threshold is an important factor affecting the tracking results. Due to the complexity of flow field data, simple threshold judgment cannot explain various flow field motion phenomena, and tracking errors are prone to occur; unlike traditional methods, the flow field feature tracking method of this application avoids the threshold selection part in the traditional tracking method, so that the accuracy of this method has been relatively improved, and it is more advantageous in actual use. Its purpose is to improve the accuracy of flow field feature tracking by designing a new flow field feature tracking method, which is specifically reflected in the fact that there is no need to manually set the tracking threshold and the tracking is more accurate for fluid features with small volume and fast movement.
[0123] Furthermore, the present application selects multiple vortices in the flow field and combines them with the corresponding drawing results to prove the effectiveness of the method of the present application.
[0124] (I) Verification test 1:
[0125] Experiment 1 uses numerical simulation data around a semi-cylinder. The incompressible 3D flow numerical simulation data around the semi-cylinder contains 1501 time steps in total. The dimension of each time step is 640×240×80, and the size of a single time step is about 1.2GB. Figure 7 The figure shows the plotting results of the flow field data around the semi-cylinder, and the figure shows the plotting results of the 100th, 250th, 500th, 600th, 700th, 800th, and 1500th time steps; it can be seen from the figure that the first 500 time steps are in the initialization stage, and the flow field changes slightly. From the time step 500 onwards, the vortex begins to fall off from the left side, which corresponds to the time axis part of the right half, that is, the time period of high flow field change;
[0126] like Figure 8The figure shows the distribution diagram of the vortex area in the flow field obtained by processing the original data of the numerical simulation data around the semi-cylinder using the method for extracting vortices in the process of step S1 of the present application; it can be seen from the figure that the direct visualization result based on the original volume data cannot clearly display the internal structure of the data; in contrast, the flow field visualization result based on spatial feature extraction can effectively reduce visual interference and reduce the complexity of visual analysis.
[0127] For the tracking results, take a vortex in the 560th time step of the numerical simulation data around a semi-cylinder as an example, Fig. 9 As shown in the figure, the vortex splits into two smaller vortices at the 580th time step, represented by a black vortex and a gray vortex respectively; the two vortices evolve independently in the following time steps. At the same time, after the vortex splits at the 580th time step, the sum of the volumes of the two vortices after the split is similar to the volume of the vortex before the split, and the vortex position does not move much; the rationality of the tracking results can be further verified by combining the drawing results. At the same time, Fig. 9 In order to more clearly show the interaction between the features, a DAG diagram based on the Sankey diagram is also used to represent the various characteristic events in the vortex evolution process; among them, in the DAG diagram, the dark square column corresponding to each time step represents the change in the volume size of the vortex.
[0128] (II) Verification Test 2:
[0129] Experiment 1 uses numerical simulation data of the water flow around the tail of the research vessel Tangaroa. The simulation data contains a total of 201 time steps, each time step data dimension is 300×180×120, the single time step data size is about 0.7GB, and the drawing time is about 3.2s. The figure shows the drawing results of the 0th, 50th, 100th, and 150th time steps. Fig.10 Figure 1 shows a plot of simulated flow data around the wake of the research vessel Tangaroa, showing the plots at time steps 0, 25, 50, 75, 100, 150, and 200.
[0130] like Fig.11 The figure shows the distribution map of the vortex area in the flow field obtained by processing the original data of the numerical simulation data of the stern water flow around the research vessel Tangaroa using the method for extracting vortices in the process according to step S1 of the present application; as can be seen from the figure, the direct visualization result based on the original volume data cannot clearly display the internal structure of the data; in contrast, the flow field visualization result based on spatial feature extraction can effectively reduce visual interference and reduce the complexity of visual analysis.
[0131] like Fig.12As shown in the figure, three vortices are selected from the 40th time step of the numerical simulation data of the stern water flow around the research vessel Tangaroa. From the drawing results, it can be seen that in the subsequent evolution process, the positions of the three vortices gradually approached each other, and finally merged into a larger vortex. As shown in the figure, the two vortices initially located on the right side merged into one vortex at the 41st time step, and the vortex initially located on the left side tended to be close to the position of the newly merged vortex, and merged at the 43rd time step; the corresponding figure below is the corresponding feature tracking result figure, and it can be seen that the tracking result is consistent with the drawing result.
[0132] To further verify the accuracy of the feature tracking method, this application uses the feature tracking method based on graph optimization and the traditional feature tracking method to track the numerical simulation data of the semi-cylinder, and sets the thresholds of the traditional feature tracking method Toverlap = 0.65, Tsize = 0.8; As mentioned above, the feature tracking method based on graph optimization proposed in this application mainly solves the problem of tracking interruption caused by threshold selection in the traditional feature tracking method. Therefore, statistics are collected for events with a front-to-back correspondence, including continuous events, splitting and merging events, and newborn and decaying events without a front-to-back correspondence, and the results are shown in Table 3.
[0133] Table 3 below shows the comparison of feature tracking results of different feature tracking methods.
[0134] Table 3:
[0135]
[0136] From the tracking results in Table 3, it can be seen that the traditional feature tracking method detects more birth and decay events than the feature tracking method based on graph optimization. Correspondingly, the feature tracking method based on graph optimization detects more continuous, split, and events; at the same time, by locating the differences in the detection results of the two methods, it is found that the traditional feature tracking method has a problem of tracking interruption when the vortex is small in volume, thereby detecting multiple continuous events as birth and decay events.
[0137] like Fig.13As shown in the figure, in order to better describe the evolution of the vortex over time, this paper uses gradient colors to show its dynamic evolution. The vortex of the time step is gray to dilute its visual effect, and the vortex of the intermediate time step is drawn with transitional colors from light to dark, so that the time-varying information of the feature can be presented in a single image. From the drawing results, it can be seen that the evolution process of the vortex should correspond to a continuous event, that is, the tracking result of the graph optimization. Further analysis of the error causes of the traditional tracking method shows that compared with other vortices in the same time step, the vortex volume corresponding to the interrupted part of the tracking is smaller, which causes the calculated overlap to be lower than the set threshold, causing the problem of no correspondence. In contrast, the feature tracking method based on bipartite graph optimization proposed in this paper does not need to use relevant thresholds for feature correspondence judgment, which can effectively avoid such problems.
[0138] To summarize, the graph optimization-based flow field feature tracking method of the present application improves the accuracy of flow field feature tracking by designing a new flow field feature tracking method. Specifically, there is no need to manually set the tracking threshold and the tracking is more accurate for fluid features with small volume and fast movement.
Claims
1. A flow field feature tracking method based on graph optimization, It is characterized in that Here are the steps: S1, extracting vortex regions in the flow field by determining characteristic data points that meet the vortex characteristics from sampling points evenly distributed in the flow field at intervals of time steps; at the same time, drawing a distribution map of vortex regions in the flow field, and obtaining characteristic attributes of each vortex region, including the size of the characteristic region, the center of the characteristic region, and a numerical matrix of the characteristic region; S2. According to the vortex area distribution map in the flow field, a feature matching bipartite graph G = (U∪V, E, w) is constructed to realize the construction of tracking relationships between feature areas at different time steps; wherein each feature area is represented by a feature node, based on which U represents the set of all feature nodes in time step t, and V represents the set of all feature nodes in time step t+1; the similarity value w of the correlation between two feature nodes; The calculation formula of w is: In the formula, u i ∩v j Indicates the number of feature data points that overlap between two feature nodes, Max(u i ,v j ) represents the maximum number of feature data points contained in two feature nodes; when w>0, it indicates that there is a correlation between the two feature nodes, then an edge E is inserted between the two nodes obtained by the above calculation, and the similarity value w is marked on it; when w<0, w is reassigned to 0; S3. Based on the optimal matching of the feature matching bipartite graph, complete the tracking of the largest components in the continuous events and split-merge events in the flow field feature tracking; the specific steps are: S301, based on the maximum similarity value w matching principle, the feature nodes in U are matched with the feature nodes in V in a one-to-one correspondence; S302, determining the final tracking result based on the characteristic events divided by the conflict domain; 1) Define candidate points and find the candidate point set of each feature node in the previous time step in the adjacent time step; when each fully matched feature node u in U i ∈U enumerates all points v in V that currently have no complete match j ∈V, and satisfies w(u,v)>0, then this type of point v j ∈V is called a candidate point; correspondingly, this type of point v j The set of nodes is called feature node u i The candidate point set of 2) Based on each complete match and its corresponding candidate point set, construct all possible split events and possible merge events; 3) According to the similarity calculation formula: Calculate the probability of all consecutive events, possible split events, and possible merge events; 4) Define conflict domain: Only one event can be selected from the possible events containing the feature node corresponding to any feature node at the same time step as the final tracking result. Therefore, the possible events containing the feature node are called a conflict domain; all possible events obtained according to the above step 2) are divided into several conflict domains, and the maximum value of the possibility calculation based on several possible events in the same conflict domain is selected as the tracking result, and the possible events corresponding to the same feature node as the tracking result in different conflict domains are deleted, and the possible events in each conflict domain are determined in turn; finally, the feature node of the unmatched event is determined as a new event or an extinction event according to its appearance time.
2. According to the flow field feature tracking method based on graph optimization according to claim 1, It is characterized in that In step S1, the method for extracting the vortex region in the flow field is: The vector at each sampling point in the flow field is usually represented by the following three-variable function: F(x,y,z)=(F 1 (w,y,z),F 2 (x,y,z),F 3 (x,y,z)), In the formula, F 1 , F 2 and F 3 They are the three-dimensional coordinate directions of F(x,y,z) in the Cartesian coordinate system. and The three elements on , then the gradient value is expressed as: The eddy current area is defined as S 2 +Ω 2 The two eigenvalues of are negative in the region, where S and Ω are the symmetric and antisymmetric parts of the Jacobian matrix vector field, that is: Calculate the vorticity value Q at each sampling point in the flow field to locate the vortex in the flow field: Q=0.5(||Ω|| 2 -||S|| 2 ), Among them, the larger the Q value is, the higher the vorticity value of the sampling point is, that is, the higher the probability of being detected as a vortex area in the flow field; Define the threshold Tho, and extract the sampling points that satisfy Q>Tho as feature data points to extract the vortices in the flow field.
3. The flow field feature tracking method based on graph optimization according to claim 2, It is characterized in that In step S1, the method for drawing the vortex area distribution map in the flow field is: randomly select a number of data points from all the feature data points as seed points, and maintain a queue based on each seed point; then, search the area adjacent to the seed point, push the feature data points that meet the threshold conditions into the corresponding queue, and mark the feature data points belonging to the same connected area with the same value until the search stops when the elements in the queue are empty; finally, the area corresponding to each queue is a feature area extracted from the flow field, that is, the vortex area in the flow field.
Citation Information
Patent Citations
Bayes-based vortex automatic tracking method
CN105894535A
Flow field vortex detection method based on neural network
CN111414720A
Multi-target tracking method, device and system and computer readable storage medium
CN112883819A
Turbulence simulation method
JP2017004103A
Spatio-temporal forecasting of future risk from past events
US20160104071A1
