A method for judging correspondence relation of dense crowd based on graph structure

Abstract

The application discloses a dense crowd corresponding relation judgment method based on a graph structure, which comprises a dense crowd distribution graph construction method, a pedestrian node feature descriptor construction method, a pedestrian node corresponding relation judgment method based on the graph structure and a judgment result correction method. The method solves the problem of how to quickly judge the monitoring conditions of the same target under different modes, that is, quickly judges the corresponding relation between dense crowds in multiple different mode sensor data.

Description

TECHNICAL FIELD

[0001] The present application relates to pedestrian re-identification technology, in particular to a dense crowd correspondence relationship judgment method based on a graph structure. BACKGROUND

[0002] In recent years, with the increasing perfection of the social security prevention and control system, pedestrian re-identification has become one of the popular applications in the field of computer vision, and its goal is to match pedestrians with the same identity in different sensor devices. Since different modal sensors include a large amount of information of monitored targets, which helps to improve the supervision of dense crowds and the risk handling capability, it is more and more common to use multiple sensors of different modalities to monitor dense crowds, such as using infrared and visible light cameras to monitor the same place. In this case, how to quickly judge the monitoring situation of the same target under different modalities requires a cross-modal dense crowd correspondence relationship judgment method, that is, to quickly judge the correspondence relationship between dense crowds in multiple different modal sensor data. The present technology realizes the judgment of the correspondence relationship of dense crowds in a cross-modal scene, that is, to quickly and accurately judge the correspondence relationship of multiple pedestrian targets in a dense scene.

[0003] Some people have proposed a dynamic refinement network (DRN) which contains two newly proposed component modules, namely a feature selection module (FSM) and a dynamic reference head (DRH).

[0004] Some people have proposed a multi-scale target detection method which effectively solves the problem of target detection in dense crowd scenes to a certain extent. The method is divided into coarse scale detection and fine scale detection. In the coarse scale, the whole body is detected, and the optimized DPM (Deformable Part Model) detection method is used to detect sparse targets in the whole scene; then in the fine scale, the head of the human body is detected, and in order to improve the accuracy, the Faster R-CNN (Region-based Convolutional Neural Network) network is used to detect the targets in the dense crowd.

[0005] Image matching technology is mainly applied in pose estimation and other aspects. In the last century, the research level of computer science was low, and the imaging tool was relatively backward, which led to the tortuous and slow progress of the research on image matching. Most of the traditional registration algorithms used gray gradient or Fourier transform domain. With the continuous improvement of device accuracy and the rapid development of computer technology, feature-based image matching technology appears and is applied more and more in actual occasions.

[0006] In 2004, SIFT (Scale-Invariant Feature Transform) algorithm was proposed, which has obvious advantages in feature-based algorithms. SIFT algorithm has the characteristics of scale invariance and is not affected by factors such as light and noise to a certain extent. SIFT algorithm is widely used in image feature extraction and has excellent performance in image matching application. Subsequently, on the basis of SIFT algorithm, numerous improved algorithms were born like bamboo shoots after a spring rain, which all have excellent performance in a certain field. Representative algorithms include SURF, BRISK, Kaze and the like.

[0007] The correspondence relationship of the cross-modal scene of the dense crowd aims to find the one-to-one correspondence relationship of the individuals in the dense crowd in two modalities. The existing dense crowd detection technology can only locate the position information of the pedestrian targets in different modalities, but cannot obtain the correspondence relationship of the pedestrians in different modalities. Image registration cannot locate all pedestrians, which greatly hinders the subsequent correspondence relationship judgment. SUMMARY

[0008] The main purpose of the present application is to provide a dense crowd correspondence relationship judgment method based on graph structure, which solves the problem of how to quickly judge the monitoring situation of the same target in different modalities, that is, quickly judge the correspondence relationship between the dense crowds in multiple different modal sensor data.

[0009] The technical scheme adopted by the present application is: a dense crowd correspondence relationship judgment method based on graph structure, comprising:

[0010] A dense crowd distribution graph construction method;

[0011] A pedestrian node feature descriptor construction method;

[0012] A pedestrian node correspondence relationship judgment method based on graph structure;

[0013] A judgment result correction method.

[0014] Further, the dense crowd distribution graph construction method comprises:

[0015] A method for constructing dense population distribution maps based on node neighborhoods is adopted. The formula for calculating node neighborhoods is as follows:

[0016] (1)

[0017] In the formula:

[0018] — Node neighborhood radius;

[0019] — Pixel width of images of densely populated scenes;

[0020] — The pixel height of the image;

[0021] — The total number of densely populated pedestrian nodes in this scene;

[0022] After determining the neighborhood radius, the neighborhood of each pedestrian node can be calculated, and the neighboring nodes can be identified.

[0023] Each node in the domain; by connecting each pedestrian node to other nodes in its neighborhood one by one, we can obtain a dense crowd distribution map for this scene.

[0024] Furthermore, the method for constructing the pedestrian node feature descriptor includes:

[0025] After constructing the dense crowd distribution map, feature descriptors for pedestrian nodes were built. Constructing feature descriptors for pedestrian nodes is an important prerequisite for obtaining the features of the nodes. The specific representation method is shown below:

[0026]

[0027] in, The degree of a node representing a pedestrian node. This represents the shortest path pass rate. Represents the distance from the central rate. Represents the value of neighboring nodes. The five features, along with their calculation formulas, represent the importance of nodes.

[0028] 1) Pedestrian node degree Node degree This is used to represent the number of nodes directly connected to a pedestrian node in a dense crowd distribution map; as can be seen from the construction process of the dense crowd distribution map, the degree of a pedestrian node is... It is also equal to the total number of nodes in the neighborhood of that node; for a node, the total number of nodes is A dense crowd distribution map, showing a specific pedestrian node. Node degree The calculation formula is shown as formula (2):

[0029] (2)

[0030] In the formula:

[0031] — constant, indicating whether the i-th node and the j-th node are directly connected, if directly connected, then ; if not directly connected, then ;

[0032] — total number of nodes;

[0033] 2) Shortest path passing rate of pedestrian node The shortest path passing rate of a pedestrian node refers to the ratio of the number of shortest paths passing through the node to the total number of shortest paths in the entire distribution map; the shortest path passing rate The calculation formula is shown as formula (3):

[0034] (3)

[0035] In the formula:

[0036] — total number of paths passing through node i in the shortest paths of other nodes;

[0037] — total number of shortest paths in the entire dense crowd distribution map;

[0038] 3) Distance center rate of pedestrian node The distance center rate of a pedestrian node refers to the inverse of the average value of the inverse tangent of the shortest path length from the node to all other nodes; for a dense crowd distribution map with a total number of nodes , the distance center rate of a certain pedestrian node i is shown as formula (4):

[0039] (4)

[0040] In the formula:

[0041] — length of the shortest path between node i and node j;

[0042] ​​​​​​​4) neighborhood node value of pedestrian node The neighborhood node value of a certain node is based on its node degree The neighborhood node value matrix is calculated The calculation formula is shown in (5):

[0043] (5)

[0044] In the formula:

[0045] The neighborhood matrix of dense crowd The order of

[0046] C - node degree matrix (C), that is, the value of the first row and the first column in the matrix represents the node degree of the first node;

[0047] Node importance of pedestrian node .

[0048] Further, the pedestrian node correspondence relationship determination method based on the graph structure comprises:

[0049] Taking a local-to-global idea, the pedestrian nodes in the local area are determined first, and then the pedestrian nodes in the global area are determined; after obtaining the features of each node, the following method is adopted to start screening and matching the nodes:

[0050] First, the quasi-core nodes, i.e., the pedestrian nodes in the local area, are screened out; the determination formula of whether a certain node is a quasi-core node is shown in (6):

[0051] (6)

[0052] In the formula:

[0053] Feature weight

[0054] Feature of node

[0055] Average feature of corresponding scene If

[0056] <0, it is determined that the node is not a quasi-core node; if >0, it is determined that the node is a quasi-core node

[0057] ​​Next, each quasi-core node is matched to obtain the judgment result of the local pedestrian node correspondence;

[0058] Then, starting from the local correspondences, the global correspondence results are obtained; this step is completed using a dense population correspondence interpretation method based on region growth, which specifically includes the following steps:

[0059] The core nodes in the two scenarios are reconciled according to the original relationships in the dense crowd distribution map to form two core local structure graphs; each core local structure graph only includes core nodes and edges;

[0060] For each scenario, the non-core nodes that are directly connected to the core nodes are first matched, and then added to the corresponding core local structure graph in turn.

[0061] Repeat the previous step until no new nodes can be added.

[0062] Furthermore, the method for correcting the interpretation result includes:

[0063] The judgment result is corrected by using the position angle of the pedestrian node. First, the formula for calculating the position angle is given:

[0064] In their respective coordinate systems, the angle between the positions of the two pedestrian nodes The calculation method is shown in (7):

[0065] (7)

[0066] In the formula:

[0067] — The vertical coordinate of the pixels in the scene of the second pedestrian node;

[0068] — The vertical coordinate of the pixels in the scene of the first pedestrian node;

[0069] — The x-coordinate of the pixels in the scene of the second pedestrian node;

[0070] — The x-coordinate of the pixels in the scene of the first pedestrian node.

[0071] Advantages of this invention:

[0072] The method of the present application plays a significant role in the construction and perfection of social security prevention and control system, and is beneficial to maintaining social stability and promoting social harmonious development. Most of the factors leading to the danger of dense crowds come from the inside of the dense crowds, and in most cases, a few individuals appear panic and spread the psychology, thus leading to large-scale out of control and causing a large number of casualties. The method has practical value in analyzing the trend of multiple behavioral individuals in dense crowds and preventing major public safety accidents, and is closely related to the construction of a safe city and the realization of intelligent monitoring of dense crowds and handling of public dangers.

[0073] In addition to the purposes, features and advantages described above, the present application has other purposes, features and advantages. The present application will be further described below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0074] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, illustrate the preferred embodiments of the application, and assist in the explanation of the application. They should not be considered as limiting the present application.

[0075] Figure 1 is a flow chart of the dense crowd corresponding relationship interpretation method based on the graph structure of the present application;

[0076] Figure 2 is a generated cross-modal dense crowd distribution map of the present application;

[0077] Figure 3 is a flow chart of the node importance calculation method of the present application

[0078] Figure 4 is a specific registration flow chart of the quasi-core node of the present application

[0079] Figure 5 is a flow chart of the result correction method of the present application;

[0080] Figure 6 is a schematic diagram of each pedestrian node and neighborhood of the present application;

[0081] Figure 7 is a cross-modal corresponding relationship interpretation result map of the present application. DETAILED DESCRIPTION

[0082] In order to make the purposes, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0083] The problem solved by the graph structure-based dense crowd correspondence interpretation method of the present application is how to quickly judge the monitoring situation of the same target under different modalities, that is, quickly judge the correspondence between dense crowds in multiple different modal sensor data. In the process of solving this problem, the difficulties and challenges encountered include the following aspects:

[0084] 1) The universality problem of cross-modal interpretation. The research method of this paper should minimize the influence of the difference in visual features of the target under different modalities. Although the current cross-modal monitoring system is mostly a multi-modal monitoring system composed of infrared cameras (or thermal imagers) and visible light monitoring cameras (including black and white and color cameras), with the development and progress of science and technology, more and more visual modalities will emerge. At the same time, as human needs continue to increase, the application scenarios of this topic will become increasingly rich and complex. Therefore, in order to ensure universality, the method of this paper should minimize the dependence on modalities and minimize the impact of different modalities in the research process as much as possible.

[0085] 2) The problem of interpreting the same pedestrian in different positions with binocular disparity. It goes without saying that when different modal cameras at the same height shoot the same scene, there will be a binocular perspective difference, which may cause many influences including the following situations:

[0086] (1) The same pedestrian is in different positions in the images presented by different modalities;

[0087] (2) Due to the difference caused by different perspectives, some pedestrians present in the image of one modality are missing in the image of another modality; and the image under another modality will usually have some pedestrians. This is particularly true for pedestrians in or near the edge zone of the scene.

[0088] The algorithm uses the distribution structure features between nodes in different scenes for matching. This method first constructs dense crowd distribution graphs, then builds pedestrian node feature descriptors according to these distribution graphs, and then interprets the correspondence of dense crowds. The framework diagram of this method is shown in Figure 1 . In Figure 1 , the purpose of the dense crowd pedestrian target detection algorithm is to obtain the positions of all pedestrians in the dense scene, and this algorithm can be realized by a deep learning target detection algorithm.

[0089] (1) Dense crowd distribution graph construction method:

[0090] This method uses a dense crowd distribution graph construction method based on node neighborhood. The calculation formula of the node neighborhood is as follows:

[0091] (1)

[0092] In the formula:

[0093] — Node neighborhood radius;

[0094] — The pixel width of images of densely populated scenes (i.e., images captured by surveillance cameras);

[0095] — The pixel height of the image;

[0096] — The total number of densely populated pedestrian nodes in this scene.

[0097] Once the neighborhood radius is determined, the neighborhood of each pedestrian node can be calculated (see reference). Figure 6 This involves identifying each pedestrian node and its neighboring nodes. Connecting each pedestrian node to all other nodes in its neighborhood yields a dense crowd distribution map for that scene. For example... Figure 2 The image shown is an example of a dense population distribution map and a schematic diagram.

[0098] (2) Pedestrian node feature descriptor construction method:

[0099] After constructing the dense crowd distribution map, this section builds the feature descriptors for pedestrian nodes. Constructing these feature descriptors is a crucial prerequisite for obtaining node features and forms the basis for subsequent node interpretation. The specific representation method is shown below:

[0100]

[0101] in, The degree of a node representing a pedestrian node. This represents the shortest path pass rate. Represents the distance from the central rate. Represents the value of neighboring nodes. This indicates the importance of a node. The five features and their calculation formulas are as follows:

[0102] 1) Pedestrian node degree Node degree This represents the number of nodes directly connected to a pedestrian node in a dense crowd distribution map. From the construction process of the dense crowd distribution map, we know that the degree of a pedestrian node... This is also equal to the total number of nodes in the neighboring region of that node. For a given node, the total number of nodes is... A dense crowd distribution map, showing a specific pedestrian node. Node degree The calculation formula is shown in (2):

[0103] (2)

[0104] wherein:

[0105] — constant, indicating whether the th node is directly connected with the th node (i.e. without passing through any intermediate node), if directly connected, then ; if not directly connected, then ;

[0106] — total number of nodes.

[0107] 2) Shortest path passing rate of pedestrian node The shortest path passing rate of a pedestrian node refers to the ratio of the number of shortest paths passing through the node to the total number of shortest paths in the whole distribution map. The calculation formula of the shortest path passing rate is shown in (3):

[0108] (3)

[0109] wherein:

[0110] — the total number of paths passing through the node from the shortest paths of other nodes;

[0111] — the total number of shortest paths in the whole dense crowd distribution map. 3) Distance centrality of pedestrian node

[0112] The distance centrality of a pedestrian node refers to the inverse of the average value of the inverse tangent of the shortest path length from the node to all other nodes. For a dense crowd distribution map with a total number of nodes , the calculation formula of the distance centrality of a certain pedestrian node is shown in (4):

[0113] (4) wherein:

[0114]

[0115] — the length of the shortest path between the node and the node . 4) Neighborhood node value of pedestrian node

[0116] ​The value of a node's neighboring nodes is based on the node's degree. Perform calculations. Neighborhood node value matrix. The calculation formula is shown in (5):

[0117] (5)

[0118] In the formula:

[0119] — Adjacency matrix of densely populated populations of Rank;

[0120] C — Node degree matrix ( ), that is, the first in the matrix The value represents the first... The degree of each node.

[0121] 6) Importance of pedestrian nodes The method for calculating the node importance of L is as follows: Figure 3 As shown:

[0122] S1, set the initial node importance Lo=1;

[0123] S2, calculate the degree C of each pedestrian node in the graph;

[0124] S3. Determine if there are pedestrian nodes with C ≤ Lo. If so, set the importance L of these nodes to Lo, delete them, and recalculate the C value of each pedestrian node in the remaining graph structure. Otherwise, set Lo = Lo + 1, and determine if there are still pedestrian nodes in the dense crowd distribution graph.

[0125] If yes, proceed to step S2; otherwise, output the node importance result.

[0126] (1) A method for determining the correspondence between pedestrian nodes based on graph structure:

[0127] This method adopts a local-to-global approach, first analyzing pedestrian nodes in local areas, and then analyzing pedestrian nodes globally. After obtaining the features of each node, the following method is used to filter and match the nodes.

[0128] First, identify the quasi-core nodes, i.e., pedestrian nodes in local areas. Then, determine the location of a specific node. The formula for determining whether a node is a quasi-core node is shown in (6).

[0129] (6)

[0130] In the formula:

[0131] feature weight;

[0132] feature of the node ;

[0133] average feature of the corresponding scene;

[0134] If <0, it is determined that the node is not a quasi-core node; if >0, it is determined that the node is a quasi-core node.

[0135] Then, the respective quasi-core nodes are matched to obtain the interpretation result of the local pedestrian node correspondence. The specific matching process is shown in the flowchart of Figure 4 ;

[0136] Then, the global correspondence result is obtained from the local correspondence. In this method, a dense crowd correspondence interpretation method based on region growing is used to complete this step, which is specifically divided into the following steps:

[0137] 1) The core nodes in the two scenes are constructed into two core local structure graphs according to the original association relationship in the dense crowd distribution graph. In each core local structure graph, only core nodes and edges are included.

[0138] 2) The non-core nodes directly connected to the core nodes in each scene are matched first, and then are sequentially added to the corresponding core local structure graph. The matching method is the same as the flowchart. Figure 4

[0139] 3) Repeat 2) until no new node can be added.

[0140] At this point, the preliminary pedestrian node correspondence is obtained by the method. In order to increase the accuracy of the interpretation result, a method for correcting the interpretation result needs to be designed.

[0141] (4) Interpretation result correction method:

[0142] The position angle of the pedestrian node is used to correct the interpretation result. First, the calculation formula of the position angle is given:

[0143] In the respective coordinate system, the calculation method of the position angle of two pedestrian nodes is shown in (7):

[0144] (7)

[0145] In the formula, the position angle of the pedestrian node is calculated as follows:

[0146] ​ — the pixel vertical coordinate in the second pedestrian node scene;

[0147] — the pixel vertical coordinate in the first pedestrian node scene;

[0148] — the pixel horizontal coordinate in the second pedestrian node scene;

[0149] — the pixel horizontal coordinate in the first pedestrian node scene.

[0150] The specific method of position angle correction is shown in the flowchart Figure 5 .

[0151] The dense crowd correspondence relationship interpretation based on the graph structure mainly includes the following steps: constructing a dense crowd distribution graph, constructing a pedestrian node feature descriptor, interpreting the pedestrian node correspondence relationship based on the graph structure, and correcting the correspondence result.

[0152] 1) Constructing a dense crowd distribution graph: regarding the dense crowd as a zero graph, regarding the pedestrian as a node in the zero graph, and connecting part of the nodes in the zero graph into a dense crowd distribution graph.

[0153] 2) Constructing a pedestrian node feature descriptor: constructing a node feature descriptor for each node in the dense crowd distribution graph.

[0154] 3) Pedestrian node correspondence relationship interpretation method based on the graph structure: after obtaining the feature descriptor of the pedestrian node in the dense crowd distribution graph, the features of each node are calculated according to the feature descriptor, and the nodes with similar features in different modalities are corresponded. This method first selects the core nodes in the pedestrian nodes, and then matches the core nodes, and then uses the dense crowd correspondence relationship interpretation method based on region growing or the dense crowd correspondence relationship interpretation method based on multi-hypothesis graph to match as many pedestrian nodes as possible.

[0155] 4) Correcting the matching result, the purpose of this step is to ensure the accuracy of the matching result, so as to obtain a more accurate interpretation result of the dense crowd in the cross-modal scene (for reference Figure 7 ).

[0156] The dense crowd correspondence relationship interpretation method based on the graph structure uses a target detection algorithm to obtain the position information of all pedestrians, and then performs subsequent registration steps, so it mainly involves target detection technology and image matching technology.

[0157] Target detection plays an important role in computer vision and pattern recognition, which aims to detect specified targets in various scene images, mark the target positions, and label the corresponding categories for each detected target. The detection of dense targets is always a popular topic in computer vision. There are many applications in reality. For example, the detection of dense crowds, the detection and identification of dense vehicles, etc.

[0158] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A dense crowd correspondence determination method based on a graph structure, characterized in that it comprises: a dense crowd distribution graph construction method; a pedestrian node feature descriptor construction method; a pedestrian node correspondence determination method based on a graph structure; a determination result correction method; the pedestrian node feature descriptor construction method comprises: after the construction of the dense crowd distribution graph is completed, a feature descriptor of a pedestrian node is constructed; the construction of the feature descriptor of the pedestrian node is an important prerequisite for obtaining the features of the node, and the specific representation method is as shown below: in the formula: ； wherein, denotes the node degree of a node representing a pedestrian node, denotes the shortest path passing rate, denotes the distance center rate, denotes the neighborhood node value, denotes the node importance degree; the five features and the calculation formula are as follows: 1) Pedestrian node degree : Node degree The number of nodes directly connected to a certain pedestrian node in the dense crowd distribution map; from the construction process of the dense crowd distribution map, the pedestrian node degree is also equal to the total number of nodes in the neighborhood of the node; for a dense crowd distribution map with a total number of nodes , the node degree of a certain pedestrian node is calculated as shown in equation (2): （2） in the formula: — Constant, indicating whether the th node is directly connected to the th node, if directly connected, then ; if not directly connected, then ; — total number of junctions; 2) Shortest path passing rate of pedestrian node The shortest path passing rate of a certain pedestrian node refers to the ratio of the number of shortest paths passing through the node to the total number of shortest paths in the entire distribution graph. The shortest path passing rate The calculation formula of the shortest path passing rate is shown in equation (3): （3） in the formula: — the number of shortest paths from node to other nodes, passing through node ; - the total number of shortest paths in the entire dense crowd distribution map; 3) the distance centrality of a pedestrian node The distance centrality of a pedestrian node is defined as the inverse of the average of the inverse tangent of the shortest path length from the node to all other nodes. For a dense crowd distribution map with a total number of nodes , the distance centrality of a pedestrian node is calculated as shown in equation (4): ​ （4） in the formula: —— node to the node between the shortest path of the length; 4) Neighborhood node value of a pedestrian node The neighborhood node value of a certain node is based on its node degree The calculation is performed; neighborhood node value matrix The calculation formula is shown in (5): （5） S1, set an initial node importance Lo=1; - Adjacency matrix of dense crowds of order; C — Node degree matrix ( ), that is, the first in the matrix The value represents the first... The degree of each node; Node importance of a pedestrian node The node importance calculation method of L is: S2, calculate the node degree C of each pedestrian node in the graph; S3, determine whether there is a pedestrian node with C≤Lo, if yes, set the node importance L of the node with C≤Lo to Lo, delete these nodes, and recalculate the C value of each pedestrian node in the remaining graph structure; if not, Lo=Lo+1, and determine whether there is still a pedestrian node in the dense crowd distribution graph, if yes, execute step S2, and if not, output the node importance result; the pedestrian node correspondence determination method based on a graph structure comprises: adopting a local-to-global idea, the pedestrian nodes in a local area are determined first, and then the pedestrian nodes in the global are determined; after obtaining the features of each node, the following method is adopted to start screening and matching the nodes: in the formula: First, the quasi-core nodes, i.e. the pedestrian nodes in local areas, are selected. The formula for determining whether a node is a quasi-core node is shown in (6): (6) （6） then, each quasi-core node is matched to obtain the determination result of the local pedestrian node correspondence; - feature weights; - knots characteristics; — average features of the respective scene; If <0, then determine that the node is not a quasi-core node; if >0, then determine that the node is a quasi-core node; then, the global correspondence result is obtained from the local correspondence; the dense crowd correspondence determination method based on region growing is adopted to complete this step, which comprises the following steps: the core nodes in the two scenes are connected according to the original association relationship in the dense crowd distribution graph to form two core local structure graphs; each core local structure graph only includes core nodes and edges; the non-core nodes directly connected to the core nodes in each scene are matched first, and then are sequentially added to the corresponding core local structure graph; repeat the previous step until no new node can be added. the dense crowd distribution graph construction method comprises:

2. The graph structure-based dense crowd correspondence interpretation method of claim 1, wherein, the dense crowd distribution graph construction method based on the node neighborhood is adopted, and the calculation formula of the node neighborhood is as follows: in the formula: （1） after the neighborhood radius is determined, the neighborhood of each pedestrian node can be obtained, and each node in the neighborhood can be found; each pedestrian node and other nodes in its neighborhood are connected one by one to obtain the dense crowd distribution graph in the scene. — node neighborhood radius; — the pixel width of the dense crowd scene image; — the pixel height of the image; - the total number of dense crowd pedestrian junctions contained in the scene; the determination result correction method comprises:

3. The method according to claim 1, wherein, the position angle of the pedestrian node is used to correct the determination result, and the calculation formula of the position angle is as follows: in the formula: The position of two pedestrian nodes in the respective coordinate system is calculated by the included angle as shown in (7): （7） ​ — pixel ordinate in the second pedestrian junction scenario; — the vertical coordinate of the pixel in the first pedestrian junction scenario; — pixel abscissa in the second pedestrian junction scenario; - the pixel abscissa in the first pedestrian junction scenario.

Images