A Multi-Target Association Method Based on Clustered Images

Abstract

This invention discloses a multi-target association method based on clustered images, relating to the field of multi-target association. It employs a target recognition technology based on the DBSCAN algorithm to cluster the dataset according to density, aggregating points with similar densities into the same target cluster. This algorithm does not require a pre-defined number of generated targets, adapts to arbitrary shape distributions, exhibits good stability, and possesses strong noise resistance. It can effectively filter stray signals even without prior information, automatically achieving target recognition and segmentation, and obtaining the number of targets and spatiotemporal localization results. Compared with traditional parametric recognition methods, it can automatically achieve target association, overcome dependence on prior information, reduce the impact of subjective factors in human interpretation on the association effect, and demonstrates better target recognition and association effects in multi-target association scenarios, improving the efficiency of multi-target association. It is suitable for widespread promotion and application.

Description

Technical Field

[0001] This invention relates to the field of multi-object association, and more specifically to a multi-object association method based on clustered images. Background Technology

[0002] As is known, a light curtain target is an instrument that uses photoelectric conversion technology to detect the arrival time of a flying target at a designated location in space. Two arrayed light curtain targets work in conjunction with a timing instrument to calculate the time it takes for the target to pass over the two targets based on the signals output by the two targets. The average velocity of the target as it passes over the two targets can be calculated based on the distance between them. However, when multiple targets pass through the light curtain target simultaneously, the problem of multi-target correlation between the preceding and following targets must first be solved in order to effectively calculate the velocities of different targets.

[0003] Currently, the commonly used multi-target association method is the parametric recognition method. The core of this method is to determine the parameters of the generated pulse signals based on prior information about the targets (target shape, size, velocity range, etc.). Since multiple targets have different sizes and velocities, different pre-set rule parameters (e.g., velocity range, contact time, similarity of target position coordinates before and after) are manually preset to filter target signals, determining pulse signals belonging to the same target, and further calculating the target's velocity. This method is highly dependent on the accuracy of prior knowledge. Manually filtering signals using prior information means that a lack of prior conditions or biased prior information directly affects the accuracy of the recognition results, making it impossible to correctly identify valid targets. Furthermore, manually associating targets requires extensive human experience. When multiple targets pass through the target surface simultaneously, there are various combinations of target position coordinates before and after the targets, requiring complex rule settings, leading to low efficiency and decreased accuracy in manual interpretation.

[0004] Therefore, it is particularly important to provide a multi-objective association method based on clustered images. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a multi-target association method based on clustered images. The present invention can accurately and efficiently perform multi-target association under the condition of lack of prior information.

[0006] To achieve the aforementioned inventive objectives, the present invention employs the following technical solution:

[0007] A multi-objective association method based on clustered images, the multi-objective association method specifically includes the following steps:

[0008] The first step is to obtain the time points and coordinates in the x and y directions of the pulse signals generated by the front and rear targets.

[0009] The second step involves using the target recognition technology based on the DBSCAN algorithm on the data obtained in the previous step. Each cluster generated is used as a target for recognition, and the time and coordinate range of each cluster are calculated as parameters of the target.

[0010] The third step is to draw a spatiotemporal distribution map of the same size for each target individually.

[0011] Step 4: Perform image preprocessing on the spatiotemporal distribution maps of multiple targets in front and behind the target.

[0012] Step 5: When there is more than one current and subsequent target, use the Kuhn-Munkres algorithm for multi-target association.

[0013] The multi-target association method based on clustered images, in the first step, converts the data into a two-dimensional vector consisting of the sum of coordinates in the time, x, and y directions, reducing computational complexity and improving computational speed without affecting the accuracy of target range identification.

[0014] In the multi-target association method based on clustered images, each data point in the pulse signals of the front and rear targets in the first step is in vector form, containing the time and position information of the target passing through the front and rear targets.

[0015] The multi-target association method based on clustered images, specifically includes the following steps in the second step of target recognition based on the DBSCAN algorithm:

[0016] The first step is to set the parameters ε and M of the DBSCAN algorithm, where ε is the neighborhood radius and M is the minimum number of samples in the v neighborhood of a sample point. This algorithm is based on sample density clustering. When the number of samples in the neighborhood of a sample point exceeds the threshold M, it is added to the cluster with similar samples. Setting different parameters will affect the number of clusters generated, i.e. the number of targets identified.

[0017] Based on the distribution of the experimental data sample points, ε and M are set. When ε increases, the number of generated clusters may decrease, and vice versa. When M is too large, the number of clusters increases, and vice versa. When M is too small, the number of clusters will be too small.

[0018] The second step is to determine the formula for calculating the distance between two points in the DBSCAN algorithm. Usually, the Euclidean distance formula can be chosen.

[0019] The third step is to cluster the front and rear target data according to the set parameters. Each cluster generated is a valid target. The output of the front and rear target clustering results is used to calculate the time and coordinate range of each cluster as the positioning parameters of each target.

[0020] The multi-target association method based on clustered images, in the third step, the spatiotemporal distribution map is a graphic drawn in a two-dimensional planar manner using the target time and coordinate range obtained by clustering. This graphic can reflect the time range and coordinate range of the target signal.

[0021] The multi-target association method based on clustered images, in the fourth step of image preprocessing, includes generating a grayscale image from the original image, binarizing the grayscale image, calculating the Hu moments of the target images, and obtaining possible associated target combinations according to the time sequence.

[0022] The multi-target association method based on clustered images, in its fifth step using the Kuhn-Munkres algorithm for multi-target association, specifically includes the following steps:

[0023] The first step is to use the front and rear target groups as the left and right vertices of the Kuhn-Munkres algorithm, respectively, with each vertex representing a target;

[0024] When the number of targets in the front and rear targets is not equal, virtual vertices need to be added to make the number of targets in the front and rear targets equal.

[0025] The second step is to filter out target combinations that may have a relationship based on the time order in which the preceding and following targets appear, and set them as edges in the Kuhn-Munkres algorithm.

[0026] The third step is to calculate the Hu moments of the target image. For target combinations that are related, calculate the target association value based on the Hu moments. The greater the similarity, the smaller the target association value. The calculation result of the target association value is not affected by the displacement and rotation of the image.

[0027] Step 4: Set the vertex weights and the weights of all edges connected to that vertex in the Kuhn-Munkres algorithm based on the target association value;

[0028] Step 5: Construct a weighted bipartite graph according to the set vertex weights and edge weights, and use the Kuhn-Munkres algorithm to perform maximum matching. The maximum matching is the matching with the most edges and the largest sum of weights among all matchings. Output the association results.

[0029] By employing the technical solution described above, the present invention has the following advantages:

[0030] Compared with traditional parametric recognition methods, this invention can automatically achieve target association, overcome the dependence on prior information, reduce the impact of subjective factors of human interpretation on the association effect, and has better target recognition and association effects in multi-target association scenarios, improving the efficiency of multi-target association, etc., making it suitable for widespread promotion and application. Attached Figure Description

[0031] Figure 1 This is a flowchart of the multi-target association process in this invention;

[0032] Figure 2 This is a schematic diagram of the overall spatiotemporal distribution of the target group generated based on the clustering results in an embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram of the spatiotemporal distribution of the target in the embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of the spatiotemporal distribution of the target in an embodiment of the present invention. Detailed Implementation

[0035] The present invention can be explained in more detail through the following embodiments, but the present invention is not limited to the following embodiments;

[0036] Combined with appendix Figures 1-4 The multi-objective association method based on clustered images includes the following steps:

[0037] The first step is to obtain the time points and x and y coordinates of the pulse signals generated by the front and rear targets. In practice, the data is converted into a two-dimensional vector summing the time, x, and y coordinates to reduce computational complexity and improve computational speed without affecting the accuracy of target range identification. Each data point in the pulse signals of the front and rear targets is in vector form and contains the time and position information of the target passing through the front and rear targets.

[0038] The second step involves applying the DBSCAN algorithm-based target recognition technology to the data obtained in the previous step. Each generated cluster is treated as a target for recognition, and the time and coordinate range of each cluster are calculated as parameters of the target. In practice, the target recognition based on the DBSCAN algorithm specifically includes the following steps:

[0039] The first step is to set the parameters ε and M of the DBSCAN algorithm, where ε is the neighborhood radius and M is the minimum number of samples in the ε neighborhood of a sample point. This algorithm is based on sample density clustering. When the number of samples in the neighborhood of a sample point exceeds the threshold M, it is added to the cluster with similar samples. Setting different parameters will affect the number of clusters generated, i.e. the number of targets identified.

[0040] Based on the distribution of the experimental data sample points, ε and M are set. When ε increases, the number of generated clusters may decrease, and vice versa. When M is too large, the number of clusters increases, and vice versa. When M is too small, the number of clusters will be too small.

[0041] The second step is to determine the formula for calculating the distance between two points in the DBSCAN algorithm. Usually, the Euclidean distance formula can be chosen.

[0042] The third step is to cluster the front and rear target data according to the set parameters. Each cluster generated is a valid target. The output of the front and rear target clustering results is used to calculate the time and coordinate range of each cluster as the positioning parameters of each target.

[0043] The third step is to draw a spatiotemporal distribution map of the same specifications for each target. In practice, the spatiotemporal distribution map is a graphic drawn in a two-dimensional planar manner using the target time and coordinate range obtained by clustering. This graphic can reflect the time range and coordinate range of the target signal.

[0044] The fourth step is to perform image preprocessing on the spatiotemporal distribution maps of multiple targets in front and behind the target; the image preprocessing includes generating grayscale images from the original images, binarizing the grayscale images, calculating the Hu moments of the target images, and obtaining possible related target combinations according to the time sequence.

[0045] Step 5: When there is more than one current and subsequent target, use the Kuhn-Munkres algorithm for multi-target association. In practice, the use of the Kuhn-Munkres algorithm for multi-target association specifically includes the following steps:

[0046] The first step is to use the front and rear target groups as the left and right vertices of the Kuhn-Munkres algorithm, respectively, with each vertex representing a target;

[0047] When the number of targets in the front and rear targets is not equal, virtual vertices need to be added to make the number of targets in the front and rear targets equal.

[0048] The second step is to filter out target combinations that may have a relationship based on the time order in which the preceding and following targets appear, and set them as edges in the Kuhn-Munkres algorithm.

[0049] The third step is to calculate the Hu moments of the target image. For target combinations that are related, calculate the target association value based on the Hu moments. The greater the similarity, the smaller the target association value. The calculation result of the target association value is not affected by the displacement and rotation of the image.

[0050] Step 4: Set the vertex weights and the weights of all edges connected to that vertex in the Kuhn-Munkres algorithm based on the target association value;

[0051] Step 5: Construct a weighted bipartite graph according to the set vertex weights and edge weights, and use the Kuhn-Munkres algorithm to perform maximum matching. The maximum matching is the matching with the most edges and the largest sum of weights among all matchings. Output the association results.

[0052] Combined with appendix Figure 1 The specific embodiments of the present invention are as follows:

[0053] The first step is to obtain the target dataset, including the time points when the pulse signals from the front and rear targets are generated, and the sum of the x and y coordinates corresponding to each time point. This is specifically represented as a two-dimensional vector, denoted as...

[0054] (x+y,t)

[0055] The second step involves using the DBSCAN algorithm-based target recognition technology on the data from the previous step. Each cluster generated is treated as a target for recognition, and the time and coordinate range of each cluster are calculated as parameters of the target.

[0056] 1) Based on the sampling frequency, target size and target signal density, set the parameters ε and M of the DBSCAN algorithm, where ε is the neighborhood radius and M is the minimum number of samples in the ε neighborhood of a certain sample point;

[0057] (ii) Determine the distance calculation formula between two points in the DBSCAN algorithm, and choose the Euclidean distance formula. For example, the distance between the i-th sample and the j-th sample is...

[0058] Distance(i,j)=||(x i +y i , t i )-(x j +y j , t j )||;

[0059] (iii) Cluster the pre- and post-target data separately, with each category representing a valid target, and output the pre- and post-target clustering results; for example, the clustering results of a certain experiment are as follows: Figure 2 As shown;

[0060] The third step is to draw a spatiotemporal distribution map of the same size for each target individually.

[0061] The spatiotemporal distribution map is drawn using a two-dimensional planar diagram. This map can represent the time and coordinate range of the target signal. The spatiotemporal distribution maps of the target before and after are shown below. Figure 3 , Figure 4 As shown;

[0062] Step 4: Perform image preprocessing on the spatiotemporal distribution maps of multiple targets in front of and behind the target.

[0063] The image preprocessing report includes: generating a grayscale image from the original image, and then performing binarization on the grayscale image. The purpose of this processing is to separate the target and the background in the image, and at the same time improve the speed of subsequent image processing.

[0064] Step 5: When there is more than one current and subsequent target, use the Kuhn-Munkres algorithm for multi-target association;

[0065] The target groups before and after are respectively used as the left and right vertex sets of the Kuhn-Munkres algorithm, with each vertex representing a target, denoted as .

[0066] U = {u1, ..., u} n},

[0067] V = {v1, ..., v} n};

[0068] When the number of targets in the front and rear targets is not equal, virtual vertices are added to make the number of targets in the front and rear targets equal.

[0069] Calculate the Hu moments for each target image before and after the target image. Image Hu moments are image features that are invariant to translation, rotation, and scale. Seven Hu moments can be calculated for each binary image; these remain invariant under image translation, rotation, and scale changes. Let the Hu moments of target u be denoted as... The Hu moment of the target v is Where k = 1, ..., 7;

[0070] Calculate the target correlation value I for the target combination. Let the formula for calculating the correlation value of target u and target v be:

[0071]

[0072] in

[0073]

[0074]

[0075] The smaller the I value, the higher the target correlation. The sum of the correlation values ​​for all target combinations is calculated and denoted as I. total ;

[0076] Connecting the two vertices of the front and rear targets forms an edge. For each edge (u i v j ), with weight w(u) i v j ), where the edge weight between vertices with no relation is 0, w(u i v j The calculation formula is as follows: w(u) i v j The larger u is i and v j The greater the correlation;

[0077]

[0078] The association relationship is determined based on the time order of appearance of the preceding and following targets. If the appearance time of a subsequent target is less than that of a preceding target, there is no association relationship, the edge weight is set to 0, and it will not be included in the matching results. For example, in this embodiment, the appearance times of all subsequent targets are greater than those of the preceding targets, so there may be association relationships between the targets.

[0079] Let the weight of the front target vertex be the maximum weight among all edges connected to that vertex, the weight of the back target vertex be 0, and the weight of the virtual vertex be assigned to 0, denoted as .

[0080] l(u i )=max{w(u i v i )|v i ∈V},u i ∈U

[0081] l(v i ) = 0, v i ∈V

[0082] Construct a weighted bipartite graph based on the set vertex and edge weights, denoted as .

[0083] G(U,V), where U={u1,...,u n}, V = {v1, ..., v n};

[0084] The Kuhn-Munkres algorithm is used to perform target association and find the maximum matching of G(U,V). The maximum matching is the matching with the most edges and the largest sum of weights among all matchings. The maximum matching result is output as the target association result.

[0085] Compared with the prior art, the present invention has the following advantages:

[0086] (1) It overcomes the dependence on prior information and eliminates the need for manual signal filtering. Using the target recognition technology based on the DBSCAN algorithm, the dataset is divided into clusters based on density, and points with similar densities are aggregated into the same target cluster. This algorithm does not require a preset number of generated targets, adapts to arbitrary shape distributions, has good stability, and has good noise resistance. It can effectively filter stray signals under the condition of no prior information, automatically realize target recognition and segmentation, and obtain the number of targets and spatiotemporal positioning results;

[0087] (2) Define the target association value, use a standardized formula to determine the target similarity, draw the spatiotemporal distribution map of each target based on the clustering results, and calculate the association value based on the Hu moments of the target images for possible related target combinations.

[0088] (3) Based on the target association value, Kuhn_Munkras is used to automatically associate the target to obtain the optimal association result. This reduces the impact of subjective factors in human interpretation on the association effect and can automatically achieve target association.

[0089] This invention has good correlation efficiency in multi-target scenarios, improving the ability to identify and distinguish multiple targets.

[0090] The parts of this invention not described in detail are prior art.

[0091] The embodiments selected herein for the purpose of disclosing the inventive objectives are currently considered suitable; however, it should be understood that the invention is intended to include all variations and modifications of the embodiments that fall within the scope of this concept and invention.

Claims

1. A multi-objective association method based on clustered images, characterized by: The multi-target association method specifically includes the following steps: The first step is to obtain the time points and coordinates in the x and y directions of the pulse signals generated by the front and rear targets. The second step involves using the target recognition technology based on the DBSCAN algorithm on the data obtained in the previous step. Each cluster generated is used as a target for recognition, and the time and coordinate range of each cluster are calculated as parameters of the target. The third step is to draw a spatiotemporal distribution map of the same size for each target individually. The fourth step is to perform image preprocessing on the spatiotemporal distribution maps of multiple targets in front and behind the target. Image preprocessing includes generating grayscale images from the original images, binarizing the grayscale images, calculating the Hu moments of the target images, and obtaining possible combinations of related targets according to the time sequence. Step 5: When there is more than one current and subsequent target, use the Kuhn-Munkres algorithm for multi-target association.

2. The multi-objective association method based on clustered images according to claim 1, characterized in that: In the first step, the data is converted into a two-dimensional vector consisting of the sum of time, x, and y coordinates, which reduces computational complexity and improves computational speed without affecting the accuracy of target range identification.

3. The multi-objective association method based on clustered images according to claim 1, characterized in that: In the first step, each data point in the pulse signals of the front and rear targets is in vector form, containing the time and position information of the target passing through the front and rear targets.

4. The multi-objective association method based on clustered images according to claim 1, characterized in that: The second step, target recognition based on the DBSCAN algorithm, specifically includes the following steps: Step 1: Set the parameters of the DBSCAN algorithm and ,in The neighborhood radius, For a certain sample point The algorithm is based on sample density clustering and finds the minimum number of samples in the neighborhood of a sample point. This is because the number of samples in the neighborhood of a sample point exceeds a certain threshold. Add it to the cluster that is close to it. Setting different parameters will affect the number of clusters generated, that is, the number of targets identified. Based on the distribution of the experimental data sample points, set and ;when Increasing the number of clusters may decrease the number of clusters generated, and vice versa. If the size is too large, the number of clusters increases, and vice versa. If the value is too small, it will result in too few clusters; The second step is to determine the formula for calculating the distance between two points in the DBSCAN algorithm, and to choose the Euclidean distance formula. The third step is to cluster the front and rear target data according to the set parameters. Each cluster generated is a valid target. The output of the front and rear target clustering results is used to calculate the time and coordinate range of each cluster as the positioning parameters of each target.

5. The multi-objective association method based on clustered images according to claim 1, characterized in that: The spatiotemporal distribution map mentioned in the third step is a graphic drawn in a two-dimensional planar manner using the target time and coordinate range obtained by clustering. This graphic can reflect the time range and coordinate range of the target signal.

6. The multi-objective association method based on clustered images according to claim 1, characterized in that: When using the Kuhn-Munkres algorithm for multi-target association in the fifth step, the specific steps include: The first step is to use the front and rear target groups as the left and right vertices of the Kuhn-Munkres algorithm, respectively, with each vertex representing a target; When the number of targets in the front and rear targets is not equal, virtual vertices need to be added to make the number of targets in the front and rear targets equal. The second step is to filter out target combinations that may have a relationship based on the time order in which the preceding and following targets appear, and set them as edges in the Kuhn-Munkres algorithm. The third step is to calculate the Hu moments of the target image. For target combinations that are related, calculate the target association value based on the Hu moments. The greater the similarity, the smaller the target association value. The calculation result of the target association value is not affected by the displacement and rotation of the image. Step 4: Set the vertex weights and the weights of all edges connected to that vertex in the Kuhn-Munkres algorithm based on the target association value; Step 5: Construct a weighted bipartite graph according to the set vertex weights and edge weights, and use the Kuhn-Munkres algorithm to perform maximum matching. The maximum matching is the matching with the most edges and the largest sum of weights among all matchings. Output the association results.

Images