Search

CN-121074452-B - Multi-target merging method based on effective cut point under cross-camera scene

CN121074452BCN 121074452 BCN121074452 BCN 121074452BCN-121074452-B

Abstract

The invention discloses a multi-target merging method, medium and equipment based on an effective cut point under a cross-camera scene, and relates to the technical field of image analysis. And combining the specific similar targets under different cameras into a cluster by means of hierarchical clustering and combining, performing threshold constraint by utilizing the feature similarity among the clusters, combining into an undirected graph, and adopting a graph analysis method to divide the undirected graph through effective cut points so as to reduce the miscombination phenomenon caused by jump noise and further improve the accuracy of multi-target combination under a cross-camera scene.

Inventors

  • MAN QINGKUI
  • LIU JING

Assignees

  • 杭州云栖智慧视通科技有限公司

Dates

Publication Date
20260512
Application Date
20230925

Claims (8)

  1. 1. The multi-target merging method based on the effective cut point under the cross-camera scene is characterized by comprising the following steps of: s101, acquiring a plurality of targets in images shot by a plurality of cameras within a preset duration; S102, clustering is carried out based on the feature similarity between the targets to obtain a plurality of clusters, wherein each cluster comprises a clustering center; S103, calculating feature similarity aiming at the clustering centers of every arbitrary two clusters to obtain a similarity matrix, wherein each row/column of the similarity matrix corresponds to one clustering center, and the value of an element of an ith row/column in the similarity matrix represents the feature similarity between the ith clustering center and the jth clustering center; S104, constructing one or more undirected graphs based on the values of elements in the similarity matrix, wherein the undirected graphs comprise a plurality of nodes and edges, the nodes of the undirected graphs are in one-to-one correspondence with the clustering centers of the clusters, and when the values of the elements are larger than a first preset threshold value, edges exist between two nodes corresponding to the elements; the flow of constructing the undirected graph is as follows: From the cluster center I=0, if the current merge flag is false, the cluster center is indicated After the traversal analysis, the next cluster center traversal is performed, if not found When the clustering center is true, the clustering center indicates that all clusters in the current clustering result are processed; If it is For true, the cluster center is represented No relation analysis with other clustering centers is performed, and the clustering centers are used for Adding node set Vertexes as node Vertex i , traversing and clustering the center All adjacent cluster result members, and the cluster center meeting the condition is used as a new node to be added to the node set Vertexes; S105, traversing each undirected graph, calculating the degrees of all cutting points in the undirected graph, and screening out the cutting points with the degrees larger than n, wherein n is larger than or equal to 2; S106, screening out a cutting point with the characteristic similarity corresponding to the cutting edge smaller than a second preset threshold value as an effective cutting point according to the characteristic similarity corresponding to the cutting edge corresponding to the screened out cutting point; S107, dividing the undirected graph based on the effective cut points to obtain a plurality of sub-graphs, and taking the nodes corresponding to each sub-graph as a combination result; the step S107 includes: traversing the undirected graph, and not traversing the effective cut points in the traversing process until each node in the undirected graph is traversed to obtain a plurality of subgraphs; obtaining a first association degree of the effective cutting point and each sub-graph based on the degree of the effective cutting point in each sub-graph; When the maximum value of the first association degree is only one, deleting the effective cut point in the subgraph of which the first association degree is not the maximum value; When the maximum value of the first association degree is p, deleting the effective cut point in any p-1 sub-graphs with the maximum first association degree, and deleting the effective cut point in sub-graphs with the first association degree which is not the maximum value, wherein the effective cut point is determined by the method ; And taking the node corresponding to each sub-graph as a merging result.
  2. 2. The multi-objective merging method in a cross-camera scene based on effective cut points according to claim 1, wherein after S104, further comprising: And for each undirected graph, judging whether the number of nodes of the undirected graph is greater than a third preset threshold, if so, executing S105-S107, and if not, taking all the nodes corresponding to the undirected graph as merging results.
  3. 3. The multi-target merging method under the cross-camera scene based on the effective cutting point according to claim 1, wherein when the maximum value of the first association degree is p, calculating the sum of edge values of the effective cutting point in the corresponding p subgraphs respectively to obtain a second association degree of the effective cutting point and the p subgraphs, wherein the edge values are feature similarity corresponding to edges; if the maximum value of the second association degree is only one, deleting the effective cut point in the subgraph of which the second association degree is not the maximum value; if the maximum value of the second association degree is q, deleting the effective cut point in any q-1 sub-graphs with the maximum second association degree, and deleting the effective cut point in sub-graphs with the second association degree which is not the maximum value, wherein the effective cut point is deleted by the sub-graphs with the maximum second association degree 。
  4. 4. The multi-objective merging method in a cross-camera scene based on effective cut points according to claim 1, wherein the calculation formula of the feature similarity is as follows: ; Wherein, the The feature similarity between the object X and the object Y is represented, and the features of the object X, Y are K-dimensional vectors, which are respectively recorded as: , 。
  5. 5. the multi-target merging method in a cross-camera scene based on effective cut points according to claim 1, wherein the feature of the cluster center is the average number of features of all targets of the cluster in which the cluster center is located or the feature of the target closest to the cluster center.
  6. 6. The multi-target merging method in a cross-camera scene based on effective cut points according to claim 1, wherein in S102, a hierarchical clustering algorithm is adopted as the clustering algorithm.
  7. 7. The multi-objective merging method in a cross-camera scene based on effective cut points according to claim 1, wherein in S105, the traversal undirected graph adopts depth-first traversal.
  8. 8. The multi-objective merging method in a cross-camera scene based on effective cut points according to claim 1, wherein in S106, for the found cut points, traversing the cut edges around the cut points, if the cut edges exist, analyzing the feature similarity corresponding to the cut edges corresponding to the cut points, if one of the cut edges is smaller than a second preset threshold, determining that the cut points are effective, and adding the cut point set.

Description

Multi-target merging method based on effective cut point under cross-camera scene Cross Reference to Related Applications The application relates to a split application of Chinese patent application, which is based on the application number 2023112418830 and the application date 2023, 09 and 25, and has the name of 'a multi-target merging method, medium and equipment under a cross-camera scene'. Technical Field The present invention relates to the field of image analysis technologies, and in particular, to a method, medium, and apparatus for multi-target merging in a cross-camera scene. Background In the field of video image analysis, multi-object merging across cameras is an important algorithmic field. Particularly in the field of multi-target tracking of video analysis, the targets which are same at the same time and under different cameras are required to be classified and combined, so that different angles and postures of the same target are obtained, and subsequent logic analysis processing is performed. Under different cameras, factors such as angles, body postures, illumination brightness and the like of targets can cause great differences of targets, and how to effectively analyze logical relations on the basis of the positions and the characteristics of the targets is the key of an algorithm. At present, when targets are combined across cameras, feature distances and space-time constraints among clusters are constrained by utilizing different threshold conditions through multiple hierarchical clustering, so that the same targets are combined and different targets are distinguished, but the threshold conditions have the phenomenon of 'one cut', and a plurality of different targets are combined into one category by some jump noise images, so that the phenomenon of frequent error combination occurs. Therefore, how to solve the problem of error merging caused by jump noise is a key problem to be solved by the technical scheme. Disclosure of Invention In order to solve at least one technical problem in the background art, the invention aims to provide a multi-target merging method, medium and equipment under a cross-camera scene, which can reduce the error merging phenomenon caused by jump noise and improve the accuracy of cluster merging. In order to achieve the above purpose, the present invention provides the following technical solutions: in a first aspect, an embodiment of the present invention provides a multi-target merging method in a trans-camera scene, including: s101, acquiring a plurality of targets in images shot by a plurality of cameras within a preset duration; S102, clustering is carried out based on the feature similarity between the targets to obtain a plurality of clusters, wherein each cluster comprises a clustering center; S103, calculating feature similarity aiming at the clustering centers of every arbitrary two clusters to obtain a similarity matrix, wherein each row/column of the similarity matrix corresponds to one clustering center, and the value of an element of the ith row/column in the similarity matrix represents the feature similarity between the ith clustering center and the jth clustering center; S104, constructing one or more undirected graphs based on the values of elements in the similarity matrix, wherein the undirected graphs comprise a plurality of nodes and edges, the nodes of the undirected graphs are in one-to-one correspondence with the clustering centers of the clusters, and when the values of the elements are larger than a first preset threshold value, edges exist between two nodes corresponding to the elements; S105, traversing each undirected graph, calculating the degrees of all cutting points in the undirected graph, and screening out the cutting points with the degrees larger than n, wherein n is larger than or equal to 2; S106, screening out a cutting point with the characteristic similarity corresponding to the cutting edge smaller than a second preset threshold value as an effective cutting point according to the characteristic similarity corresponding to the cutting edge corresponding to the screened out cutting point; And S107, dividing the undirected graph based on the effective cut points to obtain a plurality of sub-graphs, and taking the nodes corresponding to each sub-graph as a merging result. Further, after S104, the method further includes: And for each undirected graph, judging whether the number of nodes of the undirected graph is greater than a third preset threshold, if so, executing S105-S107, and if not, taking all the nodes corresponding to the undirected graph as merging results. Further, the step S107 includes: traversing the undirected graph, and not traversing the effective cut points in the traversing process until each node in the undirected graph is traversed to obtain a plurality of subgraphs; obtaining a first association degree of the effective cutting point and each sub-graph based on the degree of the effective cutting