CN104065932B - A kind of non-overlapping visual field target matching method based on amendment weighting bigraph (bipartite graph) - Google Patents

Abstract

The invention proposes a non-overlapping view target matching method based on a modified weighted bipartite graph, which relates to the field of computer vision. The present invention expresses the matching problem of non-overlapping sight domain targets as a maximum a posteriori probability problem, thereby combining the target observation model with the space-time constraint information of the monitoring network, and solves the maximum a posteriori probability problem by solving the maximum weight matching of a weighted bipartite graph . Aiming at the problem that it is easy to introduce error matching in the construction of ordinary weighted bipartite graph, the present invention proposes a modified weighted bipartite graph construction method based on adaptive threshold, so as to avoid introducing errors in the process of constructing weighted bipartite graph as much as possible match. Aiming at the drawbacks of the traditional KM method, which is too computationally intensive when dealing with large-scale weighted bipartite graph matching, a method based on MH sampling is proposed to approximately solve the maximum weight matching of weighted bipartite graphs, so as to obtain non-overlapping view-shed target matching. relation.

Description

A Non-overlapping Viewshed Target Matching Method Based on Modified Weighted Bipartite Graph

technical field

The invention belongs to the field of computer vision, specifically relates to the field of intelligent monitoring, in particular to a non-overlapping view target matching method based on a modified weighted bipartite graph.

Background technique

With the development of camera monitoring technology, monitoring a large area has become an important means to ensure the safety of people's lives and property. However, for a monitoring situation with a large area, it is unrealistic to use cameras to cover all the monitoring areas. Therefore, the method of covering key areas is usually used to build a multi-camera surveillance system with non-overlapping fields of view. For non-overlapping viewshed monitoring networks, only by determining the matching relationship of targets in different viewsheds can we better understand the behavior of targets in the entire surveillance scene. However, interference factors such as illumination changes and scene changes in a non-overlapping view environment will cause changes in the observed values of the target under different view areas, thereby reducing the probability of correct matching of the target. Therefore, it is necessary to comprehensively consider the target information and scene information, so as to improve the target matching rate in the non-overlapping view as much as possible.

The problem of target matching in non-overlapping horizons is essentially a data fusion problem. The purpose is to fuse the similarity information of the observation model with the spatio-temporal constraint information of the monitoring network, so as to achieve target matching in non-overlapping horizons. Researchers at home and abroad have studied this data fusion problem and achieved certain results. For example, the document "Huang T, RussellA. Object identification: A Bayesian analysis with application to trafficsurveillance [J]. Artificial Intelligence. 1998, 103(1): 1-17" establishes a Bayesian framework to match adjacent Vehicles detected by two cameras. The document "Kettnaker V, Zabih R.Bayesianmulti-camera surveillance[C]. IEEE Computer Society Conference on ComputerVision and Pattern Recognition. Fort Collins: IEEE, 1999.2010-2016" also uses the Bayesian framework to achieve indoor human matching and recognition , approximate the maximum a posteriori probability estimation problem as a linear programming problem. The document "Javed O, Shafique K. Tracking across multiple cameras with disjoint views[C]. Proceedings of the IEEE International Conference on Computer Vision. Nice: IEEE, 2003.952-957" not only gives a more detailed and in-depth appearance model and topological relationship The estimation method also uses the method of bipartite graph matching to calculate the maximum posterior probability, and the calculation process meets the requirements of polynomial time.

Domestic research on relevant data fusion issues is mainly based on the document "Javed O, Shafique K. Tracking across multiple cameras with disjoint views[C]. Proceedings of the IEEE International Conference on Computer Vision. Nice: IEEE, 2003.952-957" of. For example, the literature "Liu Shaohua, Zhang Maojun, Chen Wang. Data association method for multi-cameras with non-overlapping views [J]. Computer Applications, 2009, 29(9): 2378-2382" proposed a weighted bipartite graph-based association The strategy uses the target as a node, combines time constraints and space constraints to construct edges, the similarity of the target is the weight of the edge, and obtains the optimal association result by finding the maximum weight matching of the bipartite graph. The literature "Liu Shaohua, Lai Shiming, Zhang Maojun. Non-overlapping multi-camera target association method based on the minimum cost flow model [J]. Association method, which corrects the shortcoming of over-reliance on utility function in bipartite graph association. However, the time complexity of the cost flow association problem is larger than that of the bipartite graph association problem, so it is not suitable for dealing with large-scale matching problems.

Therefore, using an unsupervised learning method to adaptively learn the topology of a multi-camera surveillance network is crucial for the practical application of a multi-camera surveillance system. In the non-overlapping view-shed target matching method, there are three main points to be noted: (1) How to integrate the target’s own information and external environment information to build a non-overlapping view-shed target matching method framework; (2) How to use mature method to solve the non-overlapping horizon object matching problem; (3) how to avoid the introduction of wrong matching relationship as much as possible and improve the matching speed when using the graph theory method to solve the non-overlapping horizon object matching problem.

Contents of the invention

The purpose of the present invention is to propose a non-overlapping view target matching method based on modified weighted bipartite graph. This method has strong robustness to the impact of illumination changes and environmental differences on target matching under non-overlapping view fields.

The technical solution of the present invention is: the non-overlapping sight target matching is expressed as a maximum a posteriori probability problem, and when constructing the maximum a posteriori probability problem, the target observation model and the space-time constraints of the monitoring network are comprehensively considered. Construct a weighted bipartite graph, and solve the maximum a posteriori probability problem by solving the maximum weight matching of the bipartite graph. Aiming at the problem that it is easy to introduce wrong matching in the construction of ordinary weighted bipartite graph, a modified weighted bipartite graph construction method based on adaptive threshold is proposed, so as to avoid introducing wrong matching in the process of constructing weighted bipartite graph as much as possible. Aiming at the drawbacks of the traditional KM method, which has too much calculation when dealing with large-scale weighted bipartite graph matching, a method based on MH sampling is proposed to approximate the maximum weight matching of weighted bipartite graphs, so as to obtain the non-overlapping visual domain target matching relationship .

Concrete implementation steps of the present invention are as follows:

1) Determine the target observation model

2) Determine the time and space constraints of the monitoring network

3) Divide the smallest unit of time and space

4) Construct the maximum a posteriori probability problem

5) Calculate the adaptive threshold

6) Construct a modified weighted bipartite graph

7) The MH method approximates the maximum weight matching of the weighted bipartite graph

Description of drawings

Fig. 1 is a system flow chart of the non-overlapping view target matching method based on the modified weighted bipartite graph of the present invention

Fig. 2 (a) is the extraction schematic diagram of the main color feature that the present invention uses

Fig. 2 (b) is the extraction schematic diagram of the space texture feature that the present invention uses

Fig. 3 is a schematic diagram of monitoring network topology in the present invention

Fig. 4 is a schematic diagram of the division of the smallest space unit in the present invention

Fig. 5 is a schematic diagram of time minimum unit division in the present invention

Fig. 6 is a schematic diagram of the modified weighted bipartite graph in the present invention

Detailed ways

Figure 1 shows the system flow chart of the non-overlapping view-shed target matching method based on the modified weighted bipartite graph: using the maximum a posteriori probability framework to describe the non-overlapping view-shed target matching problem, when constructing the maximum a posteriori probability framework, comprehensive consideration The target observation model and the space-time constraints of the monitoring network are established. The invention extracts the main color feature and the space texture feature of the target, and constructs a target observation model with strong robustness to illumination differences and environmental changes. On the premise that the topology of the monitoring network is known, the invention obtains the space-time constraints of the monitoring network from the topology. According to the space-time constraints of the monitoring network, the monitoring network topology is divided into multiple minimum units according to the principle of independence and minimization. For any target that disappears in the minimum unit, its reappearance is still located in the minimum unit. For the target matching problem in any smallest unit, it is expressed as a maximum a posteriori probability problem by using the Bayesian inference criterion. The invention uses graph theory to solve the maximum posterior probability problem, transforms the maximum posterior probability problem into the maximum weight matching problem of the weighted bipartite graph, and then obtains the matching relationship of non-overlapping visual domain objects. Aiming at the problem that it is easy to introduce error matching in the construction of ordinary weighted bipartite graph, the present invention proposes a modified weighted bipartite graph construction method based on adaptive threshold, so as to avoid introducing errors in the process of constructing weighted bipartite graph as much as possible match. Aiming at the drawbacks of the traditional KM method, which has too much calculation when dealing with large-scale weighted bipartite graph matching, a method based on MH sampling is proposed to approximate the maximum weight matching of weighted bipartite graphs, so as to obtain non-overlapping view-shed target matching relation.

Concrete operation steps of the present invention

1) Construct target observation model

Use the mature feature extraction technology to extract the target features, and then construct the target observation model. Considering that environmental factors such as illumination changes and background differences in non-overlapping fields of view will affect the target imaging, thereby reducing the discrimination ability of the target observation model, the present invention comprehensively considers the main color features and spatial texture features of the target, and constructs an image of the environment. Variation factors are more robust to target observation models. The main color features describe the macroscopic information of the target. The present invention uses the K-means clustering method to extract the main color features of the target, and uses the standardized RGB distance to calculate the similarity between the main colors, thereby improving the robustness of the main color features to illumination changes. sex. The spatial texture feature describes the detailed information of the target. The invention divides the target into blocks, and counts the number of pixels satisfying the gradient condition in different directions of each sub-block, thereby obtaining the spatial texture feature represented by a histogram. The invention uses a weighting method to fuse main color features and spatial texture features to construct a target observation model. The extraction schematic diagram of the main color feature used in the present invention is as shown in Figure 2 (a), wherein the left side of any sub-graph is the detected target, and the right side is the extracted 15 kinds of main colors; the space texture used in the present invention The schematic diagram of feature extraction is shown in Figure 2(b), in which the detected target is on the left side of any subgraph, the gradient image and block situation are in the middle, and the histogram representing spatial texture features is on the right side.

2) Determine the time and space constraints of the monitoring network

On the premise that the topology of the monitoring network is known, the invention obtains the time-space constraint information of the network from the topology of the monitoring network. The space-time constraints of the monitoring network include the reachability of the nodes in the network and the average transit time. The reachability of the nodes can be determined by monitoring the connected edges in the network topology. If there is a connected edge between two nodes in the topology, then the monitoring area corresponding to the two nodes is reachable. For two connected nodes in the topology, the transfer time distribution is usually determined by a single Gaussian distribution, then the average transit time between the network nodes corresponding to these two points is the mean value of the single Gaussian distribution. The schematic diagram of the monitoring network topology used in the present invention is shown in FIG. 3 , where the black arrows represent the connectivity between topology nodes.

3) Divide the smallest unit of time and space

For the non-overlapping viewshed target matching problem, the same target may appear at multiple locations and at multiple times, resulting in greater uncertainty in the target matching problem. The present invention proposes the concept of the smallest unit of time and space. In any smallest unit of time and space, when the target corresponding to the disappearance observation value reappears, it is only possible within the space range determined by the smallest unit, and the time of appearance must be determined by the smallest unit. within a certain time frame. The division of the smallest space-time unit decomposes the entire monitoring network and the target matching problem in all time periods into independent pairwise matching problems of observations, thereby eliminating the uncertainty in target matching to the greatest extent. The division of the smallest space-time unit must follow two principles, namely, the principle of completeness and the principle of minimization.

Assume that for an independent unit U (U={node _i |i=1,2,…,k}) containing k nodes, the division of the smallest spatial unit of U must satisfy the following two principles:

①Completeness principle: for a target that disappears from any node _i in unit U, the node _j where it reappears satisfies node _j ∈ U; similarly, for a target that appears at any node _j in unit U, its last The node where it disappeared satisfies node _i ∈ U.

②Minimization principle: On the basis of satisfying completeness, let each unit contain the least number of nodes.

The principle of completeness and minimization in the division of the smallest unit of time is similar to the principle of completeness and minimization in the division of the smallest unit of space.

As shown in Figure 4, it is a schematic diagram of the division of the smallest spatial unit in the present invention. The division process is as follows: first remove the dotted line connection in the figure, and the original connected graph becomes a non-connected graph. At this time, each connected subgraph in the figure is Represents a minimum unit of space. Since each unit is a connected graph, for any node in the unit, the connected nodes must still be in the unit, thus satisfying the completeness principle of unit segmentation. If any node in the unit is separated, then for other nodes connected to the node, the connected nodes are not in the same unit, which violates the completeness principle. Therefore, no node in the unit can be separated, which satisfies the minimum principle of unit division. For the topology shown in Figure 4, two independent units can be obtained, which are {3,8,10,11,14} and {4,7,9} respectively.

FIG. 5 is a schematic diagram of the division of the smallest time unit in the present invention. The division process is as follows: First, determine the disappearance observation value and the appearance observation value in the figure. Assuming that the observation value O ₄ ¹ represents the first observation value detected by node 4 at time t1, it is divided into O ₄ ^1- and O ₄ ¹⁺ , which are located on both sides of the graph. After adding all observations in the same way, the connectivity of different observations is listed using the time accessibility between topological nodes, and each sub-connected space is regarded as a time minimum unit. Taking FIG. 5 as an example, the two minimum time units are {t1-t1:t1-t4} and {t2-t5:t6-t8} respectively.

The connectivity between observations is estimated by the time accessibility of topological nodes. Assuming that the transition time distribution of nodes 4 and 9 obeys N(μ,σ ² ), then for observations O ₄ ^1- and O ₉ ¹⁺ , connectivity can be defined as follows:

Among them, t4-t9 represents the time when the target passes through the node, and the threshold 3σ is obtained according to the 3σ criterion in the Gaussian distribution.

4) Construct the maximum a posteriori probability problem

Non-overlapping visual domain target matching is to fuse the target observation model with the space-time constraints of the monitoring network through a specific data fusion method, so as to obtain accurate target matching information. Bayesian inference is a commonly used data fusion method. The invention uses the Bayesian inference criterion to express the non-overlapping view target matching problem as a maximum posterior probability problem. Assuming that in a certain spatio-temporal smallest unit, Mi disappearing observations are detected, recorded as O _i ＝{O _i ¹ ,O _i ² ,…,O _i ^Mi }, where O _i ^m represents the i-th observation value in the smallest unit The information of the mth target captured by a camera includes the target observation model O _i ^m (app) and the space-time constraint relationship O _i ^m (st). Since the observation model of the target has nothing to do with the time and position of the target in the monitoring network, it can be considered that O _im (app) and _{O im} ^{( st} ) are independent of each other. Similarly, when N _j observed values are detected, it is denoted as O _j ={O _j ¹ ,O _j ² ,...,O _j ^Nj }. make Indicates that the observation pairs (O _i ^m , O _j ⁿ ) are related to each other, then the non-overlapping viewshed target matching problem can be expressed as finding a matching set The following conditions:

① ∈A if and only if O _i ^m and O _j ⁿ belong to the same target;

②An observation value has at most one predecessor and one successor, that is, for all Have

suppose is a solution to the target matching problem, and all matches are independent of each other, then the target matching problem can be expressed as:

in, Indicates that after the observed values O _i ^m and O _j ⁿ are detected, match probability of occurrence. Through Bayesian inference, we can get:

The target observation model O _i ^m (app) and the space-time constraint information O _i ^m (st) can be introduced into the target matching problem, defined as follows:

in, Indicates the observation model matching rate of the target, Represents the spatiotemporal constraint information of the topology, Represents the transition probability of the target from camera C _i to C _j , p(O _i ^m , O _j ⁿ ) is represented by a constant scale factor (assuming that the target appearance belongs to a uniform distribution).

The non-overlapping field of view target matching problem is to find the solution A* that maximizes the posterior probability:

5) Calculate the adaptive threshold

In the traditional weighted bipartite graph construction process, only connected edges are added to the bipartite graph according to the spatiotemporal constraints of the topology, and the weight of the edges is the similarity of the pairs of observations to be matched. However, some erroneous observation pairs may also meet the spatio-temporal constraints and thus be added to the weighted bipartite graph and eventually lead to association errors. The present invention proposes an adaptive threshold to control the addition of connected edges in a weighted bipartite graph, and the threshold must satisfy the following two criteria:

① If the similarity of the observations to be matched is generally large, it means that the change of the environment has little impact on the target matching. At this time, the threshold should be increased to exclude the wrong pairs of observations; Generally small, the threshold should be reduced at this time, so as to prevent the correct matching relationship from being excluded.

②If the number of disappearing observations in a minimum unit is inconsistent with the number of appearing observations, it means that false detection or blind area replacement may have occurred. At this time, the threshold should be increased to exclude non-ideal pairs of observations.

Based on the above two principles, the calculation process of the adaptive threshold proposed by the present invention is as follows: for any disappearing observed value, calculate its similarity with the appearing observed value, and add the maximum similarity to the set U. After similar processing for each vanishing observed value, arrange U in descending order to obtain a set containing n elements. Assuming that the correct observation value pair is more than half of the total number of observation value pairs (the probability of false detection or blind area replacement is relatively small), take out the first half of elements in U {u ₀ ,u ₁ ,…,u _m }, dynamic threshold Φ It is defined as follows:

in, Indicates the mean value of the selected elements, which is used to represent the influence of environmental factors on the dynamic threshold setting. n represents the number of vanishing target observations in the smallest unit. P _a represents an adaptive parameter, which is used to adjust the threshold. If the number of observations of the disappearing target in the smallest unit is inconsistent with the number of observations of the appearance of the target, it is necessary to increase the value of P _a to increase the threshold.

6) Construct a modified weighted bipartite graph

Under ideal conditions (one-to-one correspondence between the disappearance observation value and the appearance observation value in the smallest space-time unit), the description process of transforming the maximum a posteriori probability problem into a weighted bipartite graph is as follows: Assume that in a certain smallest unit, the disappearance observation value Two point sets of the weighted bipartite graph are constituted with the observed values, expressed as X={x ₁ ,x ₂ ,...,x _m } and V={v ₁ ,v ₂ ,...,v _m }. Use the space-time constraints of the monitoring network to judge the connectivity of any observation pair x _i -v _j , thus forming the edges of the weighted bipartite graph. For a connected edge x _i -v _j , calculate the matching rate of the observation model between x _i and v _j as the weight of the edge. In this way, the original maximum a posteriori probability problem is transformed into a weighted bipartite graph problem. For the construction of the modified weighted bipartite graph proposed by the present invention, the observation value pairs satisfying the space-time constraints are not simply added to the weighted bipartite graph, but the adaptive threshold Φ is calculated according to the method in 1), and then the observation model Observation pairs with a matching rate ≥ Φ are added to the weighted bipartite graph to construct a modified weighted bipartite graph. The schematic diagram of the corrected weighted bipartite graph constructed by the present invention is shown in Figure 6. This graph contains 4 pairs of disappearance and appearance observation values. The black connecting line represents the edge of the weighted bipartite graph, and the edge weight w _ij represents the relationship between x _i and v _j Observe the model matching rate.

7) The MH method approximates the maximum weight matching of the weighted bipartite graph

Using the MH method to calculate the maximum weight matching of a bipartite graph, the weighted bipartite graph problem must first be transformed into the form required by the MH method, that is, the target state x, the "proposed function" Q(x, ) and the probability distribution function P (x). Assume that for the weighted bipartite graph G={U, V, E}, the point set U is selected for matching. Let x represent the current state space, x∈{0,1}|E|, xij represents the matching relationship of the edge (ui,vj), xij=1 means that ui and vj match, that is, there is a connected edge between ui and vj; Conversely, xij=0 means that ui does not match vj. Let Q(x'|xt) represent the "proposed function", that is, how to infer the possible state x' at the next moment from the current state xt. Since what is to be solved here is the bipartite graph weighted matching problem, the present invention defines a "proposed function" as randomly taking two pairs of observations and exchanging their connected relations, that is: take out ui and vj, um and vn, and the current state is xij=1 and xmn=1, xin=1 and xmj=1 after being processed by the "proposed function", and the connectivity relations of the other points remain unchanged. When calculating the acceptance probability, the present invention comprehensively considers the "proposed function" Q(x'|xt) and the probability distribution function P(x), that is, f(x')=Q(x'|xt)P(x') . f(x) is defined as follows:

① When the randomly selected two pairs of points contain shared points (that is, two points on one side are connected to the same point on the other side), f(x)=0;

②When the randomly selected two pairs of points do not contain shared points, That is, the weighted value of the current state.

In the actual calculation, for the calculation of the weighted value of the current state in step ②, it is only necessary to consider the influence of the two pairs of observations that transform the connectivity relationship on the overall weighted value. If the selected pair of points is impossible to connect, then the weight of the pair of points is -1 to punish the wrong connection. Method 2 presents a bipartite graph maximum weight matching process based on MH sampling approximation.

Method 1 (Bipartite Graph Maximum Weight Matching Based on MH Sampling Approximation)

1: Initializing: x ₀ , t=0, f(x _t )=0;

2:loop

3:Sample x'from Q(x'|x _t );

4: Sample U from U(0,1);

5: Calculate the acceptance ratio

6:if U≤α(x _t |x')then

7: x _t+1 = x';

8: else

9: x _t+1 = x _t ;

10: end if

11:t=t+1;

12: end loop.

Claims (2)

1. A non-overlapping view-shed target matching method based on a modified weighted bipartite graph, characterized in that the maximum a posteriori probability problem is used to describe the non-overlapping view-shed target matching, and the target observation is comprehensively considered when constructing the maximum a posteriori probability problem Model and monitoring network space-time constraints; construct a weighted bipartite graph, solve the maximum a posteriori probability problem by solving the maximum weight matching of the bipartite graph; specifically include the following steps: 1) Determine the target observation model; The structure of the target observation model can be combined with one or more of color features, texture features, and shape features. For any of the above-mentioned features, a large number of sub-features are included, and the sub-features include but are not limited to : Color features include main color features, RGB histogram features, HSV histogram features, texture features include edge features, HOG features, LBP features, shape features include area features, and discreteness features; 2) Determine the time and space constraints of the monitoring network; In the determination of the space-time constraints of the monitoring network, the space-time constraints of the monitoring network include the reachability of nodes in the network and the average transit time. These two kinds of information can be given artificially or obtained from the topology of the monitoring network; 3) Divide the smallest space-time unit; According to the space-time constraints of the monitoring network, the monitoring network topology is divided into multiple minimum units according to the principle of completeness and minimization. For any disappearing target in the minimum unit, its reappearance is still located in the minimum unit; 4) Construct the maximum a posteriori probability problem; Using the Bayesian inference criterion, the non-overlapping view target matching problem is expressed as a maximum a posteriori probability problem. Assume that i, j, k, and l represent the serial numbers of cameras in the surveillance network, and m, n, c, and d represent different cameras capturing The serial number of the detected target, in a certain space-time smallest unit, detects M _i disappearing observations, recorded as O _i ＝{O _i ¹ ,O _i ² ,…,O _i ^Mi }, where O _i ^m represents The information of the m-th target captured by the i-th camera in the smallest unit includes the target observation model O _i ^m (app) and the space-time constraint relationship O _i ^m (st). The time that appears in has nothing to do with the position, so it can be considered that O _im (app) and O _im (st) are independent of each other; similarly, N ^j observations that appear are _detected , recorded as O ^j _＝ {O _j ¹ , O _j ² ,…,O _j ^Nj }, let Indicates that the observation pairs (O _i ^m , O _j ⁿ ) are related to each other, then the non-overlapping viewshed target matching problem can be expressed as finding a matching set The following conditions: ① If and only if O _i ^m and O _j ⁿ belong to the same target; ②An observation value has at most one predecessor and one successor, that is, for all Have suppose is a solution to the target matching problem, and all matches are independent of each other, then the target matching problem can be expressed as: in, Indicates that after the observed values O _i ^m and O _j ⁿ are detected, match The probability of occurrence; through Bayesian inference, we can get: The target observation model O _i ^m (app) and the space-time constraint information O _i ^m (st) can be introduced into the target matching problem, defined as follows: in, Indicates the observation model matching rate of the target, Represents the spatiotemporal constraint information of the topology, Denotes the transition probability of the target from camera C _i to C _j , Use a constant scale factor, assuming that the target appears to belong to a uniform distribution; The non-overlapping horizon target matching problem is to find the solution A ^* that maximizes the posterior probability: 5) Calculate the adaptive threshold; The threshold must meet the following two criteria: ① If the similarity of the observations to be matched is generally large, it means that the change of the environment has little impact on the target matching. At this time, the threshold should be increased to exclude the wrong pairs of observations; Smaller, the threshold should be reduced at this time, so as to prevent the correct matching relationship from being excluded; ②If the number of disappearing observations in a minimum unit is inconsistent with the number of appearing observations, it means that false detection or blind area replacement may have occurred. At this time, the threshold should be increased to exclude non-ideal observation pairs; Based on the threshold must meet two criteria ①②, the calculation process of the adaptive threshold is as follows: for any disappearance observation value, calculate its similarity with the appearance observation value, and add the maximum similarity to the set U, After similar processing for each disappearing observation value, arrange U in descending order to obtain a set containing q elements; assuming that the correct observation value pair is more than half of the total number of observation value pairs, the probability of false detection or blind replacement is relatively small , take out the first half of elements {u ₀ ,u ₁ ,…,u _r } in U, and the dynamic threshold Φ is defined as follows: where r=(q/2)-1, Indicates the mean value of the selected element, which is used to indicate the influence of environmental factors on the dynamic threshold setting, q indicates the number of observed values of the disappearing target in the smallest unit, P _n indicates the adaptive parameter, which is used to adjust the threshold, if the observation of the disappeared target in the smallest unit The number of values is inconsistent with the number of observations where the target appears, and the value of P _n needs to be increased to increase the threshold; 6) Construct a modified weighted bipartite graph; In the case of one-to-one correspondence between disappearing observations and appearing observations in the smallest space-time unit under ideal conditions, the description process of transforming the maximum a posteriori probability problem into a weighted bipartite graph is as follows: Assume that in a certain smallest unit, the disappearing observation and the observed values constitute two point sets of weighted bipartite graph respectively, expressed as X={x ₁ ,x ₂ ,…,x _λ } and V={v ₁ ,v ₂ ,…,v _λ }; The spatio-temporal constraints of the network judge the connectivity of any observation pair x _b -v _h to form the edges of the weighted bipartite graph. For a connected edge x _b -v _h , calculate the matching rate of the observation model between x _b and v _h , as the weight of the edge, so that the original maximum a posteriori probability problem is transformed into a weighted bipartite graph problem; According to the method in step 5), the adaptive threshold Φ is calculated, and then the observation value pair of the observation model matching rate is added to the weighted bipartite graph, thereby constructing the modified weighted bipartite graph; 7) The main steps of the maximum weight matching approximate solution of the weighted bipartite graph of MH sampling: a) MH (Metropolis-Hastings) sampling method The "proposed function" Q(x, y) in the MH sampling method must satisfy the following three conditions: i, when x is fixed, Q(x, ) is a probability density function; ii, for any x, y, Q(x, y) can be calculated; iii, when x is fixed, random numbers can be easily generated from Q(x,y); Under the premise of satisfying the above conditions, the "proposed function" Q(x, y) can be selected arbitrarily in theory; the MH sampling method uses the "proposed function" Q(x, ) to generate the next moment from the current distribution x _t The possible distribution x', and define the acceptance probability to judge whether the received state x' is a new state; the MH sampling method updates the state through the "proposal function", and calculates the acceptance probability to judge whether to accept the new state, if not accept the new state, Then the current state is still the state of the previous moment, and the acceptance probability reveals which of the current state and the new state is closer to the real state of the target. By continuously updating the state and judging whether to accept the new state, it can gradually approach the real state of the target; b) Maximum weight matching solution of approximate bipartite graph based on MH sampling Using the MH sampling method to calculate the maximum weight matching of the bipartite graph, the weighted bipartite graph problem must first be transformed into the form required by the MH sampling method, that is, the target state x, the "proposed function" Q(x, ) and the probability distribution are defined Function P(x), assuming that for the weighted bipartite graph G={U,V,E}, select the point set U for matching, let x represent the current state space, x∈{0,1} ^|E| , x _bh Indicates the matching relationship of the edge (ub, v _h ), x _bh = 1 means that u _b matches v _h , that is, there is _{a connected edge between u b and v h; otherwise, x bh = 0 means that u b and v h are not} match; let Q(x'|x _t ) represent the "proposal function", that is, how to deduce the possible state x' at the next moment from the current state x _t ; since the bipartite graph weighted matching problem is to be solved here, Therefore, the "proposed function" is defined as randomly taking two pairs of observations and exchanging their connected relations, that is, taking out u _b and v _h , u _λ and v _z , the current state is x _bh =1 and x _λz =1, after " Proposal function" after processing x _bz = 1 and x _λh = 1, the connection relationship of the remaining points remains unchanged; when calculating the acceptance probability, the "proposed function"Q(x'|x _t ) and the probability distribution function P( x), that is, f(x')=Q(x'|x _t )P(x'), the definition of f(x) is as follows: (b-1) When two pairs of randomly selected points contain shared points, that is, two points on one side are connected to the same point on the other side, f(x)=0; (b-2) When the randomly selected two pairs of points do not contain shared points, That is, the weighted value of the current state; For the calculation of the weighted value of the current state in step (b-2), it is only necessary to consider the influence of the two pairs of observations that transform the connectivity relationship on the overall weighted value. If the selected pair of points is impossible to connect, Then let the weight of the pair be -1, so as to punish the wrong connection. 2. The non-overlapping view target matching method based on the modified weighted bipartite graph according to claim 1, wherein the sub-features in the step 1) use a combination of main color features and HOG features to establish a target observation model.