CN103810266B - Semantic network target recognition sentences card method - Google Patents

Abstract

The present invention proposes a semantic network object recognition judgment method, which aims to provide a comprehensive judgment method that can reduce the calculation amount of evidence synthesis and update, reduce the conflict among multi-attribute evidences, and improve the consistency of recognition results. The present invention is realized through the following technical solutions: between evidence receiving and evidence output, a target semantic database and an evidence receiving module for interacting data with the target semantic database, an evidence semantic knowledge extraction module and an evidence semantic knowledge expansion module are created, and the evidence receiving module real-time Receiving target recognition evidence from different types of sensor recognition sources, the evidence semantic knowledge clustering module clusters the elements in the multi-attribute set involved in the extended attribute constraint relationship, and obtains several attribute classifications, and the evidence synthesis update module from multiple In the constraint relationship between attributes, the identification evidence of different levels of attributes and the classification evidence of each attribute are orthogonally calculated and updated by orthogonal synthesis to obtain the updated identification evidence result.

Description

A Discriminative Method for Object Recognition in Semantic Web

technical field

The invention relates to a target recognition method based on multi-sensor data fusion in the field of target recognition and tracking pattern recognition, in particular to a comprehensive identification method for multiple identification sources and multiple attributes under the multi-identification framework of an integrated target identification system.

Background technique

With the development of target recognition technology, it is a development trend to use multiple types of sensors and multiple recognition techniques to comprehensively identify targets. At present, the target recognition and tracking system based on multi-sensor (radar and infrared) signal fusion, because different types of recognition sources independently recognize the target, the recognition results given are often different at the attribute level, and the recognition results at the same recognition level are often inconsistent. In order to make the integrated recognition system have a unified output, it is necessary to conduct judgment and decision-making on the target recognition results given by various recognition methods. The decision-making process of judgment includes single-attribute or single-identification-frame discrimination, and multi-attribute or multi-identification-frame discrimination. However, at present, it is difficult to reduce the conflict among multiple attribute evidences and improve the consistency of identification results by conducting comprehensive judgments based on multiple identification sources and multiple attributes.

In the existing technology, multi-sensor data fusion, as an emerging interdisciplinary subject, has received extensive attention and rapid development in recent years. Multi-sensor data fusion technology can integrate various aspects of information provided by multiple sensors, and can obtain more comprehensive and accurate information of the observed object, so as to obtain accurate and fast decision-making and judgment. Object recognition is an important part of data fusion technology. The general definition of target recognition is to make some kind of discrimination on the target type or attribute. Target recognition is also called attribute classification or identity estimation. The output data from multiple target sources can be both dynamic information and identity information. Dynamic information is the dynamic parameters of target motion, usually including position, velocity and acceleration. Identity information is obtained from multiple target sources to help The relevant information of the proposition or statement to establish the target identity. Since the target identity information is composed of sensor signals, attribute information, and identity description. Since single-sensor systems usually only provide partial information for identifying and tracking objects, multi-sensor systems use data fusion technology to synthesize information from different sources to overcome the defects of single sensors. Measure space to obtain information. Since the target is constantly moving and its posture is constantly changing, the images of postures are very different, which greatly increases the difficulty of target recognition in three-dimensional space. Traditional multi-sensor data fusion is carried out on data level, feature level and decision level. In multi-sensor target recognition, the traditional method is to directly send multiple local decisions to the fusion center for the final overall decision. Existing methods for discriminating between multiple identification sources and multiple attributes can be roughly divided into three categories: (1) Bayesian network reasoning; (2) evidence theory reasoning; (3) heuristic judgment. Bayesian network is a probabilistic network based on the mathematical model of probabilistic inference. It is a graphical network based on probabilistic inference, and Bayesian formula is the basis of this probabilistic network. The so-called probabilistic reasoning is the process of obtaining other probability information through the information of some variables. The Bayesian network based on probabilistic reasoning is proposed to solve the problems of uncertainty and incompleteness. It has a good advantage in solving the failures caused by the uncertainty and correlation of complex equipment, and has been widely used in many fields. . The Bayesian network, also known as the belief network, is itself an uncertain causal association model. Bayesian network is different from other decision-making models. It is a probabilistic knowledge representation and reasoning model that visualizes multivariate knowledge diagrams. It is one of the most effective theoretical models in the field of uncertain knowledge representation and reasoning. It is a directed acyclic graph (Directed Acyclic Graph, DAG), which consists of nodes representing variables and directed edges connecting these nodes. Nodes represent random variables, and the directed edges between nodes represent the mutual relationship between nodes (from the parent node to its descendant nodes). The conditional probability is used to express the relationship strength, and the prior probability is used to express information if there is no parent node. A node variable can be an abstraction of any problem, such as: test value, observation phenomenon, opinion consultation, etc. Applicable to the expression and analysis of uncertain and probabilistic events, applied to decision-making conditionally dependent on multiple control factors, can never be complete. Making inferences from imprecise or uncertain knowledge or information. The construction of a Bayesian network is a complex task that requires the participation of knowledge engineers and domain experts. It is necessary to establish a conditional probability table between various attributes, and to represent the constraint relationship between different attributes in a quantitative form. In reality, , this quantitative relationship is not easy to obtain; at the same time, Bayesian network reasoning cannot solve the "unknown" situation. Evidence theory reasoning is relatively simple for the single identification framework, but for the multi-identification framework of the multi-information fusion identification framework, there is no clear way to realize it. Although the DS method of evidence theory reasoning has been widely used in various data fusion systems, its algorithm implementation has become a difficult problem due to the computational complexity of the Dempster composition rules, the core of the DS method. Heuristic judgment mainly uses the constraint knowledge between attributes to update the reliability of attributes according to certain rules such as scoring, so as to make judgment decisions. Although this method is simple, it does not have a strict mathematical basis. Moreover, the existing judgment methods do not make good use of the semantic knowledge between target attributes, and do not have a clear mathematical realization of multi-attribute judgment, so the reliability is poor, and the rationality of the results cannot be guaranteed.

Contents of the invention

The task of the present invention is to provide a more reasonable judgment for the deficiencies of multi-target identification sources, different sources of comprehensive information, and multi-attribute comprehensive identification systems, which can reduce the amount of evidence synthesis and update calculations, and can identify evidences from multiple categories. The identification source contains a comprehensive judgment method of multi-attribute evidence such as target attributes, types, and models of "unknown" statements, so as to reduce the conflict between multi-attribute evidence and improve the consistency of identification results.

The solution adopted by the present invention to solve the problems of the prior art is: a semantic network target recognition and discrimination method, which has the following technical features: creating and storing various target entities between evidence receiving and evidence output, and semantic subordination knowledge between multiple attributes The target semantic library with attribute constraint relationship at different levels of the target, the evidence receiving module, the evidence semantic knowledge extraction module, the evidence semantic knowledge extension module, and the sequential The evidence output module that outputs the results of the judgment data through the evidence semantic knowledge clustering module and the evidence synthesis update module; the evidence receiving module receives target recognition evidence from different types of sensor recognition sources in real time, including target attributes, target types, target models, and target attributes The identification statement of one or more attributes in the hierarchy and/or the "unknown" identification evidence statement contained in it; the evidence semantic knowledge extraction module extracts the affiliation relationship between various target attributes stored in the target semantic library, and extracts from the evidence receiving module The semantic knowledge involved in the received recognition evidence statement; the evidence semantic knowledge extension module expands the extracted semantic knowledge according to the transfer rules of the affiliation relationship between target attributes, and obtains the extended attribute support constraint relationship and attribute conflict constraint relationship; evidence The semantic knowledge clustering module clusters the elements in the multi-attribute set S involved in the extended attribute constraint relationship, and obtains several attribute categories, and the attributes in each category are the minimum attribute sets with mutual constraint relationships. The evidence synthesis update module performs orthogonal calculation and orthogonal synthesis update on the identification evidence of different levels of attributes and the classification evidence of each attribute from the constraint relationship between multiple attributes, and obtains the updated identification evidence judgment result; the evidence output module will update The final identification evidence judgment result is output to other modules that call the method of the present invention.

Compared with the prior art, the present invention has the following beneficial effects.

The present invention introduces between evidence receiving and evidence output including: target semantic library, evidence receiving module, evidence semantic knowledge extraction module, evidence semantic knowledge expansion module, evidence semantic knowledge clustering module, evidence synthesis update module and evidence output module semantic network , through the constraint relationship between the attributes of different levels of the target described by the target semantic library, as the basis of the judgment process, making the judgment more reasonable. The target semantic library is stored independently during the judgment process, which is convenient for maintenance; the three modules of evidence semantic knowledge extraction module, evidence semantic knowledge expansion module, and evidence semantic knowledge clustering module classify the attributes involved in the received evidence, reducing the The amount of calculation for evidence synthesis and update; the evidence synthesis update module starts from the global constraint relationship between multiple attributes, and performs strict orthogonal calculation and orthogonal synthesis on the identification evidence of different levels of attributes, so as to obtain globally better judgment results ; Participating in the judgment of "unknown" attributes in multi-category attributes makes the description of the judgment results more reasonable.

The present invention uses the target semantic database and the evidence receiving module, the evidence semantic knowledge extraction module, and the evidence semantic knowledge expansion module to classify and process the attributes involved in the received evidence, and to classify the attributes from multiple types of identification sources including The target attribute, type, model, etc. of the "unknown" statement are comprehensively judged by multi-identification sources and multi-attribute evidence, which further reduces the conflict between multi-attribute evidence and improves the consistency of identification results.

Description of drawings

In order to understand the present invention more clearly, the present invention will now be described through the embodiments of the present invention with reference to the accompanying drawings, wherein:

Fig. 1 is a schematic diagram of the principle of semantic network object recognition discrimination in the present invention.

Fig. 2 is a flow chart of the evidence semantic knowledge clustering module involved in the present invention.

Fig. 3 is a flow chart of the evidence synthesis update module involved in the present invention.

detailed description

See Figure 1. In the semantic network target recognition and judgment embodiment described below, between evidence receiving and evidence output includes: target semantic database, evidence receiving module, evidence semantic knowledge extraction module, evidence semantic knowledge extension module, evidence semantic knowledge clustering module , an evidence synthesis update module, and an evidence output module. The semantic affiliation knowledge of multiple attributes of various target entities is stored in the target semantic library; the evidence receiving module receives in real time the target recognition evidence of multiple attributes that can contain "unknown" statements from different types of sensor recognition sources; the evidence semantic knowledge The extraction module extracts the semantic knowledge involved in the received recognition evidence statement from the affiliation relationship between various target attributes stored in the target semantic library; the evidence semantic knowledge extension module extracts the extracted Extend the semantic knowledge to obtain the extended attribute support constraint relationship and attribute conflict constraint relationship; the evidence semantic knowledge clustering module clusters the elements in the attribute type set involved in the expanded attribute constraint relationship to obtain several attribute classifications , the attributes in each category are the minimum set of attributes with mutual constraints; the evidence synthesis update module performs orthogonal calculation and orthogonal calculation on the identification evidence of different levels of attributes and the evidence of each attribute classification from the constraint relationship between multiple attributes. Synthesize and update to obtain the updated identification evidence judgment result; the evidence output module outputs the updated identification evidence judgment result. The multi-attribute set is S={model, size type, platform type, environment type, attribute, nationality, ...}; the specific implementation steps are as follows: in step S0, the target semantic database stores different objects related to model, size type, and platform type , attribute and other semantic descriptions between objects contained in target attributes of each level. Taking the object "F16" contained in the target model as an example, its semantic affiliation knowledge is described as: small air target; taking the object contained in the platform type as an example, its semantic affiliation knowledge is described as: fixed-wing aircraft, small air target. For the convenience of description, the constraint relationship between different target attribute objects is abbreviated as R[(T ₁ ,t ₁ ),r,(T ₂ ,t ₂ )]. Among them, R represents a constraint knowledge; (T ₁ , t ₁ ), (T ₂ , t ₂ ) represent object t ₁ of attribute T ₁ and object t ₂ of attribute T ₂ respectively; r represents (T ₁ , t ₁ ) and (T ₂ , t ₂ ), there are three kinds of relationships: "support", "conflict", and "not declare", which are respectively denoted as _rs , r _c , and r _u . Wherein, T ₁ and T ₂ are the target attributes respectively; t ₁ and t ₂ are the attribute objects corresponding to the target attribute T ₁ and the attribute objects corresponding to the target attribute T ₂ respectively.

In step S1, the evidence receiving module receives in real time target identification evidence of multiple attributes that may contain "unknown" statements from different types of sensor identification sources such as radar, optoelectronics, electronic reconnaissance, and identifiers, and stores them in the cache.

In step S2, the evidence semantic knowledge extraction module extracts the multi-attribute recognition evidence received in the current shot from the cache, and then extracts the semantics involved in the received recognition evidence from the affiliation relationship between various target attributes stored in the target semantic library. Knowledge.

In step S3, the evidence semantic knowledge expansion module expands the extracted semantic knowledge according to the transfer rules of the affiliation relationship between target attributes, and obtains the extended attribute support constraint relationship and attribute conflict constraint relationship; the involved target attribute affiliation relationship The delivery rules are as follows:

[Support extended rules]: If R[(T ₁ , t ₁ ), r _s , (T ₂ ,t ₂ )] and R[(T ₂ ,t ₂ ), r _s ,(T ₃ ,t ₃ )] , then R[(T ₁ , t ₁ ), r _s , (T ₃ , t ₃ )];

[Conflict extension rule]: If R[(T ₁ , t ₁ ), _rs , (T ₂ , t ₂ )] and R[(T ₂ , t ₂ ), r _c , (T ₃ ,t ₃ )] , then R[(T ₁ , _{t 1} ), rc , (T ₃ , t ₃ )].

In step S4, the evidence semantic knowledge clustering module clusters the elements in the set of attribute types involved in the expanded attribute constraint relationship.

See Figure 2. In step S41, the clustering module initializes the clustering results, and regards each target attribute as a cluster; the clustering module traverses all the current clustering results in step S42, selects a cluster pair, and uses one of them as a reference class, Taking the other as the class to be compared, the clustering module judges whether there is a constraint relationship between the objects involved in the reference class and the objects involved in the class to be compared in the cluster pair selected in step S43. If there is a constraint relationship, go to step S44 and the clustering module merges the attributes in the class to be compared into the reference class; otherwise, go to step S45 and judge whether the clustering module has traversed all the cluster pairs formed in the existing clusters. If the traversal is complete, then enter step S46 and the clustering module will store the obtained clustering results in the cache; otherwise, proceed to step S42;

See Figure 3. In step S5, the evidence synthesis update module combines each cluster in the clustering results in the cache with the evidence involved in the attribute set corresponding to the cluster, and performs orthogonal operations and orthogonal synthesis on the input identification evidence. In the process, the confidence of each attribute classification is synthetically updated to obtain the updated identification evidence. In step S51, the evidence synthesis update module selects an unprocessed cluster among the attribute clusters in the cache, and implements the following steps for the attributes involved in the cluster: Orthogonal calculation of attributes, the calculation method includes: the selected cluster C _q contains m attributes, the i-th attribute contains n _i definite variables and 1 "unknown" variable, a total of (n _i +1) attributes object. The confidence assigned by the input evidence about the jth object corresponding to the ith attribute is denoted as Compute the elements in the orthogonal product matrix

is the validity flag of the orthogonal product, which takes a value of 0 or 1, and the determination method is: if there is p∈[1,m] and q∈[1,m], there is a relationship R[( ],but otherwise Among them, m is the number of attributes; j ₁ , j ₂ , and j _m represent the jth object corresponding to the _first attribute, the j2th object corresponding to the _second attribute, and the jth object corresponding to the mth attribute, respectively. _m objects; is the confidence degree of the j-th object corresponding to the i-th attribute; is an orthogonal element in the orthogonal matrix obtained by step S52; is the validity flag of the orthogonal product, which takes a value of 0 or 1. The determination method is: if there is p∈[1,m] and q∈[1,m], there is a relationship R[( ],but otherwise Among them, T _p is the pth attribute, is the j _pth attribute object in the pth attribute; r _c means "conflict" relationship.

In step S53, the evidence synthesis update module normalizes each element in the orthogonal matrix to obtain a normalized matrix, and calculates the normalized value in the normalized matrix

In the formula, is the orthogonal element in the orthogonal matrix obtained in step S52.

In step S54, the evidence synthesis update module marginalizes the orthogonal elements in the normalized orthogonal matrix to obtain the confidence value of each attribute object after updating

In step S55, the evidence synthesis update module uses the following formula to estimate the credibility value of each object contained in each attribute according to the updated evidence of each attribute; for the j variable object of the i attribute Indicates an "unknown" variable. The reliability is estimated to be

In step S56, for each type of attribute in the cluster, determine the maximum credibility value of all objects; the maximum credibility value of all objects for the i-th attribute is And determines The storage number j ^* of the corresponding object in all objects of the attribute; given a threshold γ _th , if then will As the final judgment of the i-th attribute; otherwise, t ₀ , that is, "unknown" is taken as the final judgment of the i-th attribute. in, is the normalized value of the element in the normalized matrix obtained in step S53, is the updated confidence of the j-th object corresponding to the i-th attribute, the jth object representing the kth attribute, Indicates the j-th object of the i-th attribute, and n _i is the number of objects other than "unknown" of the i-th attribute, Indicates the credibility of the j-th object of the i-th attribute, Indicates the maximum credibility value of all objects in the i-th attribute, and γ _th indicates a decision threshold.

In step S57, it is judged whether the evidence synthesis update module has processed all the attribute clusters, if not, return to step S51; otherwise, go to step S58, update the confidence of all clustered attribute objects at each level The result of attribute judgment of the hierarchy is saved to the designated buffer area.

In step S6, the evidence output module outputs the attribute judgment result saved in the cache area in step S58 to the network, so that other devices can receive it.

What has been described above are only preferred embodiments of the present invention. It should be pointed out that those skilled in the art can make some modifications and improvements without departing from the principles of the present invention, and these changes and changes should be deemed to belong to the protection scope of the present invention.

Claims (9)

1. A semantic network target recognition and discrimination method, which has the following technical features: between evidence receiving and evidence output, a target semantic library is created to store various target entities and semantic membership knowledge between multiple attributes, as well as attribute constraints at different levels of the target , and classify the attributes involved in the received evidence; introduce between evidence receiving and evidence output including: target semantic library of interactive data, evidence receiving module, evidence semantic knowledge extraction module, evidence semantic knowledge extension module, evidence semantics Knowledge clustering module, evidence synthesis update module, and evidence output module that outputs forensic data results, wherein the evidence receiving module receives target recognition evidence from different types of sensor recognition sources in real time, including target attributes, target types, target models, and target attributes The identification statement of one or more attributes in the hierarchy and/or the "unknown" identification evidence statement contained in it; the evidence semantic knowledge extraction module extracts the affiliation relationship between various target attributes stored in the target semantic library, and extracts from the evidence receiving module The semantic knowledge involved in the received recognition evidence statement; the evidence semantic knowledge extension module expands the extracted semantic knowledge according to the transfer rules of the affiliation relationship between target attributes, and obtains the extended attribute support constraint relationship and attribute conflict constraint relationship; evidence The semantic knowledge clustering module clusters the elements in the multi-attribute set S involved in the extended attribute constraint relationship, and obtains several attribute categories, and the attributes in each category are the minimum attribute sets with mutual constraint relationships. The evidence synthesis update module synthesizes and updates the confidence of each attribute classification from each cluster in the clustering results in the cache, combined with the evidence involved in the attribute set corresponding to the cluster, and obtains the updated identification evidence. And from the constraint relationship between multiple attributes, the identification evidence of different levels of attributes and the classification evidence of each attribute are orthogonally calculated, and the orthogonal synthesis is updated to obtain the updated identification evidence judgment result; the evidence output module judges the evidence synthesis update module Whether to complete all attribute clustering processing, and output the identification evidence judgment result updated after attribute clustering processing to other calling modules. 2. The semantic network object recognition and discrimination method according to claim 1, characterized in that: the multi-attribute set is S={model, size type, platform type, environment type, attribute, nationality, ...}. 3. The semantic network target identification and discrimination method as claimed in claim 1, characterized in that: each object contained in different targets about model, size type, platform type, attribute and each level target attribute is stored in the target semantic base Semantic description between. 4. The semantic network target identification method according to claim 1, characterized in that: the constraint relationship between different target attribute objects is R[(T ₁ ,t ₁ ),r,(T ₂ ,t ₂ ) ], where R represents a constraint knowledge; (T ₁ , t ₁ ), (T ₂ , t ₂ ) represent object t ₁ of attribute T ₁ and object t ₂ of attribute T ₂ respectively; r represents (T ₁ , The relationship between t ₁ ) and (T ₂ , t ₂ ) has three relationships: "support", "conflict", and "not declare", where T ₁ and T ₂ are target attributes respectively; t ₁ and t ₂ are respectively the attribute object corresponding to the target attribute T ₁ and the attribute object corresponding to the target attribute T ₂ . 5. The semantic network target identification method as claimed in claim 1, characterized in that: the evidence semantic knowledge clustering module clusters the elements in the set of attribute types involved in the extended attribute constraint relationship, clustering The result is initialized, and each target attribute is regarded as a cluster, traverse all the current clustering results, select a cluster pair, use one of them as a reference class, and use the other as a class to be compared, and judge the selected cluster pair , whether there is a constraint relationship between the objects involved in the reference class and the objects involved in the class to be compared, if there is a constraint relationship, merge the attributes in the class to be compared into the reference class; otherwise, determine whether all existing clusters have been traversed The clustering pairs formed in the class; if the traversal is complete, the obtained clustering results will be saved in the cache. 6. The semantic network target identification and discrimination method as claimed in claim 1, characterized in that: the evidence synthesis update module selects an unprocessed cluster in the attribute class clusters in the cache, and for the clusters involved in the cluster The attributes in the selected cluster are orthogonally calculated. The calculation method includes: the selected cluster C _q contains m attributes, the i-th attribute contains n _i definite variables and 1 "unknown" variable, There are a total of (n _i + 1) attribute objects, and the confidence assigned by the input evidence about the jth object corresponding to the ith attribute is recorded as Compute the elements in the orthogonal product matrix 7. the semantic network object identification discrimination method as claimed in claim 6, is characterized in that: the element in the orthogonal product matrix Among them, m is the number of attributes; j ₁ , j ₂ , and j _m represent the jth object corresponding to the _first attribute, the j2th object corresponding to the _second attribute, and the jth object corresponding to the mth attribute, respectively. _m objects; is the confidence degree of the j-th object corresponding to the i-th attribute; is an orthogonal element in the orthogonal matrix obtained by step S52; is the validity flag of the orthogonal product, which takes a value of 0 or 1, and the determination method is: if there is a relationship between p∈[1,m] and q∈[1,m] but otherwise Among them, T _p is the pth attribute, is the j _pth attribute object in the pth attribute; r _c means "conflict" relationship. 8. The semantic network target identification and discrimination method as claimed in claim 7, characterized in that: the evidence synthesis update module normalizes each element in the orthogonal matrix to obtain a normalized matrix, and calculates the normalized matrix in the normalized matrix. normalized value In the formula, are the orthogonal elements in the orthogonal matrix obtained in step S52, and j ₁ , j ₂ , and j _m represent the j _1st object corresponding to the first attribute, the j ₂ th object corresponding to the second attribute, and the j th object corresponding to the second attribute, respectively. The j _mth object corresponding to the m attributes. 9. The semantic network target identification method as claimed in claim 8, characterized in that: the evidence synthesis update module marginalizes the orthogonal elements in the normalized orthogonal matrix to obtain the confidence of each attribute object after the update value The evidence synthesis update module uses the following formula to estimate the credibility value of each object contained in each attribute according to the updated evidence of each attribute; for the jth variable object of the ith attribute j=0 means "unknown" variable, The reliability is estimated to be For each type of attribute in the cluster, determine the maximum credibility of all objects; the maximum credibility of all objects in the i-th attribute is And determines The storage sequence number j ^* of the corresponding object in all objects of the attribute, given a threshold γ _th , if then will As the final judgment of the i-th attribute; otherwise, t ₀ , that is, "unknown" is taken as the final judgment of the i-th attribute; where, is the normalized value of the element in the normalized matrix obtained in step S53, is the updated confidence of the j-th object corresponding to the i-th attribute, the jth object representing the kth attribute, Indicates the j-th object of the i-th attribute, and n _i is the number of objects other than "unknown" of the i-th attribute, Indicates the credibility of the j-th object of the i-th attribute, Indicates the maximum credibility value of all objects in the i-th attribute, and γ _th indicates a decision threshold.