JP7278763B2 - State quantity data classification device and state quantity data classification method - Google Patents

Description

The present invention mainly relates to a state quantity data classifying device.

Conventionally, techniques have been proposed for estimating the state quantity of a target and classifying information for each target using acquired information.

The target state quantity estimation device disclosed in Patent Document 1 includes an observation update calculation device, a time extrapolation calculation device, and a pattern matching calculation device. The Observation Update Calculator updates the estimator using the observables obtained from the Seeker. The time extrapolation calculation device performs time extrapolation calculation using the estimated state quantity obtained from the observation update calculation device and its error covariance matrix. The pattern matching calculation device uses the target acceleration estimated value obtained from the time extrapolation calculation device, compares it with preset target turning pattern data, and estimates the motion pattern of the target.

According to Patent Literature 1, this allows the estimation filter calculation to quickly converge with sufficient accuracy by the pattern matching calculation process.

Patent Literature 2 discloses a target state quantity estimating device that estimates a target state quantity using a Kalman filter. This target state quantity estimation device includes an observation noise matrix calculation device, an observation update calculation device, a time extrapolation calculation device, and a process noise matrix calculation device. The observation noise matrix calculator calculates an observation noise matrix representing the magnitude of error in the observed quantity based on the observed quantity, which is data relating to the distance and direction to the target, and the S/N ratio of the observed quantity. The observation update calculation device uses the observed quantity and the observation noise matrix to update the target state quantity and the error covariance matrix output by the time extrapolation calculation device. The time extrapolation calculation device calculates the target state quantity and error by time extrapolation based on the updated target state quantity and error covariance matrix, and the process noise matrix that is the magnitude of the error of the target motion model. Calculate the covariance matrix. The process noise matrix calculator calculates a process noise matrix based on the target state quantity calculated by the time extrapolation calculator.

According to Patent Document 2, even in the situation of estimating the motion of a target with a small radar reflection cross-section and a low S/N ratio of the observed value, the filter characteristics satisfy the performance index required on the spot. (that is, an appropriate process noise matrix, which is the characteristic of the Kalman filter, etc.) can be obtained, and the target state quantity can be estimated with high accuracy.

Patent Literature 3 discloses a target correlation integration device that includes a target observation information input device, a motion parameter correlator, and a first correlation group generator. The target observation information input device receives target information from the target observation device that has captured the target as observation information. The motion specification correlator determines whether the observation information is correlated with the managed target, and the observation information falls within a target existence prediction range centered on the predicted position of the already managed target. or not. A first correlation group generator generates a first correlation group using observation information of observation targets expected to be correlated according to the determination of the motion specification correlator and prediction information of the managed targets as one correlation group. I do.

According to Japanese Patent Laid-Open No. 2002-200013, this enables targets having similar properties (position, velocity, etc.) to some extent to be reintegrated as one target.

JP 2013-195104 A JP 2016-161269 A JP-A-9-171072

However, the configurations of Patent Documents 1 and 2 do not classify information for each target, so when estimating the state quantity of a target, information related to a different target may be mixed. As a result, it may become impossible to accurately estimate the state quantity of the target.

In the configuration of Patent Document 3, when the obtained observation information and the information about the managed target are similar, it becomes difficult to perform accurate correlation determination, and it may not be possible to generate an accurate correlation group. There was room for improvement.

The present invention has been made in view of the above circumstances, and its object is to classify acquired state quantity data into state quantity data for each target with high accuracy. is to provide

The problems to be solved by the present invention are as described above. Next, the means for solving the problems and the effects thereof will be described.

A first aspect of the present invention provides a state quantity data classification device having the following configuration. That is, this state quantity data classification device includes a state quantity data acquisition section and a state quantity data classification section. The state quantity data acquisition unit acquires state quantity data output by detecting the state quantity of the target by a plurality of sensors. The state quantity data classification unit uses the state quantity data and category affiliation information indicating to which category the state quantity data relates to the target when the target is classified into a plurality of categories from some point of view, to classify the Classify state quantity data. The state quantity data classifying unit classifies state quantity data indicating state quantities that are similar to each other into the same cluster, classifies state quantity data that have the same section belonging information into the same cluster, classifies the state quantity data into the same cluster, classifies the section belonging information state quantity data with different are classified into different clusters. When the state quantity data classifying unit classifies the state quantity data, a plurality of the state quantity data having the same degree of similarity between the state quantity data classified into each cluster and the classification belonging information are classified into different clusters. A first penalty regarding classification evaluation is taken into consideration when classifying a plurality of state quantity data having different classification affiliation information into the same cluster.

According to a second aspect of the present invention, the following state quantity data classification method is provided. That is, this state quantity data classification method includes a state quantity data acquisition step and a state quantity data classification step. In the state quantity data obtaining step, the state quantity data output by detecting the state quantity of the target by a plurality of sensors is obtained. In the state quantity data classification step, the state quantity data and category affiliation information indicating which category the state quantity data belongs to when the targets are classified into a plurality of categories from some point of view are used to classify the Classify state quantity data. In the state quantity data classification step, state quantity data indicating state quantities that are similar to each other are classified into the same cluster, state quantity data with matching section belonging information are classified into the same cluster, and the section belonging information is classified into the same cluster. state quantity data with different are classified into different clusters. When classifying the state quantity data in the state quantity data classifying step, the plurality of state quantity data having the same degree of similarity between the state quantity data classified into each cluster and the classification belonging information are classified into different clusters. A first penalty regarding classification evaluation is taken into consideration when classifying a plurality of state quantity data having different classification affiliation information into the same cluster.

As a result, in addition to the similarity of the state quantity data, by considering whether or not the target category affiliation information related to each state quantity data matches, the state quantity data can be obtained from the viewpoint of the physical identity of the targets. can be suitably classified. It is possible to easily classify the state quantity data by comprehensively considering the degree of similarity between the state quantity data and the penalties.

According to the present invention, state quantity data acquired from a plurality of sensors can be accurately classified from the viewpoint of physical identity of targets.

1 is a block diagram showing the overall configuration of a state quantity data classification device according to an embodiment of the present invention; FIG. The figure which shows the flow of the state quantity data in a state quantity data classification device. The conceptual diagram explaining the preprocessing of state quantity data. FIG. 4 is a conceptual diagram for explaining Expression (4), which is an evaluation function for clustering according to the first embodiment; FIG. 4 is a diagram showing an example of a display screen in a display device;

Next, embodiments of the present invention will be described with reference to the drawings. First, the first embodiment will be explained. FIG. 1 is a block diagram showing the overall configuration of a state quantity data classification device 1 according to one embodiment of the present invention.

A state quantity data classification device 1 shown in FIG. 1 is provided in an aircraft (not shown). The aircraft is equipped with a large number of sensors 2 for detecting targets, which are mobile bodies existing in the surroundings. Each of the multiple sensors 2 detects one or multiple targets. In the following description, the target detected by the sensor 2 may be called a detected target. The sensor 2 acquires the state quantity of each detection target.

Various targets are conceivable, and examples include ships and aircraft. In this embodiment, the state quantity means the position and velocity of the target. Although the configuration of the sensor 2 is not particularly limited, it is conceivable to use, for example, a radar device, an optical sensor, an acoustic sensor, or the like. The plurality of sensors 2 may all have the same configuration, or may include sensors with different configurations.

Each sensor 2 has a tracking function and can track one or more detection targets. Since each sensor 2 operates independently, it is conceivable that the detection targets tracked by the plurality of sensors 2 may be physically the same or different.

Each sensor 2 is electrically connected to the state quantity data classification device 1 . The result of detecting the state quantity of the detection target is input to the state quantity data classification device 1 from each sensor 2 as state quantity data. As shown in FIG. 2, the state quantity data classification device 1 organizes the state quantity data of each detection target input from each sensor 2 by classifying and integrating them from the viewpoint of the physical identity of the target. , outputs the state quantities of one or more physical targets. The output result of the state quantity data classification device 1 is displayed in real time on an appropriate display device. As a result, it is possible to support the situation recognition, threat judgment, action judgment, etc. of the operator on board the aircraft, and contribute to the improvement of accuracy and promptness.

The state quantity data detected by the sensor 2 for the target to be detected is input to the state quantity data classification device 1 as shown in FIG. .

The class identification information generation unit 3 generates class identification information (class belonging information). The class identification information is information indicating some attribute of the target, such as the type of the target's body, the nationality of the target, and the like. The class identification information indicates to which class the target belongs when the target is classified into a plurality of categories from some point of view (aircraft type point of view, nationality point of view, etc.).

To give an example, assume that a certain sensor 2 is configured as an image recognition device using an optical camera, and the image recognition device detects two targets (for example, ships) sailing on the sea by image recognition. The sensor 2 acquires the position obtained by image recognition processing for each detection target. Also, the sensor 2 acquires the speed of the detection target by comparing the position obtained in the image a predetermined time ago with the position obtained in the current image. The sensor 2 outputs state quantity data (specifically, position and speed data) of each detection target to the state quantity data classification device 1 .

Furthermore, the sensor 2 outputs the image of each detection target cut out from the camera-captured image by image recognition to the class identification information generation unit 3 .

A database in which images of ships of various types are registered is built in advance in the class identification information generation unit 3 . The class identification information generator 3 determines the class of the ship by comparing the input image with the database. In this example, it is assumed that one of the two detection targets is a cargo ship and the other is a fishing boat. In this case, the class identification information generation unit 3 generates class identification information indicating that the state quantity data of one of the two targets detected by the sensor 2 belongs to a cargo ship in terms of aircraft type. The state quantity data of the other detection target is generated as class identification information indicating that it belongs to a fishing boat in terms of aircraft type.

The point of view for classifying targets is arbitrary, and various points other than those described above are conceivable. The class identification information generation unit 3 generates class identification information by various methods depending on the viewpoint of target classification, the method of detecting the target by the sensor 2, and the like.

Different targets belong to different categories. Therefore, the class identification information serves as a clue for determining whether the detection targets tracked by the plurality of sensors 2 are physically the same or different. Specifically, class identification information indicating that the state quantity data of a detection target of a certain sensor 2 belongs to a cargo ship is created, and class identification information indicating that the state quantity data of a detection target of another sensor 2 belongs to a fishing boat is created. If the identification information is made, it can be said that the two sensed targets are likely to represent physically different targets. In order to make the above judgments well, it is preferable that the class identification information includes divisions based on a plurality of viewpoints.

The class identification information generation unit 3 outputs the generated class identification information to the state quantity data classification device 1 .

The state quantity data classification device 1 includes a state quantity data input unit (state quantity data acquisition unit) 11, a data preprocessing unit 21, a class identification information input unit (class belonging information acquisition unit) 31, and a cluster analysis unit (state Quantity data classification unit) 41 , target state quantity estimation unit (target state quantity acquisition unit) 71 , and display data generation unit 81 .

Specifically, the state quantity data classification device 1 is configured as a known computer, and includes a CPU, a ROM, a RAM, an HDD, and the like. The HDD includes a state quantity data input step (state quantity data acquisition step), a cluster analysis step (state quantity data classification step), a target state quantity estimation step (target state quantity acquisition step) included in the state quantity data classification method of the present embodiment. ), and a program for realizing the display data generating process and the like. Through the cooperation of the hardware and software described above, the state quantity data classification device 1 is divided into a state quantity data input unit 11, a data preprocessing unit 21, a class identification information input unit 31, a cluster analysis unit 41, and a target state quantity estimation unit 71. , and the display data generator 81 .

State quantity data obtained by each sensor 2 detecting a target is input from the sensor 2 to the state quantity data input unit 11 of FIG. The state quantity data input unit 11 outputs the acquired state quantity data to the data preprocessing unit 21 .

The data preprocessing unit 21 performs processing for synchronizing the state quantity data acquired from the plurality of sensors 2 before classifying the state quantity data by the cluster analysis unit 41 .

More specifically, each sensor 2 operates independently to detect a target. Therefore, as indicated by the circles in the conceptual diagram of FIG. It does not match. Therefore, the data preprocessing unit 21, in order to be able to compare the state quantity data among the plurality of sensors 2, from the time transition of the state quantity data acquired from each sensor 2, such as shown in FIG. A process of estimating the state quantity data at the timing (time t1, t2, . . . ) in the time interval is performed. This processing can be performed based on a known data interpolation technique.

The data preprocessing unit 21 shown in FIG. 1 performs preprocessing on the state quantity data input from each sensor 2 to the state quantity data classification device 1 in this way so that the state quantities are obtained at common timing. . The data preprocessing unit 21 outputs the preprocessed state quantity data to the cluster analysis unit 41 .

The class identification information generated by the class identification information generation unit 3 is input to the class identification information input unit 31 . The class identification information input unit 31 outputs the acquired class identification information to the cluster analysis unit 41 .

The cluster analysis unit 41 classifies a plurality of state quantity data output from the data preprocessing unit 21 .

The cluster analysis unit 41 accumulates state quantity data preprocessed by the data preprocessing unit 21 for each sensor 2 and for each detection target tracked by the sensor 2 . As a result, it is possible to generate the time transition of the state quantity data within a predetermined time from the present time.

In this embodiment, the temporal transition of this state quantity data is the object of clustering by the cluster analysis unit 41 . In the following description, state quantity data including time transition may be simply referred to as sample x.

The cluster analysis unit 41 uses each acquired sample x as a classification target set and analyzes which sample x belongs to which target (physical target) based on clustering, thereby classifying each sample x as a physical target. classified by each.

In this embodiment, the state quantity data includes position and velocity. Therefore, the state quantity data are x-coordinate, y-coordinate, and z-coordinate of position, and x-, y-, and z-components of velocity. and can be represented by a combination of six values. As the time transition of the state quantity data, for example, when three time transitions including the present are targeted, each sample x can be expressed as a 6×3=18-dimensional vector. The clustering performed by the cluster analysis unit 41 corresponds to classifying points of each sample x in an 18-dimensional space.

In this embodiment, the cluster analysis unit 41 performs clustering based on a non-hierarchical method so that the value of an evaluation function, which will be described later, is minimized. As a clustering method, a modified fuzzy c-means method is used. The fuzzy c-means method is an extension of the well-known k-means method by introducing the idea of vaguely expressing the degree to which each sample belongs to a cluster (degree of belonging).

The cluster analysis section 41 includes an initial data processing section 51 and a search section 61 .

The initial data processing unit 51 performs clustering initialization processing. The initial data processing unit 51 includes a cluster center initialization unit 52 and a match/mismatch pair creation unit 53 .

The cluster center initialization unit 52 determines the number of clusters (the number of clusters C) by an appropriate method. Then, the cluster center initializing unit 52 sets the initial value of the cluster center μ to the same number as the number of clusters C. FIG. Each cluster center μ is also represented as an 18-dimensional vector like each sample x, and indicates a point in 18-dimensional space. In this embodiment, the initial value of each cluster center μ is randomly set.

The match/mismatch pair creation unit 53 creates a pair of samples x (hereinafter sometimes referred to as a match pair) that is highly likely to represent the same physical target among the samples x input from the different sensors 2. ) to create one or more.

Matching pairs are created in consideration of both the similarity between samples x and the matching (similarity) of class identification information. Specifically, if the position and velocity of the detection target indicated by two samples x are similar to each other over time, it is highly likely that the two samples x indicate the same physical target. It can be said that Also, it can be said that two samples x belonging to the same class in the class identification information are highly likely to physically indicate the same target. The matched/mismatched pair creation unit 53 comprehensively evaluates from these points of view and creates matched pairs.

In addition, the match/mismatch pair creation unit 53 creates a pair of samples x that are highly likely to indicate physically different targets (hereinafter referred to as a mismatch pair) among the samples x input from the different sensors 2 . exist).

Similar to the matched pair, the mismatched pair is also created in consideration of both the similarity of the state quantity data and the matching (similarity) of the class identification information. More specifically, if the time transitions of the position and velocity of the detection target indicated by two samples x are not similar to each other, it is highly likely that the two samples x indicate physically different targets. be able to. Also, it can be said that two samples x belonging to different categories in the class identification information are highly likely to represent physically different targets. The matched/mismatched pair creating unit 53 comprehensively evaluates from these points of view and creates a mismatched pair.

The search unit 61 searches for favorable clustering of the sample x so as to optimize (specifically, minimize) the value of the evaluation function that defines the favorableness of the classification.

As described above, the searching unit 61 of the present embodiment performs clustering using a modified method of the known fuzzy c-means method. Therefore, first, the basic fuzzy c-means method will be briefly described.

The fuzzy c-means method performs clustering by searching for and finding the degree of belonging u _ik that reduces the value of the evaluation function shown in the following equation (1).

where x _i is the i-th sample out of the total number of N samples. u _ik is the degree of belonging of the i-th sample x _i to the k-th cluster G _k , details of which will be described later. μ _k is the above cluster center for the k-th cluster G _k . C is the total number of clusters G; In this embodiment, both the samples x _i and the cluster centers μ _k are represented as 18-dimensional vectors.

d ² ( _xi , μ _k ) means the square of the distance (Euclidean distance) between the i-th sample x _i and the cluster center μ _k of the k-th cluster G _k in an 18-dimensional space. m is a parameter that determines the magnitude of the influence given by the degree of belonging u _ik and is set to a value greater than 1 (eg, 2). Note that the distance between the sample x _i and the cluster center μ _k is not limited to the Euclidean distance, and may be the Mahalanobis distance, the L1 distance, or the like.

The degree of belonging u _ik is a concept of fuzzy inference, and indicates how much each sample x _i belongs to each cluster G _k . Unlike the k-means method, which assumes that each sample completely belongs to only one of a plurality of clusters, in the fuzzy c-means method, a sample does not belong to a certain cluster at all, It is assumed that there is an intermediate state between a state of complete affiliation and a state of affiliation, and this degree of affiliation is represented by a number from 0 to 1. Similar to the fuzzy membership function, membership u is 0 if a sample does not belong to a cluster at all, and 1 if it completely belongs. By using the degree of belonging u, a certain sample belongs to the first cluster by 0.1 and at the same time belongs to the second cluster by 0.2, and so on. can do. For one sample x _i , a value of 1 is obtained by adding the degree of belonging u _ik to each cluster G _k for all clusters.

The right-hand side of the above equation (1), which is the evaluation function, substantially means the residual sum of squares between each sample x _i and each cluster center μ _k weighted by the degree of belonging u _ik raised to the mth power. ing. In the fuzzy c-means method, the degree of belonging u _ik is updated according to the following equation (2) so as to minimize the residual sum of squares for each given cluster center μ _k .

Subsequently, in the fuzzy c-means method, the cluster center μ _k is updated according to the following equation (3) based on the updated degree of belonging u _ik .

In this formula (3), the cluster center μ _k of a certain cluster G _k is weighted by m-th power of the belonging degree u _ik of all the samples x _i that each sample x _i belongs to the cluster G _k . indicates that it is obtained by averaging.

In the fuzzy c-means method, an appropriate convergence condition (for example, each cluster center μ _k hardly changes due to updating) is satisfied for updating the degree of belonging u _ik and updating the cluster center μ _k . Repeat alternately until If the convergence condition is met, clustering is complete.

Next, the meaning of the evaluation function shown in the above equation (1) will be explained.

If a certain sample x _i belongs to a cluster G _k with a cluster center μ _k located nearby in the 18-dimensional space (in other words, if the degree of belonging u _ik is large), the value of the evaluation function will be small. On the other hand, if a certain sample x _i belongs to a cluster G _k with a cluster center μ _k located far in the 18-dimensional space, the value of the evaluation function will be large.

Therefore, in the 18-dimensional space, if the samples x _i belonging to the cluster G _k are concentrated within a range close to each cluster center μ _k , the value of the evaluation function will be small. Also, if the sample x _i is sparse in a space far from any of the plurality of cluster centers μ _k , the value of the evaluation function will be small.

Therefore, it can be said that the smaller the value of the evaluation function, the more similar samples x can be classified into the same cluster G, and the more dissimilar samples x can be classified into different clusters G. By repeatedly updating the degree of belonging u _ik and updating the cluster center μ _k to search for the degree of belonging u _ik that reduces the value of the evaluation function, it is possible to obtain clustering that is preferable from the viewpoint of similarity as described above. can.

Next, based on the above description, the clustering method of this embodiment will be described. In this embodiment, the following equation (4) is used as the evaluation function instead of the above equation (1).

where (x _i , x _j )εM means a matching pair. (x _i , x _j )εC denotes a mismatched pair. α is a parameter that determines the weight of penalties, which will be described later. As α increases, clustering becomes more sensitive to penalties.

The first term on the right side of this evaluation function is exactly the same as the right side of equation (1) of the evaluation function for the c-means method described above. The first term is the residual sum of squares between each sample x _i and each cluster center μ _k weighted by the degree of membership u _ik to the mth power. This term becomes smaller when trying to classify similar samples x into the same cluster G and classify dissimilar samples x into different clusters G, as shown in FIG. 4(a).

In FIG. 4A, each cluster G is surrounded by a dashed line, which indicates that the degree of belonging u _ik to the cluster G of the sample x surrounded by the dashed line is relatively large. Remaining, not necessarily belonging completely. That is, for example, the sample x labeled "A1" is surrounded by the dashed line of cluster _G1 , which means that while the degree of membership of sample x of "A1" to cluster _G1 is close to 1, This means that the degrees of belonging to other clusters G ₂ and G ₃ are close to zero.

The second term on the right side of equation (4) is to classify one sample (first state quantity data) x _i and the other sample (second state quantity data) x _j forming a matched pair into different clusters G. In this case, the first penalty P1 is added.

Considering that matching pairs are a type of clustering constraint, this first penalty P1 tries to classify samples (sample x _i and sample x _j ) that are likely to represent the same target into different clusters G. It can be said that it is a penalty (penalty for violation of constraints) in the case of If consistent clustering can be performed for matching pairs, the value of the first penalty P1 term of the evaluation function can be reduced.

Specifically, in the example of FIG. 4(b), the match/mismatch pair creation unit 53 creates matching pairs "A4" and "B1", "B3" and "C1", and "C2" and " Suppose that three sets of pairs of samples x are determined, such as "A1". In this situation, if "B3" and "C1" are classified into mutually different clusters G and "C2" and "A1" are classified into mutually different clusters G as shown in FIG. The value of the function is increased by the first penalty P1.

As shown in equation (4), the value of the first penalty P1 is the reciprocal of the residual sum of squares between one sample x _i of the matched pair and the cluster center μ _k for a given cluster G _k . , and the reciprocal of the residual sum of squares between the other sample x _j and the cluster center μ _l for a different cluster G _l , and any combination of two different clusters G _k and G _l is defined as the sum of

This means that the magnitude of the first penalty P1 is determined considering the distance between the position of the sample x and the cluster center μ. That is, when classifying two samples _xi and _xj included in a matched pair into mutually different clusters _Gk and _Gl , the value of the first penalty P1 _is set to The smaller the distance between the cluster center μ _k of G _k and the smaller the distance between the other sample x _j and the cluster center μ _l of the cluster G _l into which it is classified, the larger.

The third term on the right side of equation (4) classifies one sample (third state quantity data) x _i and the other sample (fourth state quantity data) x _j forming a mismatched pair into the same cluster G , the second penalty P2 is added.

Considering that mismatched pairs are a type of clustering constraint, this second penalty P2 tries to classify samples (sample x _i and sample x _j ) that are likely to represent different targets into the same cluster G. It can be said that it is a penalty (penalty for violation of constraints) in the case of If non-matching pairs can be clustered without contradiction, the value of the term of the second penalty P2 of the evaluation function can be reduced.

Specifically, in the example of FIG. 4C, the matched/mismatched pair creation unit 53 creates two mismatched pairs, such as "C1" and "A2" and "A3" and "B3". Suppose we define a pair of samples x of . In this situation, if "A3" and "B3" are classified into the same cluster G as shown in FIG. 4A, the value of the evaluation function is increased by the second penalty P2 accordingly.

As shown in Equation (4), the value of the second penalty P2 is the reciprocal of the residual sum of squares between one sample x _i of the mismatched pair and the cluster center μ _k for a given cluster G _k . , and the reciprocal of the residual sum of squares between the other sample x _j and the cluster center μ _k for the same cluster G _k , summed for all clusters G _k .

This means that the size of the second penalty P2 is determined by considering the distance between the position of the sample x and the cluster center μ. That is, when classifying two samples x _i and _{x j} included _in a mismatched pair into the same cluster G _k , the value of the second penalty P2 is The smaller the distance to the cluster center μ _k , and the smaller the distance between the other sample x _j and the cluster center μ _k of the cluster G _k into which it is classified, the larger it becomes.

In the state quantity data classification device 1 of the present embodiment, the search unit 61 includes, as shown in FIG.

The degree of belonging update unit 62 calculates the degree of belonging u _ik for each given cluster center μ _k so as to minimize the evaluation function of equation (4). Although detailed formulas are omitted, the degree of affiliation u _ik is updated by a formula modified from the above formula (2) in consideration of the first penalty P1 and the second penalty P2.

The cluster center update unit 63 updates the cluster center μ _k of each cluster G _k based on the updated degree of belonging u _ik . This update can be done by exactly the same equation as equation (3) above.

The convergence determination unit 64 determines whether clustering has converged. Although there are various determination methods, similar to the fuzzy c-means method described above, it is determined that the search has converged if the position of each cluster center μ _k in the 18-dimensional space does not change by a predetermined distance or more even after updating. can do. However, it is not limited to this, and convergence is performed when the value of the evaluation function shown in the above equation (4) becomes equal to or less than the threshold, or when the degree of belonging u _ik does not fluctuate by more than a predetermined threshold even after updating. It can also be used as a condition for determination.

The search unit 61 alternately updates the degree of belonging u _ik by the degree of belonging update unit 62 and the cluster center μ _k by the cluster center update unit 63 until the convergence determination unit 64 determines that the clustering has converged. Iterate to Each time the update is repeated, the value of the evaluation function of formula (4) becomes smaller. This means that favorable clustering is approaching from the viewpoint of comprehensively considering the similarity of sample x, the first penalty, and the second penalty.

When the convergence determining unit 64 determines that the clustering has converged, the clustering is completed. By this clustering, after considering the content of the class identification information, samples x of physically identical detection targets belong to the same cluster G, and samples x of physically different detection targets belong to different clusters G. Sample x is classified as belonging to . That is, samples x can be sorted in terms of physical identity of the target. At the same time, this means that detection targets tracked by a plurality of sensors 2 and whose physical identity was unknown can be associated in terms of their physical identity.

After clustering is completed, the search unit 61 (cluster analysis unit 41) converts the cluster center μ _k of each cluster G _k and the degree of belonging u _ik to which each sample x _i belongs to each cluster G _k into the target state quantity Output to the estimation unit 71 .

The target state quantity estimation unit 71 performs so-called track fusion using the samples x classified by the cluster analysis unit 41 for each cluster G (in other words, for each target from a physical point of view). That is, as shown in FIG. 2, the target state quantity estimator 71 estimates the state quantity of each target on the basis of integrating the samples x included in the same cluster G. As shown in FIG.

The 18-dimensional vector indicated by the cluster center μ _k of each cluster G _k obtained by the cluster analysis unit 41 can be considered to indicate state quantity data related to each physical target, including time transition. can. In other words, the cluster center μ _k is the representative value of all the samples x belonging to the cluster G _k . Therefore, the target state quantity estimator 71 extracts the current state quantity data from the 18-dimensional vector of the cluster center μ _k , and outputs it as it is as the current position and velocity estimated for each physical target. can be done.

However, considering that the position and velocity detection errors by the respective sensors 2 are variously different, the target state quantity estimator 71 uses the degree of belonging u _ik that each sample x _i belongs to each cluster G _k . may be used to estimate position and velocity for physical targets. The target state quantity estimator 71 can also estimate the position and velocity of the target using a known tracking filter using a Kalman filter, particle filter, or the like.

The display data generation unit 81 can generate display data for graphically displaying the result estimated by the target state quantity estimation unit 71 on a display device (not shown). The data generated by the display data generator 81 is output from the state quantity data classification device 1 to the display device in real time.

For example, as shown in FIG. 5, the data generated by the display data generation unit 81 indicates the position of the own machine by a symbol S1, and the position of the physical target acquired by the clustering by the state quantity data classification device 1 by a symbol S2. , S3, S4. The speed of the target is represented by the directions and lengths of arrows attached to the symbols S2, S3 and S4. Instead of displaying the positions and velocities outputted by the multiple sensors 2 as they are, the positions and velocities of the targets organized by the state quantity data classification device 1 from the viewpoint of physical identity are displayed on the display device. , the operator can simply recognize the situation.

As described above, the state quantity data classification device 1 of this embodiment includes the state quantity data input unit 11 and the cluster analysis unit 41 . The state quantity data input unit 11 acquires state quantity data output by detecting the state quantity of the target by the plurality of sensors 2 . The cluster analysis unit 41 uses the state quantity data and the class identification information indicating to which class the state quantity data relates to the target when the targets are classified into a plurality of categories from some point of view, including the time transition. Classify the state quantity data (sample x). The cluster analysis unit 41 classifies the samples x indicating mutually similar state quantities into the same cluster G, classifies the samples x with the same class identification information into the same cluster G, and classifies the samples x with different class identification information into the same cluster G. Classify the sample x such that the sample x falls into different clusters G;

As a result, in addition to the similarity of the state quantity data (sample x), by considering whether or not the class identification information of the targets related to the state quantity data match each other, the physical identity of the target can be obtained. can suitably classify the state quantity data.

Further, in the state quantity data classification device 1 of the present embodiment, the cluster analysis unit 41 differentiates the similarity between the samples x classified into the respective clusters G and the plurality of samples x having the same class identification information. A sample x is classified in consideration of a first penalty P1 for classifying into a cluster G and a second penalty P2 for classifying a plurality of samples x having different class identification information into the same cluster G.

This makes it possible to easily classify the samples x taking into consideration the degree of similarity between the samples x and penalties in a comprehensive manner.

In addition, in the state quantity data classification device 1 of the present embodiment, the first term (value based on the similarity) of the evaluation function, Equation (4), is obtained when the samples x classified into the respective clusters G are similar to each other. becomes smaller. The second term (value based on the first penalty P1) of Equation (4) becomes large when two samples x with matching class identification information are classified into different clusters G. The third term (value based on the second penalty P2) of Equation (4) becomes large when two samples x with different class identification information are classified into the same cluster G. The cluster analysis unit 41 searches for a class of the sample x that reduces the sum of the first, second, and third terms of Equation (4).

This makes it possible to classify samples x from a comprehensive point of view with simple calculations.

Further, in the state quantity data classification device 1 of the present embodiment, the first penalty P1 is, when classifying two samples x having the same class identification information into different clusters G, one of the two samples x and The smaller the distance from the cluster center μ of the cluster G to be classified, and the smaller the distance between the other sample x and the cluster center μ of the cluster G into which it is classified, the stricter the criteria.

Thereby, the weight of the first penalty P1 can be determined appropriately.

Further, in the state quantity data classification device 1 of the present embodiment, the second penalty P2, when classifying two samples x with different class identification information into the same cluster G, one of the two samples x The smaller the distance from the cluster center μ of the cluster G into which it is classified, and the smaller the distance between the other sample x and the cluster center μ of the cluster G into which it is classified, the stricter it is determined.

Thereby, the weight of the second penalty P2 can be determined appropriately.

Further, the state quantity data classification device 1 of the present embodiment performs target state quantity estimation for acquiring the state quantity of the target corresponding to the cluster G based on the samples x classified into the same cluster G by the cluster analysis unit 41. A portion 71 is provided.

As a result, the state quantity of the target can be obtained appropriately by using the clustering result.

In addition, the state quantity data classification device 1 of the present embodiment includes a data preprocessing unit 21 that obtains state quantity data at a timing common to the plurality of sensors 2 based on the state quantity data acquired by the state quantity data input unit 11. Prepare.

Accordingly, by obtaining the state quantity data at the common timing from the state quantity data output by the plurality of sensors 2, the cluster analysis unit 41 can appropriately compare the state quantity data for classification. .

Further, in the present embodiment, state quantity data is classified by a state quantity data classification method including the following state quantity data input step and state quantity data classification step. In the state quantity data input step, the plurality of sensors 2 detect the state quantity of the target and acquire the state quantity data output. In the state quantity data classification step, when the targets are classified into a plurality of categories from some point of view, class identification information indicating to which category the state quantity data relates to the target is used to classify the target including time transition. classify the state quantity data (sample x). In the state quantity data classification step, samples x indicating state quantities that are similar to each other are classified into the same cluster G, samples x having the same class identification information are classified into the same cluster G, and the class identification information is Classify samples x such that different samples x fall into different clusters G.

As a result, in addition to the similarity of the state quantity data (sample x), by considering whether or not the class identification information of the targets related to the state quantity data match each other, the physical identity of the target can be obtained. can suitably classify the state quantity data.

Next, a second embodiment of the invention will be described. In the description of this embodiment, the same or similar members as those in the above-described embodiment are denoted by the same reference numerals in the drawings, and description thereof may be omitted.

In the clustering of the second embodiment, C N-dimensional vectors v _i (i=1, 2, . . . , C) satisfying the following equation (5) are obtained by searching. where C is the total number of clusters G; N is the total number of samples x.

The C N-dimensional vectors v _i indicate which cluster G the sample x belongs to. The j-th element v _ij of each vector v _i represents by 1 or 0 whether the j-th sample x _j belongs to the i-th cluster G _i . If the j-th sample x _j belongs to the i-th cluster G _i , then v _ij =1, otherwise v _ij =0. In this embodiment, unlike the first embodiment described above, v _ij does not take intermediate values between 0 and 1 (v _ij ε{0, 1}). All samples x belong to any one cluster G without fail. Therefore, when focusing on a certain sample x _j , the value v _ij representing whether or not the sample x _j belongs to the i-th cluster G _i is added for all clusters G to be 1.

v _i ^t denotes the transposed vector of vector v _i described above.

D is an N×N distance matrix that indicates the distance between two samples x taken from all samples x. An element of the distance matrix D, d _ij =d(x _i , x _j ), is the distance (e.g., Euclidean distance) between the i-th sample x _i and the j-th sample x _j in an 18-dimensional space. there is If this distance d _ij is small, it means that sample x _i and sample x _j are similar.

The inside of the curly braces in Equation (5) corresponds to the evaluation function of this embodiment. This evaluation function will be described below.

The first term of the evaluation function is for all clusters _G , the scalar value obtained by multiplying the transposed vector v _i ^t , the distance matrix D, and the vector v _i for the i-th cluster G i . value. The above scalar value for the i-th cluster G _i corresponds to the sum of all two combinations of the distances between two samples x extracted from all samples x belonging to the cluster G _i . Therefore, the first term means the total sum of the distances between samples x in each cluster G. This first term is the basic value of the evaluation function.

The second term of the evaluation function is a term that gives a first penalty P1 when two samples _xm and _xn forming a matched pair are classified into different clusters G, respectively.

For example, although the m-th sample x _m and the n-th sample x _n form a matched pair, the m-th sample x _m is in the third cluster G ₃ and the n-th sample x _n is in the 4 Assume that they are classified into the th cluster G ₄ . In this case, the value of v _im ·v _jn is 1 when i=3 and j=4, and 0 otherwise. The innermost Σ in the term of the first penalty P1 means to calculate the sum of the values of v _im ·v _jn ·d _mn for all matched pair samples x _m , x _n . Therefore, although the value of Σ is a matched pair, two samples x _{m and} x _n , one of which is classified into the i-th cluster G _i and the other into the j-th cluster G _j , are When there is, it coincides with d _mn , the distance d(x _m , x _n ) between the two samples x _m , x _n . Otherwise, the value of Σ is zero. The outer two Σ mean the sum of all combinations of two different clusters G (i-th cluster G _i and j-th cluster G _j ). Therefore, the first penalty P1 means the sum of the distances d(x _m , x _n ) between two samples x _m , x _n of all matching pairs classified into different clusters G.

The first penalty P1 greatly increases the evaluation function as the distance between the sample x _m and the sample x _n increases. Therefore, it can be said that the first penalty P1 is a penalty defined more severely as the distance between the samples x increases.

The third term of the evaluation function is a term that gives a second penalty P2 when two samples _xk and _xl forming a mismatching pair are classified into the same cluster G. FIG.

For example, although the k-th sample x _k and the l-th sample x _l form a mismatched pair, both the k-th sample x _k and the l-th sample x _l are in the third cluster G ₃ Suppose it is classified. In this case, the value of v _ik ·v _il is 1 when i=3 and 0 otherwise. The value of Σ inside the second penalty P2 term means to calculate the sum of the values of v _ik ·v _il ·(1/d _kl ) for all discordant pair samples x _k , x _l . . Therefore, the value of Σ is 1/d _kl when there are two samples x _k and x _l both of which are classified into the i-th cluster G _i , even though they form an inconsistent pair. , corresponds to the reciprocal of the distance d(x _k , x _l ) between the two samples x _k , x _l . Otherwise, the value of Σ is zero. The outer Σ means the sum over all clusters G (i-th cluster). Therefore, the second penalty P2 means the sum of the reciprocals of the distance d(x _k , x _l ) between two samples x _k , x _l of all mismatched pairs classified into the same cluster G. .

The second penalty P2 greatly increases the evaluation function as the distance between sample x _k and sample x _l decreases. Therefore, it can be said that the second penalty P2 is a penalty defined more severely as the distance between the samples x is smaller.

The evaluation function includes coefficients α ₁ and α ₂ for adjusting the effects of the first penalty P1 and the second penalty P2. The magnitude of each coefficient can be appropriately set under the conditions of α ₁ >0 and α ₂ >0.

In the present embodiment, the cluster analysis unit 41 randomly determines C vectors v _i , and then updates the contents of the vectors v _i to recalculate the value of the evaluation function while repeating the above equation Search for C vectors v _i (i=1, 2, . This makes it possible to determine which cluster G the sample x belongs to.

As described above, according to the present embodiment, when two samples x having the same class identification information are classified into different clusters G, the greater the distance between the samples x, the more severe the first penalty P1. is determined to be

Thereby, the weight of the first penalty P1 can be determined appropriately.

Further, according to the present embodiment, when two samples x having different class identification information are classified into the same cluster G, the second penalty P2 is determined to be severer as the distance between the samples x is smaller. .

Thereby, the weight of the second penalty P2 can be determined appropriately.

Although the preferred embodiments of the present invention have been described above, the above configuration can be modified, for example, as follows.

In the above embodiment, the temporal transition of the position and velocity of the target is represented by an 18-dimensional vector, and clustering is performed on a set of these 18-dimensional vectors (samples x). However, how to represent the state quantity data is arbitrary. As data transition, clustering may be performed on a set of 9×3=27-dimensional vectors.

Of the current positions and velocities of the target, the positions and velocities in the z-axis direction (that is, altitude direction) may be omitted, and clustering may be performed on a set of 4×3=12-dimensional vectors.

Regarding the sample x, it is possible to arbitrarily determine how many times including the present time the state quantity data should be changed over time. Clustering may be performed on samples in which only the current state quantity data is represented by a high-dimensional vector without including the transition of the state quantity at a predetermined time in the past.

In the above embodiment, matched pairs and mismatched pairs are created by considering both state quantity data and class identification information. Alternatively, however, matching and non-matching pairs can be created based solely on the matching (similarity) of class identifiers.

In the first embodiment, instead of randomly initializing the cluster centers μ _k of each cluster G _k , the cluster center initialization unit 52 uses, for example, a known k-means++ method to obtain a plurality of cluster centers μ _k can be determined so as not to be dense. In this case, good clustering and early convergence of clustering can be expected.

The state quantity data classification device 1 may output not only the current state quantity of each physical target but also the transition of the past state quantity. In this case, the movement trajectory of the target can be displayed on the display device.

The state quantity data classification device 1 may be configured to output a class estimated for each physical target based on the input class identification information. In this case, for example, it is possible to display near the symbol S2 on the display screen of FIG. 5 that the target related to this symbol S2 is a fishing boat of country X.

Instead of the fuzzy c-means method or the like, other known clustering methods such as the k-means method can be used for clustering.

The state quantity data classification device 1 may include an output unit that outputs the results of clustering and the like to an external device (for example, a recording device) different from the display device.

The state quantity data classification device 1 can be installed not only in aircraft but also in other moving bodies or ground facilities.

The state quantity data classification device 1 targets various targets, such as ships, aircraft, submersibles, and vehicles. The target may be a target related to the civil sector, such as civil ships, aircraft, and automobiles. Also, the target may be a military target such as a warship, a fighter plane, or a tank.

A configuration that classifies and integrates the target state quantities detected by a plurality of sensors 2 from the viewpoint of the physical identity of the target can also be applied to, for example, automatic operation of moving bodies, autonomous movement of robots, and the like. .

1 state quantity data classification device 2 sensor 11 state quantity data input unit (state quantity data acquisition unit)
41 cluster analysis unit (state quantity data classification unit)
G cluster

Claims (7)

a state quantity data acquisition unit for acquiring state quantity data output by detecting the state quantity of the target by a plurality of sensors;
A state quantity for classifying the state quantity data by using the state quantity data and category affiliation information indicating which category the target belongs to when the targets are classified into a plurality of categories from some point of view. a data classifier;
with
The state quantity data classification unit
Similarity between the state quantity data classified into each cluster;
a first penalty relating to classification evaluation when classifying a plurality of the state quantity data having the same section belonging information into different clusters;
a second penalty regarding classification evaluation when classifying a plurality of state quantity data with different classification affiliation information into the same cluster;
in view of,
classifying the state quantity data indicating state quantities that are similar to each other into the same cluster;
classifying the state quantity data with the same category affiliation information into the same cluster;
A state quantity data classification device for classifying the state quantity data having different classification affiliation information into different clusters. The state quantity data classification device according to claim 1 ,
The value based on the degree of similarity becomes small when the state quantity data classified into the respective clusters are similar to each other,
The value based on the first penalty increases when the first state quantity data and the second state quantity data with the same section affiliation information are classified into different clusters,
The value based on the second penalty increases when the third state quantity data and the fourth state quantity data having different classification affiliation information are classified into the same cluster,
The state quantity data classification unit searches for a classification of the state quantity data in which the sum of the value based on the degree of similarity, the value based on the first penalty, and the value based on the second penalty is small. A state quantity data classification device characterized by: The state quantity data classification device according to claim 2 ,
When classifying the first state quantity data and the second state quantity data having the same classification affiliation information into different clusters, the first penalty is that the center of the cluster into which the first state quantity data and the first state quantity data are classified is classified. The smaller the distance, the smaller the distance between the second state quantity data and the center of the cluster into which it is classified, or the larger the distance between the first state quantity data and the second state quantity data, the stricter the A state quantity data classification device characterized by being defined as follows. The state quantity data classification device according to any one of claims 1 to 3 ,
When classifying the third state quantity data and the fourth state quantity data having different classification affiliation information into the same cluster, the second penalty is that when the third state quantity data and the fourth state quantity data are classified into the same cluster, the third state quantity data and the center of the cluster into which the third state quantity data are classified The smaller the distance, the smaller the distance between the fourth state quantity data and the center of the cluster into which it is classified, the smaller the distance between the third state quantity data and the fourth state quantity data, the stricter the A state quantity data classification device characterized by being defined as follows. The state quantity data classification device according to any one of claims 1 to 4 ,
A state quantity data classification comprising a target state quantity acquiring unit for acquiring a state quantity of a target corresponding to the cluster based on the state quantity data classified into the same cluster by the state quantity data classifying unit. Device. The state quantity data classification device according to any one of claims 1 to 5 ,
A state quantity data classification device, comprising a data preprocessing unit that obtains the state quantity data at a timing common to the plurality of sensors based on the state quantity data acquired by the state quantity data acquisition unit. a state quantity data acquisition step of acquiring state quantity data output by detecting the state quantity of the target by a plurality of sensors;
A state quantity for classifying the state quantity data by using the state quantity data and category affiliation information indicating which category the target belongs to when the targets are classified into a plurality of categories from some point of view. a data classification step;
including
In the state quantity data classification step,
Similarity between the state quantity data classified into each cluster;
a first penalty relating to classification evaluation when classifying a plurality of the state quantity data having the same section belonging information into different clusters;
a second penalty regarding classification evaluation when classifying a plurality of state quantity data with different classification affiliation information into the same cluster;
in view of,
classifying the state quantity data indicating state quantities that are similar to each other into the same cluster;
classifying the state quantity data with the same category affiliation information into the same cluster;
A method of classifying state quantity data, wherein the state quantity data having different classification affiliation information are classified into different clusters.

Images