CN-122017768-A - Confidence-classification-based radar target true and false track distinguishing method

CN122017768ACN 122017768 ACN122017768 ACN 122017768ACN-122017768-A

Abstract

A method for discriminating true and false tracks of radar targets based on confidence classification belongs to the field of radar target recognition. Extracting five-dimensional characteristics of time difference, displacement distance, speed, acceleration and space distance in a temporary track set of a given radar target, constructing a deep neural network classifier based on a confidence function to obtain a confidence function of a true, false and uncertain category of each temporary track, selecting a plurality of neighbor decision samples according to the characteristic distance among the temporary tracks, constructing decision evidence by means of the corresponding confidence function, fusing, dividing two subsets according to the global confidence function of each temporary track obtained by fusing, and outputting a category discrimination result of each radar target temporary track. The deep neural network classifier based on the confidence function improves the mining and learning of the internal relation between the track characteristics and the model categories, supplements the number of decision samples through two-step decision making, effectively improves the recognition accuracy of the uncertain category temporary tracks, and realizes the rapid navigation establishment and stable tracking of targets.

Inventors

ZHANG YANG
HAN CHUNLEI
LU YAO
Qiao Dianfeng
GUO FENGJUAN
JIN ZHONGQIAN
YANG DI
ZHAO WANG

Assignees

中国电子科技集团公司第二十研究所

Dates

Publication Date: 20260512
Application Date: 20251223

Claims (8)

1. A method for discriminating true and false tracks of radar targets based on confidence classification is characterized by comprising the following specific steps: Step 1, constructing a category confidence identification framework of a radar target track; Step 2, extracting radar target track characteristics; step 3, designing a deep neural network classifier based on a confidence function; setting a deep neural network classifier according to radar target track characteristics and a category confidence identification framework; step 4, constructing target track decision evidence through a deep neural network classifier, and fusing the decision evidence to obtain a global confidence function set; and 5, dividing the global confidence function set into two subsets through two-step classification decision, judging the categories of the target tracks of the two subsets according to the confidence overall distribution to obtain the target tracks of uncertain categories, and finally completing the judgment of true and false tracks.
2. The method for discriminating true and false tracks of radar targets based on confidence classification according to claim 1, wherein in step 1, the classification confidence recognition framework for constructing radar target tracks is as follows: Based on the characteristic processing advantage of the confidence function theory on uncertain information, constructing a category confidence identification framework of radar target tracks , wherein, The true track class is indicated and, Representing the category of the false track, Representing an uncertainty class; the true target track category and the false target track category are mutually exclusive, and the category intersection operation is an empty set, namely ; The uncertain classes are the union of all independent repulsive classes, and under the framework of class confidence identification, Is that Therefore, if the intersection operation of each independent repulsive category and the uncertain category is self , 。
3. The method for determining whether the radar target is true or false according to claim 1, wherein in the step2, the process of extracting the radar target track features is as follows: the radar target track features include time differences Distance of displacement Speed and velocity of Acceleration of And space distance Five-dimensional features; The radar target track is composed of M radar tracks according to a time sequence, wherein the radar track characteristic P l at the moment I is as follows: (1); Wherein, the , , Respectively measuring target space position information after being converted into northeast coordinates, wherein t l is a time sequence label; M radar trace sets The method comprises the following steps: (2); Time difference between two radar traces of k+1 and k sequences The method comprises the following steps: (3); from the kth to k+1 time sequence, the displacement distance of two radar traces The method comprises the following steps: (4); kth time series target speed The method comprises the following steps: (5); kth time series target acceleration The method comprises the following steps: (6); Spatial distance of radar trace to origin of coordinates at kth time The method comprises the following steps: (7); five-dimensional feature vector of kth time sequence radar trace The method comprises the following steps: (8); Calculating the characteristic mean value of each dimension of all time sequence radar tracks contained in the track: (9); Finally, five-dimensional characteristic vectors of the radar target track are obtained The method comprises the following steps: (10)。
4. The method for judging whether the radar target is true or false according to the confidence classification of claim 1, wherein in the step 3, the deep neural network classifier comprises an input layer, an implicit layer and an output layer; The number of the nodes of the input layer is 5,5 nodes respectively correspond to time difference, displacement distance, target speed, target acceleration and space distance, and the input of the deep neural network classifier is five-dimensional characteristic vector of radar target track ; The number of layers of the hidden layer is 5, and the nodes of the layer 5 are both set to 20; the number of the nodes of the output layer is 3, and 3 nodes respectively correspond to the real track category Category of false track And uncertain category Confidence of (1) 、 And And meet the following I.e. the deep neural network classifier gives a confidence sum of 1 for all pattern classes.
5. The method for judging true and false tracks of radar targets based on confidence classification according to claim 1, wherein in step 4, the process of constructing target track decision evidence and fusing is as follows: Step 4.1, define To make discrimination of true or false The target tracks are obtained according to the radar target track feature extraction method in the step 2 Feature set of individual target tracks And characterizing the similarity d ij between two different target tracks by using the characteristic distance: (11); Step 4.2, obtaining a feature matrix D N according to the similarity between any tracks: (12); Since the characteristic distances correspond to each other, i.e. So the feature matrix D N is a symmetric matrix, and the element values on the main diagonal are all 0; Step 4.3, based on the characteristic array D N , any target track Selecting according to the similarity d ij Set of target tracks most similar to the set And corresponding set of similar target track features , wherein, ; Step 4.4, according to the similar target track characteristic set Obtaining through a deep neural network classifier A kind of electronic device A set of decision evidence: , wherein, , The three-dimensional confidence vector characterizes decision evidence to give true, false and uncertain support; representing the K decision evidence, based on D-S rule, combining two decision evidence And Fusion is carried out, and a combination formula is as follows: (13); Wherein, the And As a label of the decision evidence, , ; And Confidence of three categories of true, false and uncertain supported by each decision evidence before fusion , ; As a function of the confidence after the fusion, The confidence that the result obtained by fusing the decision evidence is true to the target track; the confidence coefficient of the result obtained by fusing the decision evidence on the target track is false; Confidence that the result obtained by fusing the decision evidence is of an uncertain type to the target track; after finishing all decision evidence fusion in turn, obtaining a target track Is a global confidence function of (2) Wherein, the method comprises the steps of, ; Step 4.5 for Set of individual target tracks Step 4.1 to step 4.4 are executed to obtain Global confidence function set for individual target tracks 。
6. The method for judging whether the radar target is true or false according to the confidence classification of claim 1, wherein in the step 5, the classification judgment process of the target tracks for the two subsets according to the confidence overall distribution is as follows: step 5.1, dividing the global confidence function set of the target track into two subsets according to whether the maximum confidence is an uncertain class And : (14); (15); Wherein, the Is that The category with the greatest confidence in it, ; If it is Will then Divided into In (a) and (b); If it is It is divided into In (a) and (b); Step 5.2, the All target tracks in the model are directly distributed to the real category with the maximum confidence coefficient Or false category For a pair of The category label is carried out and the method comprises the steps of, Category labels of (c) The method comprises the following steps: (16); Step 5.2, pair After completing pattern classification, respectively calculating all target tracks in the model Category center of (f) And Category center of (f) : (17); Wherein, the For a maximum confidence level of a true target track number, The maximum confidence is the false target track number; step 5.3 for In the target track, calculate the global confidence function Respectively with And Confidence distance of category center of (2), and is distributed to the mode category corresponding to the nearest center The method comprises the following steps: (18); ) Is that Is defined in the category center; and obtaining class judgment according to the overall distribution of confidence when the target tracks of the classes are uncertain, so that the judgment of reality or false is completed for all the target tracks.
7. An electronic device comprising one or more processors, memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs configured to perform the method of any of claims 1-6.
8. A computer readable storage medium storing program code which is callable by a processor to perform the method according to any one of claims 1 to 6.

Description

Confidence-classification-based radar target true and false track distinguishing method Technical Field The invention relates to a radar target identification technology, in particular to a method for distinguishing true and false tracks of radar targets based on confidence classification. Background Aiming at the problems that the radar is influenced by detection precision and complex environmental factors, the acquired detection information contains a large amount of clutter and uncertain information, a large amount of false tracks are generated when a target is built, and the target tracking precision is greatly influenced, a confidence identification framework capable of effectively representing the uncertain information needs to be researched, and a support is provided for improving the true and false track discrimination of the radar target by combining the strong data feature mining capability of a deep neural network. The traditional target track recognition algorithm adopts two categories which are not the same, and the lack of characterization processing of inaccurate and incomplete data restricts the improvement of recognition accuracy. The invention provides a radar target true and false track distinguishing method based on confidence classification, which comprises the steps of firstly extracting five-dimensional characteristics of time difference, displacement distance, speed, acceleration, space distance and the like of each temporary track, secondly constructing a deep neural network classifier based on confidence functions to obtain confidence functions of true, false and uncertain classes of each track, then selecting a plurality of neighbor decision samples based on characteristic distances among tracks, constructing decision evidence and fusing, and then dividing the temporary tracks into two subsets according to global confidence functions to carry out two-step decision, wherein the maximum confidence is that the true or false temporary tracks are firstly directly judged in class, and the maximum confidence is uncertain and then judged according to the class center distance. And finally, outputting the classification discrimination result of each temporary track. The related known technology on which the invention is based is confidence function theory, also called evidence theory or D-S theory. Confidence function theory has unique advantages in the characterization and processing of inaccurate, incomplete information. The basic confidence assignment functions of multiple pieces of evidence can be combined by D-S rules to obtain a new global mass function for decision analysis. The D-S rule is simple and easy to implement, and the algorithm can be quickly converged without considering the combination sequence when multiple evidences are fused. However, D-S has a limitation in the case of high evidence conflict, and Yager, smets, dubois & Prade, PCR5, DSmT, etc. have been studied by this learner. However, the combination rules have advantages and disadvantages, and are required to be reasonably selected or combined according to specific application scenes. The closest patent of the invention is a false track identification method based on a deep neural network. The method comprises the following specific implementation steps of (1) performing off-line processing on the acquired radar measurement data, and forming a target track state through algorithms such as data association, kalman filtering and the like. (2) And recording the target track state, finishing the classification mark, and forming a training sample. (3) And performing image conversion on the target track sample by using a two-dimensional grid model to form a target track binary image. The convolution neural network receives the binary image information, carries out convolution and pooling processing, and forms an image characteristic input signal to realize characteristic extraction of a target track on an image layer. (4) And respectively carrying out second-order fitting processing aiming at different state information in the target track sample to obtain second-order fitting functions of all states. The curve change rate of the second-order fitting function is a data characteristic input signal, and characteristic extraction of a target track on a data layer is achieved. (5) And inputting the image and data characteristic signals into a two-class neural network for training, completing the construction of the deep neural network, and ending the offline processing. (6) And (3) performing on-line processing on the real-time generated target track to finish image and data feature extraction, inputting the extracted feature signals into a deep neural network to perform classification calculation to obtain an identification result, and further determining the true and false attributes of the target track. (7) And removing false tracks according to the identification result, and reserving the real tracks, thereby solving the