Target tracking maneuvering state identification method and device, and storage medium

Abstract

The invention provides a target tracking maneuvering state recognition method and device and a storage medium, and the method comprises the steps: S100, obtaining multi-frame radar target tracking data, and extracting a feature state vector of a target from the multi-frame radar target tracking data through a sliding window method; s200, building a multi-model classifier based on the improved projection twin support vector machine, and training the feature state vector; and S300, classifying and identifying the maneuvering state of the real-time test sample through a classifier, and outputting a target maneuvering state category. According to the technical scheme of the invention, multi-frame radar target detection data is utilized, through time accumulation, a sliding time window is adopted to extract plot characteristics with time correlation as features, and then a twin support vector machine identification method based on external penalty projection is adopted to carry out classification identification on a target maneuvering state. According to the method, the complicated workload and human errors of the attending personnel are remarkably reduced, so that the working process is clear, accurate and intelligent.

Description

Technical Field

[0001] The present application belongs to the field of automatic identification of target maneuvering state, and in particular relates to a target tracking maneuvering state identification method, device and storage medium. Background Art

[0002] With the rapid development of modern radar detection technology, target tracking environments are becoming highly dynamic, multi-dimensional, and data-intensive. Radar systems must process massive amounts of target information in real time, including multi-dimensional parameters such as target position, velocity, acceleration, heading, and echo intensity. Traditionally, the identification of target maneuvering states (such as starting, accelerating, descending, turning, and disappearing) relies primarily on manual experience: on-duty personnel must continuously monitor radar data streams, manually record key maneuvering nodes, and analyze trajectory intent. This approach has significant drawbacks: First, manual operation is inefficient and cannot meet the real-time requirements of high-frequency data updates. Second, subjective judgment can easily introduce human error, making it difficult to ensure accurate and consistent state identification, especially in complex maneuvering modes such as serpentine maneuvers and continuous direction changes. Third, the lack of standardized processes leads to significant differences in processing results across different scenarios.

[0003] Therefore, a target tracking maneuvering state recognition method is urgently needed to reduce manual dependence, improve decision-making efficiency, and adapt to the complex and changing radar tracking environment. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a target tracking maneuvering state identification method, equipment and storage medium, so as to reduce the complicated workload and human errors of the on-duty personnel through automatic identification and management of the target state, and make the work process clearer, more accurate and intelligent.

[0005] In a first aspect, the present application provides a target tracking maneuvering state recognition method, comprising:

[0006] Step S100: Acquire multiple frames of radar target tracking data and extract the target's characteristic state vector from the data using a sliding window method;

[0007] Step S200: constructing a multi-model classifier based on the improved projected twin support vector machine and training the feature state vector;

[0008] Step S300: classifying and identifying the maneuvering state of the real-time test sample by the classifier, and outputting a target maneuvering state category.

[0009] In an optional implementation, the target feature state vector includes the following dimensions:

[0010] a) The position, velocity, and acceleration of the target in three dimensions: range, azimuth, and elevation;

[0011] b) target heading angle;

[0012] c) Target echo intensity.

[0013] In an optional implementation, the method for configuring the measurement period of the sliding window according to the radar target tracking data acquisition frequency and extracting the characteristic state vector of the target at time i is:

[0014] v i =[r i a i e i v ri v ri v ai v ei a ri a ai a ei θ i β i ]

[0015] Where i≥1 and is a natural number; r, a, e, v r 、v r 、v a 、v e 、a r 、a a 、a e , θ, and β are the position, velocity, acceleration, target heading, and echo intensity of the target in three dimensions: distance, azimuth, and elevation, respectively.

[0016] In an optional implementation, at time i, the sliding window feature vector formed by the sliding window method is:

[0017] V i =[v i-n+1 v i-n+2 … v i ], i=n,n+1,…

[0018] Among them, vi is the test sample.

[0019] In an optional implementation, the improved projected twin support vector machine is a projected twin support vector machine based on external penalty, and its construction process includes:

[0020] Step S210: transforming the original linear programming problem into an unconstrained optimization problem through external penalty theory;

[0021] Step S220: introducing L1 norm regularization for feature selection and compressing redundant feature weights;

[0022] Step S230: Generate non-parallel interfaces to distinguish different maneuver categories.

[0023] In an optional implementation manner, step S210 includes:

[0024] Step S211: constructing the Lagrange function of the original linear programming problem;

[0025] Step S212: By introducing an external penalty term, the constraint conditions are integrated into the objective function to form an unconstrained optimization problem;

[0026] Step S213: solving the unconstrained optimization problem using the dual space to obtain analytical expressions of the interface parameters.

[0027] In an optional implementation, the L1 norm regularization term is implemented by the following objective function:

[0028]

[0029] Among them, ε is the feature compression parameter, which is used to control the weight attenuation strength of redundant features; C1 is the penalty coefficient, which is used to balance the classification error and model complexity.

[0030] In an optional implementation manner, the process of generating the non-parallel interface parameters includes:

[0031] Step S231: converting the optimization problem into a linear programming form by decomposing the weight vector wi into pi-qi;

[0032] Step S232: using the external penalty theory to solve in the dual space to obtain the interface weight vector wi and the bias term bi;

[0033] Step S233: Calculate interface parameters for different maneuver categories according to the formula:

[0034] In a second aspect, the present application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the method described in any one of the above technical solutions when executing the program.

[0035] In a third aspect, the present application provides a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the method described in any one of the above technical solutions.

[0036] Compared with the existing technology, the technical solution provided by this application has the following beneficial effects:

[0037] The target tracking and maneuvering state recognition method provided in this application utilizes multi-frame radar target detection data, accumulating data over time and using a sliding time window to extract time-correlated point trace characteristics as features. This algorithm then classifies and identifies the target's maneuvering state using an externally penalized projection twin support vector machine recognition method. This algorithm can automatically identify target maneuvering or state information, such as initiation, maneuvering, descent, turn, and disappearance, with high accuracy and recognition rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a schematic diagram of the target tracking trajectory in different maneuvering states in this application;

[0039] Figure 2 Schematic diagram of the sliding window method used in this application;

[0040] Figure 3 This is a schematic diagram of the target tracking maneuvering state recognition algorithm processing process in this application. DETAILED DESCRIPTION

[0041] Among existing automation technologies, some solutions attempt to use fixed threshold methods or single classification models (such as traditional support vector machines) for maneuver state recognition, but their limitations are obvious: the threshold method relies on preset parameters and cannot adapt to the nonlinear changes in the target's dynamic behavior. In particular, when the target's maneuver cycle varies greatly, the fixed nature of the segmentation window causes feature extraction to fail; the single classification model has insufficient generalization ability and has difficulty distinguishing similar maneuver modes (such as the combined action of acceleration and steering), and the processing efficiency of high-dimensional features is low, making it difficult to meet real-time requirements.

[0042] Furthermore, existing machine learning algorithms (such as classic support vector machines) typically require the construction of multiple binary classifiers when solving multi-classification problems, significantly increasing computational complexity. Furthermore, traditional models lack mechanisms for filtering redundant features and are susceptible to noise, further reducing classification accuracy.

[0043] Therefore, there is an urgent need for an efficient and convenient method for automatic recognition of target maneuvering status, which can extract effective features through the time correlation of multi-frame data and use an improved classification model to achieve high-precision real-time judgment, thereby reducing manual dependence, improving decision-making efficiency, and adapting to complex and changeable radar tracking environments.

[0044] The following describes the embodiments of the present invention through specific examples. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the details in this specification may be modified or altered based on different viewpoints and applications without departing from the spirit of the present invention.

[0045] It should be noted that the illustrations provided in this embodiment are only used to schematically illustrate the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0046] Example 1:

[0047] See also Figures 1 to 3 This embodiment provides a target tracking maneuvering state recognition method, including:

[0048] Step S100: Acquire multiple frames of radar target tracking data and extract the target's characteristic state vector from the data using a sliding window method.

[0049] Step S200: Construct a multi-model classifier based on the improved projected twin support vector machine and train the feature state vector.

[0050] Step S300: classifying and identifying the maneuvering state of the real-time test sample by the classifier, and outputting a target maneuvering state category.

[0051] Specifically, the sliding window method is first used to extract the characteristic state vector of the target, see Figure 1 ,The target maneuvering states mainly include acceleration, deceleration, lowering altitude, rising, turning, disappearing and other maneuvering categories.

[0052] The sliding window method, also known as the sliding window method, begins by using an inflection point detection algorithm to roughly determine the target's maneuver starting point. Detecting the starting point signifies the target has entered the maneuvering phase. Time windows are then used to continuously segment and identify the trajectory until the target's maneuvering ends. This time window mechanism is inspired by the human brain's ability to recognize continuous motion. Specifically, when faced with long-term motion changes, the human brain periodically or continuously provides feedback on the target's motion state, integrating this information to make a final judgment after the change occurs. For example, if a target performs a serpentine maneuver, the human brain will report that the target has turned, then performed the opposite turn, and then, after reviewing the overall information, conclude that the target has performed a serpentine maneuver. By simplifying this series of mental processes, a time window extraction model is formed.

[0053] Using a time window for extraction avoids the need to find a specific point at which the target's maneuver ends, and eliminates the problem of being unable to extract individual maneuver cycles from periodic maneuver patterns. However, because the time window is fixed, segmenting track types with varying time intervals (targets performing the same track type may take significantly different times) presents challenges, so the software requires a configurable time window.

[0054] Configure the sliding window measurement period according to the radar data acquisition frequency, and then extract the target's characteristic state vector, including the target's position, velocity, acceleration in three dimensions: range, azimuth, and elevation, as well as the target's heading and echo intensity. Extract the target's characteristic state vector at time i:

[0055] v i =[r i a i e i v ri v ri v ai v ei a ri a ai a ei θ i β i ]i=1,2,3,(1)

[0056] Among them, r, a, c, v r 、v r 、v a 、v e 、a r 、a a 、a e , θ, and β are the position, velocity, and acceleration of the target in three dimensions: distance, azimuth, and elevation, respectively. θ is the target heading, and β is the echo intensity.

[0057] Secondly, the sliding window feature state vector is obtained based on the target feature state vector.

[0058] The sliding window uses n measurement cycles to ensure n target states. For example, if the sliding window covers 3 frames, the target tracking estimation state at three observation moments is selected to extract the feature vector, thereby obtaining the target's sliding window feature state vector, as shown in formula (2).

[0059] At time i, the sliding window feature vector formed by the sliding window method is:

[0060] V i =[v i-n+1 v i-n+2 … v i ],i=n,n+1,…(2)

[0061] v i is the test sample x.

[0062] Subsequently, an external penalized projection twin support vector machine feature selection recognition algorithm EPFSPTSVM was constructed.

[0063] Support vector machines (SVMs), as a useful classification tool for supervised machine learning, have achieved significant results in practical applications such as drug discovery, face recognition, text classification, time series prediction, and regression estimation, demonstrating excellent generalization performance. Here, we propose an externally penalized projection twin support vector machine feature selection and recognition algorithm (EPFSPTSVM) for classifying target tracking maneuvers.

[0064] Assume an n-dimensional real space R n In the binary classification problem, we get a set of training data points, expressed as in Input j belongs to class i, and m=m1+m2, y j ∈{+1,-1} is the corresponding recognition output result. Further use the matrix Represents the input of m1 classes + 1, using the matrix represents the input of m2 classes - 1. The 2-norm of x is represented as ||x||, and the l1-norm of x is represented as ||x||1.

[0065] The method is to calculate the l1 norm distance between the samples of this class and the classification surface of this class, and at the same time incorporate the l1 norm Tikhonov regularization term into the projected twin support vector machine (PTWSVM). Specifically, EPFSPTSVM uses Instead of the PTWSVM objective function By incorporating ||w1||1 at the same time, the objective function of RFPTSVM can be constructed as follows:

[0066]

[0067] as well as

[0068]

[0069] Where ε is a feature compression parameter.

[0070] Given the following definition:

[0071]

[0072] EPFSPTSVM1(3) and EPFSPTSVM2(4) can be rewritten as:

[0073]

[0074] and

[0075]

[0076] Since the l1 norm is used, ||w iMany elements in ||1,i=1,2 will be close to zero, and the remaining non-zero elements in the weight vector are selected by the algorithm. We will discuss how to simplify and solve the formula, taking formula (7) as an example. Formula (7) will be transformed into an explicit LP problem, assuming

[0077] w1=p1-q1,p1,q1,r1,s1>0 Gw1=r1-s1#(9)

[0078]

[0079] The following LP problem can be obtained:

[0080]

[0081] where e∈R m is a unit column vector. Equation (11) is a standard LP problem, including 4(n+1)+(mm i ) variables, m1 equality constraints and 4(n+1)+(mm i ) inequality constraints, directly solving this problem requires high computational cost. Therefore, it can only be applied to small-scale classification problems. However, the external penalty theory (EP) can be used to transform it into an unconstrained minimization problem in the dual space, reducing the computational cost.

[0082] Here we discuss how to use the external penalty theory to solve the objective problem (7) of EPFSPTSVM. To calculate the dual problem of (7), we first give its Lagrange function:

[0083]

[0084] Among them, α i , i=1,2,…7 is the Lagrange multiplier vector. Solve the partial derivatives of (10) with respect to the variables (r,s,p,q,ξ) and set them to 0, and we get

[0085] e1+α2-α3=0

[0086] e1-α2-α4=0

[0087] εe-M T α1-G T α2-α5=0#(13)

[0088] εe+M T α1+G T α2-α6=0

[0089] C1e2-α1-α7=0

[0090] Substituting all equations in formula (13) into the Lagrange function (12), we can obtain the dual form of the original linear programming problem (7-8):

[0091]

[0092] set up

[0093]

[0094] Formula (14) can be rewritten as

[0095] max E T u

[0096] st-εe≤L T u≤εe#(16)

[0097] -S1≤u≤S2

[0098] in,

[0099]

[0100] In formula (17) are two zero vectors, the external penalty problem corresponding to the dual problem (16) can be expressed as:

[0101]

[0102] Where λ>0,||x|| represents the l2 norm of x, x + =max(0,x). Equation (18) is the external penalty problem of the initial equation (7).

[0103] For any λ>0, the external penalty problem (18) of the dual problem (7) is solvable. The solution u of formula (16) can produce the exact solution of the original problem (7) with respect to p, q, r, s:

[0104]

[0105] In formula (19), represents the first m1 rows of (·), and represents the last m-m1 rows of (·), and the projection vector solution is further obtained as:

[0106]

[0107] Similarly:

[0108]

[0109] In formula (21),

[0110]

[0111] The weight vectors w1 and w2 of the interface are obtained through (20) and (21) respectively.

[0112] The maneuvering states of the target include acceleration, deceleration, descent, elevation, turning, disappearance, etc. This embodiment uses the maneuvering training samples of each category and adopts formulas (20) and (21) to obtain w1 and w2 of different maneuvering interfaces respectively.

[0113] Finally, V i That is the test sample vector x, v at this time i =x The category of the test sample vector x is calculated using the following formula, that is, whether it belongs to this type of maneuver is determined using formula (22). In the formula, arg min is a small value. For example, when i = 1, the value of formula (22) is less than that of i = 2, indicating that category (x) = 1, that is, it belongs to this type of maneuver.

[0114]

[0115] In summary, see Figure 3 The main processing flow of the automatic identification of target tracking maneuver category described in this embodiment is as follows:

[0116] (1) Receive n frames (n can be 3-10) of target tracking estimated state in real time, and use the time sliding window method using formula (2) to generate the sliding window state vector V i ;

[0117] (2) Using the maneuver training samples of each category, use formulas (20) and (21) to obtain w1 and w2 of different maneuver interfaces respectively;

[0118] (3)V i That is, for the test sample vector x at this time, use formula (22) to obtain whether it belongs to this type of maneuver.

[0119] Example 2:

[0120] This embodiment provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the program, it implements the target tracking maneuver category automatic discrimination method provided in the first embodiment:

[0121] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for automatically distinguishing target maneuverability. The storage medium includes various media capable of storing program code, such as ROM, RAM, a magnetic disk, a USB flash drive, a memory card, or an optical disk.

[0122] In summary, the target tracking maneuvering state recognition method provided by this application utilizes multi-frame radar target detection data, accumulates time, and uses a sliding time window to extract time-correlated point trace characteristics as features. Subsequently, the target maneuvering state is classified and recognized through an external penalty projection twin support vector machine recognition method. This algorithm can automatically distinguish the target's starting, maneuvering, descent, turning, disappearance and other maneuvering or state information with high accuracy and recognition rate. Therefore, the technical solution provided by this application has a high industrial utilization value because it effectively overcomes the various shortcomings of the existing technology.

[0123] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical concepts disclosed in this application shall be covered by the claims of this application.

Claims

1. A target tracking maneuvering state recognition method, characterized in that: include: Step S100: Acquire multiple frames of radar target tracking data and extract the target's characteristic state vector from the data using a sliding window method; Step S200: constructing a multi-model classifier based on the improved projected twin support vector machine and training the feature state vector; Step S300: classifying and identifying the maneuvering state of the real-time test sample by the classifier, and outputting a target maneuvering state category.

2. The target tracking maneuvering state recognition method according to claim 1, characterized in that: The target feature state vector includes the following dimensions: a) The position, velocity, and acceleration of the target in three dimensions: range, azimuth, and elevation; b) target heading angle; c) Target echo intensity.

3. The target tracking maneuvering state recognition method according to claim 2, characterized in that: The method of configuring the sliding window measurement period according to the radar target tracking data acquisition frequency and extracting the characteristic state vector of the target at time i is: v i J[r i am i e i v ri v ri v ai v er am ri am ai am ei θ i β i ] Where i≥1 and is a natural number; r, a, e, v r 、v r 、v a 、v e 、a r 、a a 、a e , θ, and β are the position, velocity, acceleration, target heading, and echo intensity of the target in three dimensions: distance, azimuth, and elevation, respectively.

4. The target tracking maneuvering state recognition method according to claim 1, characterized in that: At time i, the sliding window feature vector formed by the sliding window method is: V i =[v i-n+1 v i-n+2 …v i ],i=n,n+1,… Among them, vi is the test sample.

5. The target tracking maneuvering state recognition method according to claim 1, characterized in that: The improved projected twin support vector machine is based on an externally penalized projected twin support vector machine. Its construction process includes: Step S210: transforming the original linear programming problem into an unconstrained optimization problem through external penalty theory; Step S220: introducing L1 norm regularization for feature selection and compressing redundant feature weights; Step S230: Generate non-parallel interfaces to distinguish different maneuver categories.

6. The target tracking maneuvering state recognition method according to claim 5, characterized in that: Step S210 includes: Step S211: constructing the Lagrange function of the original linear programming problem; Step S212: By introducing an external penalty term, the constraint conditions are integrated into the objective function to form an unconstrained optimization problem; Step S213: solving the unconstrained optimization problem using the dual space to obtain analytical expressions of the interface parameters.

7. The target tracking maneuvering state recognition method according to claim 5, characterized in that: The L1 norm regularization term is implemented by the following objective function: Among them, ε is the feature compression parameter, which is used to control the weight attenuation strength of redundant features; C1 is the penalty coefficient, which is used to balance the classification error and model complexity.

8. The target tracking maneuvering state recognition method according to claim 5, characterized in that: The generation process of the non-parallel interface parameters includes: Step S231: converting the optimization problem into a linear programming form by decomposing the weight vector wi into pi-qi; Step S232: using the external penalty theory to solve in the dual space to obtain the interface weight vector wi and the bias term bi; Step S233: Calculate interface parameters for different maneuver categories according to the formula:

9. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and when the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium storing computer program instructions, characterized in that: When the instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images