Track feature modeling method for air target classification and behavior recognition

Through the method of expressing speed, turning amplitude and local behavioral characteristics, the complex identification problem of aerial target trajectory information is solved, and the accurate classification and identification of aerial targets is achieved, and technical support is provided for air situation awareness.

CN119004149AActive Publication Date: 2024-11-22CHENGDU LAVENDER TECHNOLOGY CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411028295.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-30
Publication Date
2024-11-22
Estimated Expiration
2044-07-30

AI Technical Summary

Technical Problem

It is difficult for the prior art to effectively identify and classify the flight behavior and categories of aerial targets, especially when high-dimensional aerial target trajectory information changes in complex changes.

Method used

The velocity feature representation, turning amplitude feature representation and local behavior feature representation are used to preprocess and feature extraction of aerial target trajectory data, and cluster or manual identification are clustered through feature joint to achieve classification and behavior recognition of aerial targets.

Benefits of technology

Accurate identification and classification of different categories of aerial targets is achieved, and fast and interpretable algorithms are provided, suitable for air situation awareness systems, assisting the military in real-time detection and analysis of air target behavior.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004149A_ABST
    Figure CN119004149A_ABST
Patent Text Reader

Abstract

The invention relates to the field of target trajectory prediction application, discloses a trajectory feature modeling method for air target classification and behavior recognition, and aims to realize fixed-length and interpretable feature representation of high-dimensional and different-scale air target trajectory time sequence information. Flight behaviors mainly include level flight and turning of civil aircrafts, diving of bombers, complex behaviors of fighters, o-type reconnaissance of reconnaissance aircrafts and other flight behaviors. Firstly, air target trajectory data are preprocessed; performing speed feature representation on the preprocessed data based on the speed ranges of different flight targets; based on different turning amplitude ranges of different flight targets, a one-way double-walk algorithm is designed to carry out turning feature representation on a flight track; carrying out local behavior coding on the flight path in combination with target flight path behavior priori, and obtaining local behavior characteristics; and finally, combining the obtained features, and carrying out unsupervised identification on aerial target behaviors and categories by adopting a clustering algorithm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of space-based infrared data processing and target trajectory prediction applications, and provides a trajectory feature modeling method for aerial target classification and behavior recognition, in particular, a process of predicting enemy combat ideas and plans by sensing target trajectory information through various battlefield sensors to conduct a comprehensive analysis of flight behavior and target categories, and specifically involves an identification method for extracting features of aerial maneuverable target flight behavior, performing cluster analysis on target categories, and thus judging enemy combat intentions. Background Art

[0002] With the prominent position of air supremacy in modern military warfare, the continuous evolution of the aerial battlefield situation and the renewal of weapons and equipment, quickly and accurately identifying the flight behavior of enemy targets and target categories is an important part of battlefield situation assessment. The aerial battlefield has become an important field of military games among various countries. In order to fully perceive the aerial battlefield situation, the continuous emergence of visible light, radar, and infrared imaging sensors has formed a full-dimensional and multi-dimensional detection capability for aerial targets. Due to the particularity of the band used for infrared imaging, it has the advantages of low dependence on the environment, can work all day, high sensitivity, and strong resistance to electromagnetic interference. Therefore, infrared imaging detectors are widely used in military and civilian fields in various countries.

[0003] The detection system of aerial targets based on space-based infrared early warning data is a point target. It has no geometric and texture characteristics, and it is difficult to realize the category recognition of aerial targets based on traditional image processing models. The maneuverability characteristics of aerial targets are obvious, and the movement rules are complex and changeable. For example, the trajectory of civil aircraft is smooth and linear, and the speed of medium and large bombers, early warning aircraft, and tankers is relatively stable, but the flight trajectory is S-shaped or "8"-shaped. The movement trajectory of small fighters is the most complex, with large speed changes and various flight actions in three-dimensional space, such as U-turns, dives, and somersaults. Therefore, the movement trajectory information of aerial targets detected by space-based infrared can be used to express the speed, turning amplitude, and local movement behavior characteristics of high-dimensional target movement trajectories, so as to realize the behavior judgment of complex flight trajectories of aerial targets and the category recognition of aircraft. Based on the above factors, this paper proposes a trajectory feature modeling method for aerial target classification and behavior recognition, which is of great significance in the situation awareness of the aerial battlefield. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a trajectory feature modeling method for aerial target classification and behavior recognition. In combination with the aerial target flight prior, speed feature representation, turning amplitude feature representation, and local motion behavior feature representation are used to perform fixed-dimensional feature expression on the high-dimensional aerial target flight trajectory information that is difficult to model directly, which facilitates subsequent behavior recognition and target recognition model modeling. At the same time, the algorithm has fast operation speed and is interpretable, and is easily integrated into the air situation awareness system, which can assist the military in real-time detection, analysis, and identification of aerial target behavior.

[0005] In order to solve the above technical problems and achieve the above objectives, the technical solution adopted by the present invention is as follows: A trajectory feature modeling method for aerial target classification and behavior recognition comprises the following steps:

[0006] Step 1: Preprocess the aerial target trajectory data and proceed to step 2;

[0007] Step 2: Perform velocity feature representation on the preprocessed target trajectory data, and then proceed to step 3;

[0008] Step 3: Perform turning feature representation on the preprocessed target trajectory data and proceed to step 4;

[0009] Step 4: Perform local behavior feature representation on the preprocessed target trajectory data and proceed to step 5

[0010] Step 5: Combine the speed feature vector, turning feature vector, and local behavior feature vector obtained in steps 2 to 4, that is, use clustering method for unsupervised recognition, or manually design logic for target recognition. In the above technical solution, step 1 specifically includes the following steps:

[0011] Step 1.1: Remove duplicate trajectory points from the target trajectory data set and go to step 1.2;

[0012] Step 1.2: Perform an equidistant interpolation operation on the new target trajectory data set obtained in step 1.1 to ensure that the time interval between two adjacent trajectory points is uniform, and obtain the final preprocessed target trajectory data. The preprocessed target trajectory data mainly includes the number of trajectory points N, the horizontal coordinate x and the vertical coordinate y of each trajectory point on the ground observation detector. Therefore, an aerial target trajectory can be described as G = [X, Y], where X = x1, x2, ..., X N , Y=y1,y2,...,y N .

[0013] In the above technical solution, the step 2 specifically includes the following steps:

[0014] Step 2.1: Using the aerial target trajectory data G obtained in step 1.2, use equations (1) and (2) to calculate the lateral offset vector D of the first N-1 trajectory points in the target trajectory: xdev and the longitudinal offset vector D ydev Formula (1) and formula (2) are as follows:

[0015] D xdev =[dx1, dx2, ..., dx N-1 ] = [x2, x3, ..., X N ]-[x1, x2, ..., X N-1 ] (1)

[0016] D ydev =[dy1,dy2,...,dy N-1 ]=[y2,y3,...,y N ]-[y1, y2, ..., y N-1 ] (2)

[0017] Among them, dx i =x i+1 -x i ,dy i =y i+1 -y i , they represent the lateral offset and longitudinal offset of the ith trajectory point within the trajectory point interval time t, respectively.

[0018] Step 2.2: Substitute the lateral offset vector D obtained in step 2.1 xdev and the longitudinal offset vector D ydev The velocity vector V = [v1, v2, ..., v N-1 ], formula (3) is as follows:

[0019]

[0020] Where t represents the time interval between two trajectory points, and the value of i is 1, 2, ..., N-1.

[0021] Step 2.3: Perform bin counting on the velocity vector V extracted in step 2.2, bins1 = [b0, b1, ..., b m ], that is, the speed is counted in m intervals, that is, (b0, b1], (b1, b2], ..., (b m-1 , b m ] The speed statistics distribution of each interval Vc = [c1, c2, ..., c m ].

[0022] Step 2.4: The velocity statistical distribution obtained in step 2.3 is used to obtain the final velocity feature vector φ through equation (4): v , formula (4) is as follows:

[0023]

[0024] in,

[0025] In the above technical solution, the step 3 specifically includes the following steps:

[0026] Step 3.1: Perform multiple local trajectory sampling on the aerial target trajectory data G obtained in step 1.2, with the number of sampling times being T = δN, where δ∈(0, 1), N is the total number of trajectory points. The minimum number of trajectory points in the local trajectory is set to S min , the maximum value is set to S max ;

[0027] Step 3.2: In each local trajectory sampling of step 3.1, firstly, the data [1, 2, ..., NS max ] are extracted with equal probability as the local trajectory initial point ST, and the value of the initial point in each local trajectory sampling is not repeated;

[0028] Step 3.3: In each local trajectory sampling of step 3.1, obtain S min To S max A random integer R is used as the number of trajectory points of the local trajectory, and the random integer R can be repeated in each sampling, so that the end point of the local trajectory is ED=ST+R-1;

[0029] Step 3.4: The local trajectory obtained by sampling each local trajectory in step 3.1 is represented as G local =[X local , Y lcocal ], the number of trajectory points is R, where X local =x′1,x′2,...,x′ R =x ST , x ST+1 , ..., x ED , Y local =y′1,y′2,...,y′ R =y ST ,y ST+1 , ..., y ED ;

[0030] Step 3.5: The local trajectory G obtained by step 3.4 local =[X local , Y local], calculate the initial direction d0 = (x′2-x′1, y′2-y′1) and the final direction d1 = (x′ S -x′ S-1 , y′ S -y′ S-1 ), and then the turning amplitude I of the local trajectory is calculated by formula (5), which is as follows:

[0031]

[0032] Step 3.6: The turning amplitudes of the T local trajectories obtained through steps 3.1 to 3.5 are then binned and counted, bins2 = [p0, p1, ..., p e ], that is, the turning range is (p0, p1], (p1, p2], ..., (p e-1 , p e ] Count each interval and obtain the turning amplitude vector Is = [I1, I2, ..., I e ], and the final turning feature φ is obtained through formula (6) z , formula (6) is as follows:

[0033]

[0034] In the above technical solution, the step 4 specifically includes the following steps:

[0035] Step 4.1: The aerial target trajectory data G obtained in step 1.2 is divided into n local trajectory data evenly starting from the initial point with a step size of S2, where It means rounding down, that is, discarding the local trajectories whose final trajectory points are less than S2.

[0036] Step 4.2: For each segment of local trajectory data obtained in step 4.1, use steps 3.4 and 3.5 to calculate its turning amplitude I. Set the threshold Th. If θ≥Th, encode the segment of local trajectory as A, which indicates that the segment of local trajectory is linear. Th is generally set to a value greater than 0.8. If θ<Th, determine the rotation direction of the local trajectory based on the parameters d0 and d1 calculated in step 3.3. If d0×d1>0, encode it as N, and if d0×d1<0, encode it as S. N indicates that the local trajectory rotates counterclockwise, and S indicates clockwise rotation. Finally, the encoding E=[e1,e2,...,e n ], where e i Represents the encoding of the i-th local trajectory.

[0037] Step 4.3: Encode the entire trajectory obtained in step 4.2 as E = [e1, e2, ..., e n ] to perform single-segment local trajectory encoding counting statistics. At this time, bin3 is designed as {(A), (S, N)}, where (A) represents the number of e in E. i Count the number of A codes, the value is λ1, (S, N) represents the number of e in E i The encoding of S or N is counted and the value is expressed as λ2.

[0038] Step 4.4: Encode the entire trajectory obtained in step 4.2 as E = [e1, e2, ..., e n ] and make adjacent combinations to obtain E′=[e1e2, e1e2, ..., e n-1 e n ], and count the combined trajectory encoding. At this time, the box bin4 is designed as {(AS, AN, SA, NA), (SS, NN), (SN, NS)}, where (AS, AN, SA, NA) means that the combination of straight line and turn or turn and straight line is counted in the box (AS, AN, SA, NA), and the count value is λ3. (SS, NN) means that the combination of continuous clockwise turn or counterclockwise turn is counted in the box (SS, NN), and the count value is λ4. (SN, NS) means that the snake-like behavior is counted in the box (SN, NS), and the count value is λ5.

[0039] Step 4.5: The final local behavior feature φ is obtained by using formula (7) for λ1,λ2,λ3,λ4,λ5 obtained in steps 4.3 and 4.4. x , formula (7) is as follows:

[0040]

[0041] In the above technical solution, the step 5 specifically includes the following steps:

[0042] Step 5.1: The velocity characteristics φ obtained from steps 2.4, 3.6, and 4.5 respectively v 、 Turning feature φ z 、Local behavior characteristics φ x The feature representation of a trajectory is obtained by combining them, as shown in formula (8):

[0043] φ all =[φ v ,φ z ,φ x ]=[μ1,μ2,...,μ m ,η1,η2,...,η e ,κ1,κ2,...,κ5] (8)

[0044] Among them, the feature representation dimension of the entire trajectory is m+e+5.

[0045] Step 5.2: According to step 5.1, the final feature representation φ of the entire trajectory is obtained all , clustering methods can be used for classification, or the type of the trajectory and the category of the target can be identified through manual experience.

[0046] Because the present invention adopts the above technical solution, it has the following beneficial effects:

[0047] 1. Speed ​​feature representation: Through the technical solution of step 2, the present invention can accurately extract and represent the flight speed characteristics of different categories of aerial targets. Speed ​​is an important indicator of aircraft movement. Different categories of aircraft often have differences in speed due to their performance and mission requirements. By statistically analyzing the speed distribution in the flight trajectory data, different types of aircraft can be effectively identified and distinguished. For example, civil aircraft usually fly at a stable speed, while fighter jets may show large speed changes. The present invention helps to distinguish different flight target categories and flight behaviors from the speed distribution statistics through speed feature representation.

[0048] 2. Turn feature representation: The technical solution of step 3 enables the present invention to accurately extract and represent the turning features of different categories of aerial targets. The turning amplitude is an important indicator reflecting the maneuverability of an aircraft. Different categories of aircraft have significant differences in turning performance due to their design and mission requirements. By counting and analyzing the turning amplitude in the flight trajectory data, different types of aircraft can be effectively identified and distinguished. For example, fighter jets need to have strong maneuverability, so their turning amplitude is usually larger, while large transport aircraft or bombers have relatively small turning amplitudes. The present invention helps to distinguish different categories of flying targets and flight behaviors based on the turning amplitude through turning feature representation.

[0049] 3. Representation of local behavior characteristics: The technical solution of step 4 enables the present invention to accurately extract and represent the local flight behavior characteristics of different categories of aerial targets. The local behavior characteristics reflect the movement pattern of the aircraft in a specific time period. Different categories of aircraft have significant differences in local behavior characteristics due to their performance and mission requirements. By counting and analyzing the local behavior characteristics in the flight trajectory data, different types of aircraft can be effectively identified and distinguished. For example, fighter jets may exhibit complex flight maneuvers when performing specific tasks, while civil aircraft mainly fly in a smooth straight line. The present invention, through the representation of local behavior characteristics, helps to accurately identify specific categories of aerial targets through specific flight behavior models.

[0050] In summary, the present invention can effectively model and analyze the flight behavior of aerial targets by comprehensively utilizing speed characteristics, turning characteristics and local behavior characteristics, thereby realizing accurate identification and classification of different types of aircraft. This method has high practicality and reliability, and provides strong technical support for aerial battlefield situation awareness. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a schematic diagram of the aerial target behavior trajectory data used in the present invention;

[0052] Figure 2 It is a speed statistical distribution diagram of different flight speed intervals of aerial targets in the present invention;

[0053] Figure 3 It is a statistical distribution diagram of turning characteristics of aerial targets in the present invention;

[0054] Figure 4 This is a schematic diagram of the coding result of the aerial target behavior in the present invention;

[0055] Figure 5 It is a schematic diagram of the recognition result in the present invention;

[0056] Figure 6 Design a solution flow for the entire algorithm. DETAILED DESCRIPTION

[0057] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0058] The technical problem to be solved by the present invention is to provide a trajectory feature modeling method for aerial target classification and behavior recognition. The method combines the aerial target flight priors with speed feature representation, turning amplitude feature representation, and local motion behavior feature representation to perform fixed-dimensional feature expression on the high-dimensional aerial target flight trajectory information that is difficult to model directly, so as to facilitate subsequent behavior recognition and target recognition model modeling. The entire algorithm design process is as follows: Figure 6 As shown, the steps include:

[0059] Step 1.1: Remove duplicate trajectory points from the target trajectory data set and go to step 1.2;

[0060] In this example, after removing duplicate trajectory points from the target trajectory data, the number of trajectory points N=120, including 120 horizontal coordinates x and 120 vertical coordinates y.

[0061] Step 1.2: Perform an equidistant interpolation operation on the new target trajectory data set obtained in step 1.1 to ensure that the time interval between two adjacent trajectory points is uniform, and obtain the final preprocessed target trajectory data. The preprocessed target trajectory data mainly includes the number of trajectory points N, the horizontal coordinate x and the vertical coordinate y of each trajectory point on the ground observation detector. Therefore, an aerial target trajectory can be described as G = [X, Y], where X = x1, x2, ..., X N , Y=y1,y2,...,y N .

[0062] In this example, the preprocessed aerial target trajectory can be expressed as G = [X, Y], where X = x1, x2, ..., x 120 , Y=y1,y2,...,y 120 , which lists 20 flight trajectories after preprocessing such as Figure 1 shown.

[0063] Step 2.1: Using the aerial target trajectory data G obtained in step 1.2, use equations (1) and (2) to calculate the lateral offset vector D of the first N-1 trajectory points in the target trajectory: xdev and the longitudinal offset vector D ydev Formula (1) and formula (2) are as follows:

[0064] D xdev =[dx1, dx2, ..., dx N-1 ] = [x2, x3, ..., X N ]-[x1, x2, ..., X N-1 ] (1)

[0065] D ydev =[dy1,dy2,...,dy N-1 ]=[y2,y3,...,y N ]-[y1, y2, ..., y N-1 ] (2) Where dx i =x i+1 -x i ,dy i =y i+1 -y i , they represent the lateral offset and longitudinal offset of the ith trajectory point within the trajectory point interval time t, respectively.

[0066] In this example, dx i =x i+1 -x i ,dy i =y i+1 -y i, they represent the lateral and longitudinal offsets of the ith trajectory point within the trajectory point interval time t, where the time interval t = 1s. The velocity distribution results are as follows: Figure 2 shown.

[0067] Step 2.2: Substitute the lateral offset vector D obtained in step 2.1 xdev and the longitudinal offset vector D ydev The velocity vector V = [v1, v2, ..., v N-1 ], formula (3) is as follows:

[0068]

[0069] Where t represents the time interval between two trajectory points, and the value of i is 1, 2, …, N-1.

[0070] Step 2.3: Perform bin counting on the velocity vector V extracted in step 2.2, bins1 = [b0, b1, ..., b m ], that is, the speed is counted in m intervals, that is, (b0, b1], (b1, b2], ..., (b m-1 , b m ] The speed statistics distribution of each interval Vc = [c1, c2, ..., c m ].

[0071] In this example, the extracted speed vector V is binned and counted. The specific binning strategy is bins1 = [0, 100, 150, 200, 250, 300, 350, 400, 450, 500, 550], that is, the speed is counted in 10 intervals, that is, the statistical distribution Vc = [c1, c2, ..., c3, ..., c4, ..., c5, ..., c6, ..., c7, ..., c8, ..., c9, ..., c10, ..., c11, ..., c12, ..., c13, ..., c14, ..., c15, ..., c16, ..., c17, ..., c18, ..., c19, ..., c20, ..., c21, ..., c3 10 ].

[0072] Step 2.4: The velocity statistical distribution obtained in step 2.3 is used to obtain the final velocity feature vector φ through equation (4): v , formula (4) is as follows:

[0073]

[0074] in,

[0075] In this example, the velocity statistical distribution is obtained by equation (4) to obtain the final velocity 10-dimensional feature vector φ v =[μ1, μ2, ..., μ 10 ].

[0076] Step 3.1: Perform multiple local trajectory sampling on the aerial target trajectory data G obtained in step 1.2, with the number of sampling times being f = δN, where δ∈(0, 1), N is the total number of trajectory points. The minimum number of trajectory points in the local trajectory is set to S min , the maximum value is set to S max ;

[0077] In this example, the aerial target trajectory data G obtained in step 1.2 is subjected to multiple local trajectory samplings, the sampling times are T = δN, where the specific value δ = 0.5, the total number of trajectory points N = 120, S min =5, S max =10.

[0078] Step 3.2: In each local trajectory sampling of step 3.1, firstly, the data [1, 2, ..., NS max ] are extracted with equal probability as the local trajectory initial point ST, and the value of the initial point in each local trajectory sampling is not repeated;

[0079] Step 3.3: In each local trajectory sampling of step 3.1, obtain S min To S max A random integer R is used as the number of trajectory points of the local trajectory, and the random integer R can be repeated in each sampling, so that the end point of the local trajectory is ED=ST+R-1;

[0080] Step 3.4: The local trajectory obtained by sampling each local trajectory in step 3.1 is represented as G local =[X local , Y local ], the number of trajectory points is R, where X local =x′1,x′2,...,x′ R =x ST , X ST+1 , ..., x ED , Y local =y′1,y′2,...,y′ R =y ST ,y ST+1 , ..., y ED .

[0081] Step 3.5: The local trajectory G obtained by step 3.4 local =[X local , Y local ], calculate the initial direction d0 = (x′2-x′1, y′2-y′1) and the final direction d1 = (x′ S -x′ S-1 , y′ S -y′ S-1), and then the turning amplitude I of the local trajectory is calculated by formula (5), which is as follows:

[0082]

[0083] Step 3.6: The turning amplitudes of the T local trajectories obtained through steps 3.1 to 3.5 are then binned and counted, bins2 = [p0, p1, ..., p e ], that is, the turning range is (p0, p1], (p1, p2], ..., (p e-1 , p e ] Count each interval and obtain the turning amplitude vector Is = [I1, I2, ..., I e ], and the final turning feature φ is obtained through formula (6) z , formula (6) is as follows:

[0084]

[0085] In this example, the turning amplitude statistics of T = 60 local trajectories can be obtained, and then binned counting is performed, bin bins2 = [-1, 0, 0.2, 0.4, 0.6, 0.7, 0.8, 0.9, 1], that is, the turning amplitude is counted in each interval of (-1, 0], (0, 0.2], .., (0.9, 1], and the result is as follows Figure 3 As shown, the final turning feature φ is obtained by formula (6): z , and obtain the 8-dimensional feature vector Is = [I1, I2, ..., I8].

[0086] Step 4.1: The aerial target trajectory data G obtained in step 1.2 is divided into n local trajectory data evenly starting from the initial point with a step size of S2, where It means rounding down, that is, discarding the local trajectories whose final trajectory points are less than S2.

[0087] In this example, the aerial target trajectory data G obtained by step 1.2 starts from the initial point, the step length is set to S2=7, and the total number of trajectory points N=120.

[0088] Step 4.2: For each segment of local trajectory data obtained in step 4.1, use steps 3.4 and 3.5 to calculate its turning amplitude I. Set the threshold Th. If θ≥Th, encode the segment of local trajectory as A, which indicates that the segment of local trajectory is linear. Th is generally set to a value greater than 0.8. If θ<Th, determine the rotation direction of the local trajectory based on the parameters d0 and d1 calculated in step 3.3. If d0×d1>0, encode it as N, and if d0×d1<0, encode it as S. N indicates that the local trajectory rotates counterclockwise, and S indicates clockwise rotation. Finally, the encoding E=[e1,e2,...,e n ], where e i Represents the encoding of the i-th local trajectory.

[0089] In this example, for each local trajectory data obtained in step 4.1, the turning amplitude I is calculated using steps 3.4 and 3.5. The threshold Th is set to 0.9, where N indicates that the local trajectory is counterclockwise and S indicates clockwise. The encoding of the entire trajectory is obtained as E = [e1, e2, ..., e n ], where e i Represents the encoding of the i-th local trajectory, and the result is as follows Figure 4 shown.

[0090] Step 4.3: Encode the entire trajectory obtained in step 4.2 as E = [e1, e2, ..., e n ] to perform single-segment local trajectory encoding counting statistics. At this time, bin3 is designed as {(A), (S, N)}, where (A) represents the number of e in E. i Count the number of A codes, the value is λ1, (S, N) represents the number of e in E i The encoding of S or N is counted and the value is expressed as λ2.

[0091] Step 4.4: Encode the entire trajectory obtained in step 4.2 as E = [e1, e2, ..., e n ] and make adjacent combinations to obtain E′=[e1e2, e1e2, ..., e n-1 e n ], and count the combined trajectory encoding. At this time, the box bin4 is designed as {(AS, AN, SA, NA), (SS, NN), (SN, NS)}, where (AS, AN, SA, NA) means that the combination of straight line and turn or turn and straight line is counted in the box (AS, AN, SA, NA), and the count value is represented by λ3. (SS, NN) means that the combination of continuous clockwise turn or counterclockwise turn is counted in the box (SS, NN), and the count value is represented by λ4. (SN, NS) means that the snake-like behavior is counted in the box (SN, NS), and the count value is represented by λ5.

[0092] Step 4.5: The final local behavior feature φ is obtained by using formula (7) for λ1,λ2,λ3,λ4,λ5 obtained in steps 4.3 and 4.4. x , formula (7) is as follows:

[0093]

[0094] In the above technical solution, the step 5 specifically includes the following steps:

[0095] In this example, the λ1, λ2, λ3, λ4, and λ5 obtained in steps 4.3 and 4.4 are used to obtain the final 5-dimensional feature vector φ of the local behavior feature. x =[κ1,κ2,...,κ5].

[0096] Step 5.1: The velocity characteristics φ obtained from steps 2.4, 3.6, and 4.5 respectively v 、 Turning feature φ z 、Local behavior characteristics φ x The feature representation of a trajectory is obtained by combining them, as shown in formula (7):

[0097] φ all =[φ v ,φ z ,φ x ]=[μ1,μ2,...,μ m ,η1,η2,...,η e , k1, κ2, ..., κ5] (8)

[0098] Among them, the feature representation dimension of the entire trajectory is m+e+5.

[0099] Step 5.2: According to step 5.1, the final feature representation φ of the entire trajectory is obtained all , clustering methods can be used for classification, or the type of the trajectory and the category of the target can be identified through manual experience.

[0100] In this example, the k-means clustering method is used for classification, which is divided into 3 categories. The results are as follows Figure 5 shown.

Claims

1. A trajectory feature modeling method for aerial target classification and behavior recognition, characterized by: The steps include: Step 1: Preprocess the aerial target trajectory data; Step 2: Perform velocity feature representation on the preprocessed target trajectory data; Step 3: Perform turning feature representation on the preprocessed target trajectory data; Step 4: Represent the local behavior features of the preprocessed target trajectory data; Step 5: Combine the speed feature vector, turning feature vector, and local behavior feature vector obtained in steps 2 to 4, perform unsupervised recognition through clustering method, or design logic for target recognition.

2. The trajectory feature modeling method for aerial target classification and behavior recognition according to claim 1 is characterized in that: The step 1 comprises the following steps: Step 1.1: Remove duplicate trajectory points from the target trajectory data set and go to step 1.2; Step 1.2: Perform an equidistant interpolation operation on the new target trajectory data set obtained in step 1.1 to ensure that the time interval between two adjacent trajectory points is uniform, and obtain the final preprocessed target trajectory data. The preprocessed target trajectory data includes the number of trajectory points N, the horizontal coordinate x and the vertical coordinate y of each trajectory point on the ground observation detector. An aerial target trajectory is described as G = [X, Y], where X = x1, x2, ..., X N , Y=y1,y2,...,y N .

3. The trajectory feature modeling method for aerial target classification and behavior recognition according to claim 2 is characterized in that: The step 2 specifically includes the following steps: Step 2.1: Using the aerial target trajectory data G obtained in step 1.2, use equations (1) and (2) to calculate the lateral offset vector D of the first N-1 trajectory points in the target trajectory: xdev and the longitudinal offset vector D ydev , formula (1) and formula (2) are as follows: D xdev =[dx1,dx2,...,dx N-1 ]=[x2,x3,...,x N ]-[x1,x2,...,x N-1 ] (1) <h2 style=";text-align:left;direction:ltr">D<h2 style=";text-align:left;direction:ltr"> ydev <h2 style=";text-align:left;direction:ltr"> (dy1, dy2,..., dy)<h2 style=";text-align:left;direction:ltr"> N-1 <h2 style=";text-align:left;direction:ltr"> ]=[y2,y3,…,y<h2 style=";text-align:left;direction:ltr"> N <h2 style=";text-align:left;direction:ltr"> ]-[y1, y2,…, y<h2 style=";text-align:left;direction:ltr"> N-1 <h2 style=";text-align:left;direction:ltr"> ] (2) Among them, dx i =x i+1 -x i ,dy i =y i+1 -y i , they represent the lateral and longitudinal offsets of the ith trajectory point within the trajectory point interval time t, respectively; Step 2.2: Substitute the lateral offset vector D obtained in step 2.1 xdev and the longitudinal offset vector D ydev The velocity vector V = [v1, v2, ..., v N-1 ], formula (3) is as follows: Where t represents the time interval between two trajectory points, and the value of i is 1, 2, ..., N-1; Step 2.3: Perform bin counting on the velocity vector V extracted in step 2.2, bins1 = [b0, b1, ..., b m ], that is, the speed is counted in m intervals, that is, (b0, b1], (b1, b2], ..., (b m-1 , b m ] The speed statistics distribution of each interval Vc = [c1, c2, ..., c m ]; Step 2.4: The velocity statistical distribution obtained in step 2.3 is used to obtain the final velocity feature vector φ through equation (4): v , formula (4) is as follows: in, 4. The trajectory feature modeling method for aerial target classification and behavior recognition according to claim 2 is characterized in that: The step 3 specifically comprises the following steps: Step 3.1: Perform multiple local trajectory sampling on the aerial target trajectory data G obtained in step 1.

2. The number of samplings is T = δN, where δ∈(0, 1), N is the total number of trajectory points, and the minimum number of trajectory points in the local trajectory is set to S min , the maximum value is set to S max ; Step 3.2: In each local trajectory sampling of step 3.1, firstly, the data [1, 2, ..., NS max ] is extracted with equal probability as the local trajectory initial point ST, and the value of the initial point in each local trajectory sampling is not repeated, where N-Smax is the maximum starting point of the local sampling; Step 3.3: In each local trajectory sampling of step 3.1, obtain S min To S max A random integer R is used as the number of trajectory points of the local trajectory, and the random integer R can be repeated in each sampling, so that the end point of the local trajectory is ED=ST+R-1; Step 3.4: The local trajectory obtained by sampling each local trajectory in step 3.1 is represented as G local =[X local , Y local ], the number of trajectory points is R, where X locol =x′1,x′2,…,x′ R =x ST , x ST+1 , ..., x ED , Y local =y′1,y′2,...,y′ R =y ST ,y ST+1 , ..., y ED ; Step 3.5: The local trajectory G obtained by step 3.4 local =[X local , Y local ], calculate the initial direction d0 = (x′2-x′1, y′2-y′1) and the final direction d1 = (x′ S -x′ S-1 , y′ S -y′ S-1 ), and then the turning amplitude I of the local trajectory is calculated by formula (5), which is as follows: Step 3.6: The turning amplitudes of the T local trajectories obtained through steps 3.1 to 3.5 are then binned and counted, bins2 = [p0, p1, ..., p e ], that is, the turning range is divided into (p0, p1], (p1, p2], ..., (p e-1 , p e ] Count each interval and obtain the turning amplitude vector Is = [I1, I2, ..., I e ], and the final turning feature φ is obtained through formula (6) z , formula (6) is as follows:

5. The trajectory feature modeling method for aerial target classification and behavior recognition according to claim 4 is characterized in that: The step 4 specifically comprises the following steps: Step 4.1: The aerial target trajectory data G obtained in step 1.2 is divided into n local trajectory data evenly starting from the initial point with a step size of S2, where It means rounding down, that is, discarding the local trajectories with fewer final trajectory points than S2; Step 4.2: For each local trajectory data obtained in step 4.1, use steps 3.4 and 3.5 to calculate its turning amplitude I, set the threshold Th, if θ≥Th, encode the local trajectory as A, which indicates that the local trajectory is linear, Th is set to a value greater than 0.8, if θ<Th, judge the rotation direction of the local trajectory according to the parameters d0 and d1 calculated in step 3.3, if d0×d1>0, encode it as N, if d0×d1<0, encode it as S, where N indicates that the local trajectory rotates counterclockwise, and S indicates clockwise rotation, and finally the encoding E=[e1,e2,...,e n ], where e i Represents the encoding of the i-th local trajectory: Step 4.3: Encode the entire trajectory obtained in step 4.2 as E = [e1, e2, ..., e n ] to perform single-segment local trajectory encoding counting statistics. At this time, bin3 is designed as {(A), (S, N)}, where (A) represents the number of e in E. i Count the number of A codes, the value is λ1, (S, N) represents the number of e in E i Count the codes of S or N, and the value is expressed as λ2; Step 4.4: Encode the entire trajectory obtained in step 4.2 as E = [e1, e2, ..., e n ] and make adjacent combinations to obtain E′=[e1e2, e1e2, ..., e n-1 e n ], and count the combined trajectory encoding. At this time, the box bin4 is designed as {(AS, AN, SA, NA), (SS, NN), (SN, NS)}, where (AS, AN, SA, NA) means that the combination of straight line and turn or turn and straight line is counted in the box (AS, AN, SA, NA), and the count value is λ3, (SS, NN) means that the combination of continuous clockwise turn or counterclockwise turn is counted in the box (SS, NN), and the count value is λ4, (SN, NS) means that the snake-like behavior is counted in the box (SN, NS), and the count value is λ5; Step 4.5: The final local behavior feature φ is obtained by using formula (7) for λ1, λ2, λ3, λ4, and λ5 obtained in steps 4.3 and 4.

4. x , formula (7) is as follows:

6. The trajectory feature modeling method for aerial target classification and behavior recognition according to claim 1 is characterized in that: The step 5 specifically comprises the following steps: Step 5.1: The velocity characteristics φ obtained from steps 2.4, 3.6, and 4.5 respectively v 、 Turning feature φ z 、Local behavior characteristics φ x The feature representation of a trajectory is obtained by combining them, as shown in formula (7): f dl =[φ x ,f z ,f x ]=[μ1,μ2,…,μ m ,n1,n2,…,n e ,k1,k2,…,k5] (8) Among them, the feature representation dimension of the entire trajectory is m+e+5; Step 5.2: According to step 5.1, the final feature representation φ of the entire trajectory is obtained all , clustering methods are used for classification, or the type of the trajectory and the category of the target are identified through artificial experience.

Citation Information

Patent Citations

  • Target behavior clustering method and system

    CN111506627A

  • Aerial target activity rule prediction method

    CN114330509A

  • Flight target trajectory classification method based on bottleneck neural network embedding

    CN116091813A

  • Interpretable two-stage ship trajectory prediction method based on historical navigation mode mining

    CN116227652A

  • Abnormal behavior recognition model training method and device and abnormal behavior recognition method and device

    CN116451050A