A method for establishing a target tracking model for an aircraft

Abstract

The application belongs to the field of aircraft target tracking, and discloses a method for establishing a target tracking model for an aircraft. First, a target collection area is established based on a predetermined target object. Then, a sample collection circulation matrix is established based on the target collection area and the predetermined target object to collect positive and negative samples and obtain target object samples. Then, a sample area scaling model is established, and based on the target object samples of the tracking image, the flight altitude value corresponding to the tracking image, and the sample area scaling model, the real value of the target object samples of the tracking image is obtained. Then, a Kalman filter model is established, and when the tracking image has feedback on the predetermined target object, the Kalman filter model is iterated. When the number of acquired tracking images is greater than a predetermined number, and the tracking image has no feedback on the predetermined target object, corresponding predicted value data is filled based on the Kalman filter model until the tracking camera ends tracking of the predetermined target object.

Description

Technical Field

[0001] This invention belongs to the field of aircraft target tracking, and specifically relates to a method for establishing a target tracking model for aircraft. Background Technology

[0002] Aircraft can mimic the flight of birds in nature, achieving prolonged low-speed, hovering flight. With the development of microelectromechanical systems (MEMS) and artificial intelligence, aircraft-based visual target tracking systems are finding increasingly widespread applications in military reconnaissance and civilian fields.

[0003] Currently, due to the diverse scenarios in which aircraft perform missions, including varying lighting conditions, rain and snow, and different target environments, coupled with the small size of the targets being tracked, the accuracy of target tracking algorithms is relatively poor. In particular, during long-term tracking, target occlusion may occur, preventing the vision module from receiving feedback from the target object. This results in the obtained data not accurately corresponding to the target object. Furthermore, the existing recognition and tracking algorithms require a significant amount of computing resources during the aircraft's target tracking process. This not only increases the likelihood of system failures during continuous use but also makes the related hardware components bulky. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for establishing a target tracking model for aircraft. This method uses a relatively simple cyclic acquisition matrix, combined with the aircraft's built-in altitude sensor and Kalman prediction, to achieve accurate correspondence with the target object. It also greatly reduces the computing resources required in the identification and tracking calculation, thereby significantly reducing system failures during continuous use and enabling significant miniaturization of the related hardware.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for establishing a target tracking model for an aircraft, characterized by comprising the following steps:

[0007] Step S1: Based on the size of the pixel group occupied by the predetermined target object as fed back in the first frame tracking image, establish the target object acquisition area, and the pixel group occupied by the predetermined target object is contained in the pixel group occupied by the target object acquisition area.

[0008] Step S2: Establish a sample collection cycle matrix based on the ratio of the number of pixels occupied by the target collection area and the predetermined target object. Use the sample collection cycle matrix to collect positive and negative samples in the target object collection area to obtain the target object sample corresponding to the predetermined target object.

[0009] Step S3: Establish a sample area scaling model. Based on the target object sample of the current frame tracking image, the flight altitude value corresponding to the current frame tracking image, and the sample area scaling model, obtain the scaling change ratio of the target object acquisition area of ​​the current frame relative to the previous frame tracking image, and obtain the true values ​​of the target object acquisition area and the target object sample of the current frame tracking image.

[0010] Step S4: Establish a Kalman filter model. When the current frame tracking image provides feedback to the predetermined target, the Kalman filter model is iterated and the error is corrected using the Kalman filter model, the true value of the target sample in the current frame tracking image, and the predicted value. When the number of acquired tracking images is greater than a predetermined number, and the current frame tracking image provides no feedback to the predetermined target, the target sample in the current frame tracking image is filled with the corresponding predicted value data based on the current Kalman filter model.

[0011] Step S5: Repeat steps S3 and S4 until the tracking camera stops tracking the predetermined target.

[0012] Preferably, in step S2, the positive and negative sample sampling process is as follows: in the tracking image, positive and negative samples are sampled from the pixel group occupied by the target object collection area through the sample collection cycle matrix, a classifier is trained based on the obtained sample set, the multiple response values ​​obtained by the classifier from multiple sample sets are compared, and the sample set corresponding to the highest response value is taken as the target object sample.

[0013] Preferably, in step S2, the loss function for the target sample is set as follows:

[0014]

[0015] L(y i ,f(x i Let λ1 be the loss function, λ1 be the preset regularization coefficient, f(x) be the linear classifier, f(x) = w·x + b, w be the weight matrix of the function to be fitted, b be the bias vector, and (x1,y1), (x2,y2), ..., (x... m ,y m Let be the sample label, r be the scaling numerator, x be the sampled value of the cyclic matrix, and a be the class factor, where a is constant relative to (1-f(x)). i )) r Directly proportional.

[0016] Preferably, the sample area scaling model is as follows:

[0017] f(x)=ωx 2 +ψx+ζ

[0018] ψ and ω are fitting coefficients, ζ is a fitting constant, f(x) is the scaling ratio of the target acquisition area in the current frame relative to the previous frame of the tracked image, and x is the flight altitude value.

[0019] Preferably, in step S4, the iterative process of the Kalman filter model is as follows:

[0020] Process T1: Use the predicted value of the target object sample by the Kalman filter model as the sample predicted value, and use the covariance matrix of the actual value and the sample predicted value of the target object sample by the Kalman filter model as the sample covariance matrix.

[0021] Process T2: Initialize the sample predicted values ​​and covariance matrix;

[0022] Process T3: Update and iterate the initialized sample prediction values ​​and covariance matrix, and update and iterate the Kalman gain matrix;

[0023] Process T4: Repeat processes T2 and T3 until the tracking camera stops tracking the predetermined target.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] 1. Because the method for establishing a target tracking model for an aircraft according to the present invention firstly establishes a target acquisition area based on the size of the pixel group occupied by the predetermined target object as reported in the first frame tracking image; then, a sample acquisition loop matrix is ​​established based on the ratio of the number of pixels occupied by the target acquisition area to the number of pixels occupied by the predetermined target object, and positive and negative samples are acquired in the target acquisition area through the sample acquisition loop matrix to obtain the target object sample corresponding to the predetermined target object; then, a sample area scaling model is established, and the true value of the target object sample in the current frame tracking image is obtained based on the target object sample in the current frame tracking image, the flight altitude value corresponding to the current frame tracking image, and the sample area scaling model; then, a Kalman filter model is established, and when the current frame tracking image has feedback on the predetermined target object, the Kalman filter model and the current frame tracking image are used to obtain the true value of the target object sample in the current frame tracking image. The true and predicted values ​​of the target object samples are used to correct the error of the Kalman filter model through model iteration. When the number of frames of the acquired tracking images is greater than a predetermined number, and the current frame tracking image does not provide feedback to the predetermined target object, the target object samples of the current frame tracking image are filled with corresponding predicted value data based on the current Kalman filter model. Finally, steps S3 and S4 are repeated until the tracking camera ends its tracking of the predetermined target object. Therefore, this invention can achieve accurate correspondence of the target object by using a relatively simple cyclic acquisition matrix, combined with the aircraft's own altitude sensor and Kalman prediction, and greatly reduces the computing power resources required in the recognition and tracking calculation. This not only greatly reduces the system failures that occur during continuous use, but also significantly miniaturizes the related hardware.

[0026] 2. Because the sample area scaling model of the target tracking model establishment method for aircraft of the present invention is:

[0027] f(x)=ωx 2 +ψx+ζ

[0028] Therefore, this invention obtains the scaling ratio of the target object sampling area by using a fitting function with the flight altitude value as a parameter, thereby eliminating the need to establish an additional algorithm model and further reducing the complexity of the entire algorithm and the required computing resources. Attached Figure Description

[0029] Figure 1 This is a schematic diagram illustrating the steps of a method for establishing a target tracking model for an aircraft according to an embodiment of the present invention. Detailed Implementation

[0030] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate a method for establishing a target tracking model for an aircraft according to this invention. It should be noted that the description of these embodiments is for the purpose of helping to understand this invention, but does not constitute a limitation of this invention.

[0031] like Figure 1 As shown, the target tracking model establishment method S100 for an aircraft in this embodiment includes a tracking camera and an altitude sensor on the aircraft. The tracking camera is used to track and capture images of a predetermined target, thereby forming multiple frames of tracking images. The altitude sensor is used to collect the flight altitude value of the aircraft relative to a predetermined reference horizon in real time. The method includes the following steps:

[0032] Step S1: Based on the size of the pixel group occupied by the predetermined target object as fed back in the first frame tracking image, establish the target object acquisition area, and the pixel group occupied by the predetermined target object is centrally contained within the pixel group occupied by the target object acquisition area.

[0033] Specifically, when the tracking camera starts tracking the predetermined target object, the client will display the tracking image in real time. At this time, a target object acquisition area is established based on the size of the pixel group occupied by the predetermined target object as fed back in the first frame of the tracking image. This target object acquisition area is displayed on the client as a tracking box containing the predetermined target object. In this embodiment, the vision module establishes a corresponding tracking box for each moving object appearing in the tracking image. The user selects a specific tracking box by touching it. At this time, the moving object corresponding to the tracking box is the predetermined target object.

[0034] Step S2: Establish a sample acquisition cycle matrix based on the ratio of the number of pixels occupied by the target acquisition area and the predetermined target object. Use the sample acquisition cycle matrix to acquire positive and negative samples in the target object acquisition area to obtain the target object sample corresponding to the predetermined target object.

[0035] The positive and negative sample sampling process is as follows:

[0036] In the tracking image, positive and negative samples are sampled from the pixel group occupied by the target object's sampling area through a sample acquisition cyclic matrix. A classifier is trained based on the obtained multiple sample sets. The multiple response values ​​obtained by the classifier from the multiple sample sets are compared, and the sample set corresponding to the highest response value is taken as the target object sample.

[0037] The loss function for the target sample is set as follows:

[0038]

[0039] L(y i ,f(x iLet λ1 be the loss function, λ1 be the preset regularization coefficient, f(x) be the linear classifier, f(x) = w·x + b, w be the weight matrix of the function to be fitted, b be the bias vector, and (x1,y1), (x2,y2), ..., (x... m ,y m Let be the sample label, r be the scaling numerator, x be the sampled value of the cyclic matrix, and a be the class factor, where a is constant relative to (1-f(x)). i )) r Directly proportional.

[0040] Specifically, the coefficients differ for different categories; that is, the more difficult the target is to train, the higher the coefficient (1-f(x)). i )) r The larger the value of , the larger the coefficient, and the larger the overall loss function, meaning the more the model focuses on that category. Based on experience in this field, we set 'a' to 0.25 and 'r' to 2. We adjust 'a' and 'r' by adjusting the size of the loss function for each training iteration, and then use linear regression to train the classifier to obtain the optimal solution for the loss.

[0041] The solution is obtained using the linear regression least squares method.

[0042] w = (X H X+λ2I) -1 X H y

[0043] X is a cyclic matrix, where each row represents a feature vector of a sample; y is a label column vector, where each element represents a label; I is a unit cyclic matrix, generating a vector of [1,0,0,……0]; H represents the conjugate transpose (the matrix is ​​symmetrically swapped element-wise); and λ² is the coefficient.

[0044] Diagonalizing the circulant matrix using Fourier transform properties speeds up computation:

[0045]

[0046] For the Fourier transform of the original vector x, for The conjugate matrix, For X H The eigenvalues ​​of X, δ is the vector of all elements 1 generated by the Fourier transform of I, and F is the Fourier transform matrix.

[0047] Based on the property of inverting circulant matrices, matrix inversion can be transformed into eigenvalue inversion:

[0048]

[0049] Elimination using the unitary matrix property of F:

[0050]

[0051] The equations can be further simplified by inverting the diagonalization property:

[0052]

[0053] C is an anti-diagonalization operation.

[0054] Utilizing the properties of circular convolution matrices:

[0055]

[0056] because Since each element is a real number, the conjugate matrix remains unchanged, and the formula can be further simplified to:

[0057]

[0058] Finally, w can be calculated using the Fourier transform of the vector and the positional multiplication.

[0059] Step S3: Establish a sample area scaling model. Based on the target object sample of the current frame tracking image, the flight altitude value corresponding to the current frame tracking image, and the sample area scaling model, obtain the scaling change ratio of the target object acquisition area of ​​the current frame relative to the previous frame tracking image, and obtain the true values ​​of the target object acquisition area and the target object sample of the current frame tracking image.

[0060] The sample area scaling model is as follows:

[0061] f(x)=ωx 2 +ψx+ζ

[0062] ψ and ω are fitting coefficients, ζ is a fitting constant, f(x) is the scaling ratio of the target acquisition area in the current frame relative to the previous frame of the tracked image, and x is the flight altitude value.

[0063] Specifically, the scaling factor of the target acquisition area is approximately positively correlated with the target and flight altitude values. Thus, the change in the scaling factor of the target acquisition area is reflected by the change in the flight altitude value, thereby obtaining the size of the target acquisition area after the change.

[0064] Step S4: Establish a Kalman filter model. When the current frame tracking image provides feedback to the predetermined target, the Kalman filter model is iterated and the error is corrected using the Kalman filter model, the true value of the target sample in the current frame tracking image, and the predicted value. When the number of acquired tracking images is greater than a predetermined number, and the current frame tracking image provides no feedback to the predetermined target, the target sample in the current frame tracking image is filled with the corresponding predicted value data based on the current Kalman filter model.

[0065] Step S5: Repeat steps S3 and S4 until the tracking camera stops tracking the predetermined target.

[0066] The iterative process of the Kalman filter model is as follows:

[0067] Process T1: Use the predicted value of the target object sample by the Kalman filter model as the sample predicted value, and use the covariance matrix of the actual value and the sample predicted value of the target object sample by the Kalman filter model as the sample covariance matrix.

[0068] Process T2: Initialize the sample predicted values ​​and covariance matrix;

[0069] Process T3: Update and iterate the initialized sample prediction values ​​and covariance matrix, and update and iterate the Kalman gain matrix;

[0070] Process T4: Repeat processes T2 and T3 until the tracking camera stops tracking the predetermined target.

[0071] Specifically, when the predetermined target object receives feedback, that is, when the predetermined target object is not occluded and appears in the current frame tracking image, the system state difference equation and the observation equation are combined: the predicted value of the target object sample in the previous frame tracking image is compared with the true value of the target object sample in the current frame tracking image, and then the model is iterated to correct the prediction error; when the predetermined target object is occluded, the corresponding predicted value data is filled in through the system state difference equation.

[0072] The system state difference equation is:

[0073] x k =Ax k-1 +Bu k +w k

[0074] Where, x k Let u be the state at time k, let A be the n*1 state vector, let A be the transition state matrix, and let u be the state vector. k Let w be the system control vector at time k, denoted by , and B be the system control matrix. k Let be the system noise from time k-1 to time k.

[0075] The observation equation is:

[0076] z k =H k x k +v k

[0077] Where: z k H is the measurement value at time k. kLet v be the observation transformation matrix at time k. k This refers to the noise present in the observed state.

[0078] Prediction phase:

[0079]

[0080] in: The predicted value at time k;

[0081] We now have a biased prediction. and measured value z k By combining the two values, we obtain the final estimate:

[0082]

[0083] in: Let K be the estimated value at time k. k This is the Kalman gain matrix, a continuously updated matrix. Next, we will work on K... k Solution:

[0084] According to the definition of variance, the covariance matrix between the predicted and actual values ​​is:

[0085] p′ k =Ap k-1 A T +Q

[0086] Where: p' k Let p be the covariance matrix of the predicted and actual values. k-1 Let be the covariance matrix between the estimated and true values, and Q be the system noise covariance matrix under actual operating conditions.

[0087] From the above formula, we can obtain:

[0088] K k =p′ k H T (Hp′ k H T +R) -1

[0089] Where R is the system noise covariance matrix during the observation process.

[0090] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications or variations that can be made by those skilled in the art without creative effort within the scope of the appended claims are still within the scope of protection of this patent.

Claims

1. A method for establishing a target tracking model for an aircraft, wherein the aircraft is equipped with a tracking camera and an altitude sensor, the tracking camera being used to track and capture images of a predetermined target object, thereby forming multiple frames of tracking images, and the altitude sensor being used to collect the flight altitude value of the aircraft relative to a predetermined reference horizon in real time, characterized in that, Includes the following steps: Step S1: Based on the size of the pixel group occupied by the predetermined target object as fed back in the first frame tracking image, a target object acquisition area is established, and the pixel group occupied by the predetermined target object is centrally contained within the pixel group occupied by the target object acquisition area. Step S2: Establish a sample collection loop matrix based on the ratio of the number of pixels occupied by the target collection area and the predetermined target object, and collect positive and negative samples in the target object collection area through the sample collection loop matrix to obtain the target object sample corresponding to the predetermined target object; Step S3: Establish a sample area scaling ratio model. Based on the target object sample of the current frame tracking image, the flight altitude value corresponding to the current frame tracking image, and the sample area scaling ratio model, obtain the scaling change ratio of the target object acquisition area of ​​the current frame relative to the previous frame tracking image, and obtain the true values ​​of the target object acquisition area and the target object sample of the current frame tracking image. Step S4: Establish a Kalman filter model. When the current frame tracking image provides feedback to the predetermined target, the Kalman filter model is iterated and the error is corrected using the Kalman filter model, the true value of the target sample in the current frame tracking image, and the predicted value. When the number of acquired tracking images is greater than a predetermined number, and the current frame tracking image provides no feedback to the predetermined target, the target sample in the current frame tracking image is filled with corresponding predicted value data based on the current Kalman filter model. Step S5: Repeat steps S3 and S4 until the tracking camera stops tracking the predetermined target.

2. The method for establishing a target tracking model for an aircraft according to claim 1, characterized in that: in, In step S2, the positive and negative sample sampling process is as follows: In the tracking image, positive and negative samples are sampled from the pixel group occupied by the target object acquisition area through the sample acquisition cycle matrix. Multiple classifiers are trained based on the obtained multiple sample sets. The multiple response values ​​obtained by the multiple classifiers for the multiple sample sets are compared, and the sample set corresponding to the highest response value is taken as the target object sample.

3. The method for establishing a target tracking model for an aircraft according to claim 1, characterized in that: in, In step S2, the loss function for the target sample is set as follows: L(y i ,f(x i Let λ1 be the loss function, λ1 be the preset regularization factor, f(x) be the linear classifier, f(x) = w·x + b, w be the weight matrix of the function to be fitted, b be the bias vector, and (x1,y1), (x2,y2), ..., (x... m ,y m Let be the sample label, r be the scaling numerator, x be the sampled value of the cyclic matrix, and a be the class factor, where a is constant relative to (1-f(x)). i )) r Directly proportional.

4. The method for establishing a target tracking model for an aircraft according to claim 1, characterized in that: in, The sample area scaling model is as follows: f(x)=ωx 2 +ψx+ζ ψ and ω are fitting coefficients, ζ is a fitting constant, f(x) is the scaling ratio of the target acquisition area of ​​the current frame relative to the previous frame tracking image, and x is the flight altitude value.

5. The method for establishing a target tracking model for an aircraft according to claim 1, Its features are: In step S4, the iterative process of the Kalman filter model is as follows: Process T1: The predicted value of the target object sample by the Kalman filter model is used as the sample predicted value, and the covariance matrix between the actual value of the target object sample and the predicted value of the sample by the Kalman filter model is used as the sample covariance matrix. Process T2: Initialize the predicted values ​​of the samples and the covariance matrix; Process T3: Update and iterate the initialized sample prediction values ​​and the covariance matrix, and update and iterate the Kalman gain matrix; Process T4: Repeat processes T2 and T3 until the tracking camera stops tracking the predetermined target.

Images