Remote sensing image target tracking method simulating fly compound eye visual system

Abstract

The application discloses a remote sensing image target tracking method simulating fly compound eye visual system. First, the characteristics of ommatidium in fly compound eye are simulated, a luminance receiver is constructed, and image luminance information is extracted; then, the characteristics of thin plate single cell LMC and the side inhibition mechanism between LMCs are simulated, image luminance change is detected, and the contrast between a target and a background is strengthened; third, the characteristics of ON-OFF channel in medulla layer are simulated, and the luminance increase information and the luminance decrease information are processed respectively; fourth, the characteristics of STMD neuron and the side inhibition mechanism between STMD neurons are simulated, and motion information detection is performed; finally, on the basis of completing motion target detection on all frames, a minimum circumscribed rectangle method and a neighborhood threshold judgment method are designed to further analyze and judge the target trajectory, and finally, the target tracking result is obtained. The application simulates the fly compound eye visual system, can effectively track the motion target in the remote sensing image, and has good accuracy and robustness.

Description

Technical Field

[0001] This invention belongs to the field of image processing, and in particular relates to a method for tracking targets in remote sensing images that simulates the compound eye visual system of flies. Background Technology

[0002] Remote sensing image target tracking is an important research direction in the field of computer vision. It refers to the process of continuously tracking a specific target in subsequent frames after identifying it in the first frame of a remote sensing video sequence. It uses bounding boxes (usually rectangles) to label the target and simultaneously achieves target localization and scale estimation. Remote sensing image target tracking methods can be applied to many fields, such as tracking ships at sea in marine monitoring and monitoring and early warning of military targets in military strategic guidance. Existing remote sensing image target tracking methods mainly follow those used for target tracking on ordinary images and can be categorized into motion prediction-based target tracking methods, generative target tracking methods, discriminative target tracking methods, correlation filtering-based target tracking methods, deep learning-based target tracking methods, and biomimetic vision-based target tracking methods.

[0003] Motion prediction-based target tracking methods were among the earliest proposed approaches and are currently primarily used in conjunction with other target tracking methods to achieve more accurate results. Generative target tracking methods focus only on the features of the target itself; when the target features are close to the background or the target is occluded, the tracking bounding box is prone to drift. Discriminative target tracking methods typically require a large number of training samples, significantly impacting tracking efficiency and making real-time tracking difficult. Correlation filtering-based methods also require a large number of samples, leading to high computational costs and impacting tracking efficiency. Deep learning-based methods can achieve better tracking accuracy; however, most deep learning-based target tracking methods have high hardware requirements and require significant time for feature learning and model training.

[0004] Most organisms in nature possess sophisticated visual nervous systems. These systems are highly adaptable and stable, enabling them to quickly and accurately analyze and understand complex visual environments. Their exceptional ability to detect moving targets, in particular, aids them in predation, evading predators, and courtship. Current research on biomimetic vision primarily includes: mimicking human vision, mammalian vision (e.g., cats, macaques), and insect vision (e.g., locusts, dragonflies, flies). Driven by advancements in research on biological vision mechanisms, target tracking in remote sensing images, driven by biomimetic visual mechanisms, has attracted significant attention.

[0005] Authorization number CN116883458B describes a multi-target tracking system that uses a Transformer model and integrates observation-centric motion features. First, it calculates the similarity between the target trajectory and detected targets to obtain the detected targets and their confidence scores in the current frame. Then, it uses historical detection information to predict the target position in the next frame. Next, using the detection boxes of the established trajectory, it calculates the velocity and direction information of the target trajectory to obtain a virtual trajectory. Finally, it corrects the trajectory through camera motion compensation. This method has high algorithmic complexity due to the use of the Transformer deep learning model.

[0006] Authorization number CN112233252B describes an AR target tracking method and system based on feature matching and optical flow fusion. First, an augmented reality (AR) rendering target image is selected as the rendering template image. Second, feature detection is performed on the acquired video frame sequence. Next, feature matching is performed between each frame image and the template image to determine if tracking is successful. If tracking is lost, the next frame image is input for feature detection and matching; if tracking is successful, the tracking mode is maintained. Finally, real-time rendering is performed based on the estimated camera pose. This method, due to its use of optical flow and the need for real-time rendering, has relatively low efficiency.

[0007] In summary, the limitations of existing remote sensing image target tracking algorithms are mainly reflected in the following aspects:

[0008] (1) Traditional target tracking methods only use the characteristics of the target itself as the basis for judgment when tracking the target, while ignoring the background information. When the target features are similar to the background, the tracking box will drift, which will lead to a decrease in tracking performance.

[0009] (2) Deep learning-based target tracking methods require a lot of time to train deep learning models, resulting in long processing times and complex computation processes, making it difficult to meet real-time requirements. Summary of the Invention

[0010] Objective: To address the problems existing in the prior art, this invention provides a remote sensing image target tracking method that simulates the compound eye visual system of flies. This method overcomes the problems of complex model structures, high hardware requirements, and long training times inherent in deep learning-based methods. It also shortens the target tracking time while maintaining high tracking accuracy and success rate. By simulating the characteristics of neurons related to visual motion information processing and the lateral inhibition mechanism in the fly visual system, it can effectively extract moving targets from remote sensing image sequences while suppressing background clutter, thus improving the accuracy, robustness, and real-time performance of target tracking.

[0011] Technical Solution: To achieve the objective of this invention, the technical solution adopted is a remote sensing image target tracking method simulating the compound eye visual system of flies, the specific steps of which are as follows:

[0012] (1) In the retinal layer, a brightness receiver was constructed to extract image brightness information by simulating the microphthalmia of the compound eye of flies;

[0013] (2) In the thin-plate layer, the characteristics of thin-plate unipolar cells (LMCs) and the lateral inhibition mechanism between LMCs are simulated to detect changes in image brightness and enhance the contrast between the target and the background.

[0014] (3) In the medulla, the ON-OFF channel characteristics of the medulla are simulated so that the brightness increase information and the brightness decrease information are processed separately;

[0015] (4) In the lobular layer, the characteristics of STMDs neurons and the lateral inhibition mechanism between STMDs neurons are simulated to detect motor information.

[0016] (5) Based on the completion of motion target detection in all frames, the minimum bounding rectangle method and the neighborhood threshold discrimination method are designed to further analyze and judge the target trajectory and obtain the final target tracking result.

[0017] In step (1), the method for constructing a brightness receiver to extract image brightness information by simulating the microphthalmia characteristic of fly compound eyes is as follows:

[0018] Let the output of the small eye at pixel (x, y) be R(x, y), then:

[0019]

[0020] Where I(x, y) represents the brightness value at pixel (x, y) in the current image frame, and * represents the convolution operation. The Gaussian function is expressed as:

[0021]

[0022] Where σ1 represents the standard deviation of the Gaussian function.

[0023] In step (2), the method for simulating the characteristics of thin-plate unipolar cells (LMCs) and the lateral inhibition mechanism between LMCs, detecting changes in image brightness, and enhancing the contrast between the target and the background is as follows:

[0024] (2.1) Design a bandpass filter to simulate the characteristics of LMCs neurons. Let the output of the LMCs neuron be L. MC (x, y, t), then:

[0025] LMC (x,y,t)=R(x,y,t)*B(t)

[0026] Where R(x, y, t) represents the output of the small eye at time t in step (1.1) at pixel (x, y), * represents the convolution operation, and B(t) represents the bandpass filter. B(t) is implemented by the difference between two gamma kernel functions, as shown in the following formula:

[0027]

[0028] in, and The gamma kernel function is defined as follows:

[0029]

[0030] Where, n i τ represents the order of the gamma kernel function. i This represents the time constant in the gamma kernel function;

[0031] (2.2) Design a lateral inhibition function to simulate the lateral inhibition mechanism between LMCs neurons. Lateral inhibition can increase the contrast between the target and the background and improve anti-interference performance, which is beneficial for target tracking. Let the output of the LMCs neuron after lateral inhibition be L. MCI (x, y, t), then:

[0032] L MCI (x, y, t) = L MC (x, y, t)*W LMC (x, y, t)

[0033] Among them, L MC (x, y, t) represents the output of the LMCs neurons in step (2.1), * represents the convolution operation, W LMC (x, y) represents the lateral suppression kernel function. LMC (x, y) is defined as follows:

[0034] W LMC (x, y) = W ss (x, y) + W sh (x, y)

[0035] Among them, W ss (x, y) represents the excitation part of the upper-side inhibition mechanism in the spatial domain, W sh (x, y) represents the spatial domain inhibition portion of the lateral inhibition mechanism, W ss (x, y) and W sh (x, y) are all defined by a LoG distribution function, which is:

[0036]

[0037] Where σ² represents the standard deviation in the LoG distribution function. Based on the LoG distribution function, W is defined. ss (x, y) and W sh (x, y) are as follows:

[0038]

[0039]

[0040] Where, [x] + This represents max(x, 0), [x] - This means min(x, 0).

[0041] In step (3), the method for simulating the ON-OFF channel characteristics of the medullary layer to process the brightness increase information and brightness decrease information separately is as follows:

[0042] (3.1) Using a half-wave rectification method to mimic the ON-OFF channel characteristics of flies, visual information is divided into brightness increase signals and brightness decrease signals, so that subsequent processing can be performed on the two parallel channels, ON and OFF, respectively. Let M... ON (x, y, t) represents the brightness increase signal on the ON channel, M OFF (x, y, t) represents the brightness reduction signal on the OFF channel. The above process can be represented as:

[0043] M ON (x, y, t) = [L] MCI (x, y, t)] +

[0044] M OFF (x, y, t) = -[L MCI (x, y, t)] -

[0045] Among them, L MCI (x, y, t) represents the output of the LMCs neurons after lateral inhibition in step (2.2), [x] + This represents max(x, 0), [x] - This means min(x, 0);

[0046] (3.2) Simulate four motion-sensitive medullary neurons Tm1, Tm2, ​​Tm3, and Mi1. Since Tm2 and Tm3 respond immediately to increases and decreases in the brightness signal, the brightness signals on Tm2 and Tm3 are equated to the brightness signals on the OFF and ON channels, respectively. Let M...Tm2 (x, y, t) represents the output of Tm2 medullary neurons, M Tm3 (x, y, t) represents the output of the Tm3 medullary neuron, then:

[0047] M Tm2 (x, y, t) = M OFF (x, y, t)

[0048] M Tm3 (x, y, t) = M ON (x, y, t)

[0049] Among them, M OFF (x, y, t) represents the brightness reduction signal on the OFF channel in step (3.1), M ON (x, y, t) represents the brightness increase signal on the ON channel in step (3.1);

[0050] Furthermore, since the responses of Tm1 and Mi1 medullary neurons exhibit a certain delay, a design was devised to convolve them with a gamma kernel function to achieve a temporal delay. Let M... Tm1 (x, y, t) and M Mi1 (x, y, t) represent the outputs of Tm1 medullary neurons and Mi1 medullary neurons, respectively. Then:

[0051]

[0052]

[0053] in, and The gamma kernel function is defined as follows:

[0054]

[0055] Where, n i τ represents the order of the gamma kernel function. i This represents the time constant in the gamma kernel function.

[0056] In step (4), the method for detecting motion information by simulating the characteristics of STMD neurons and the lateral inhibition mechanism between STMD neurons is as follows:

[0057] (4.1) Let S TMD (x, y, t) represents the output of the STMDs neuron, then:

[0058] S TMD (x, y, t) = M Tm1 (x, y, t) × M Tm3 (x, y, t)

[0059] Among them, M Tm1 (x, y, t), M Tm3 (x, y, t) represents the output of the Tm1 and Tm3 medullary neurons in step (3.2);

[0060] (4.2) Since there is a lateral inhibition mechanism in STMDs neurons, simulating this mechanism can suppress the response to large targets in the background, thereby reducing the false detection rate. Based on this, the lateral inhibition function is implemented by convolving the output of STMDs neurons with the lateral inhibition function. First, the lateral inhibition function W in STMDs neurons is defined. STMD (x, y) is:

[0061]

[0062] Where k and m are constants, [x] + and [x] - Let max(x, 0) and min(x, 0) represent these values ​​respectively. Then, a LoG distribution function is defined as follows:

[0063]

[0064] Where σ3 represents the LoG distribution function The standard deviation in;

[0065] Then use S TMDI Let (x, y, t) represent the output of STMDs neurons after lateral inhibition, then:

[0066] S TMDI (x, y, t) = S TMD (x, y, t)*W STMD (x, y)

[0067] Finally, S is compared using a threshold judgment method. TMDI (x, y, t) and detection threshold β STMD The magnitude of S determines whether the target detected at (x, y) at time t is a real target. TMDI (x, y, t) > β STMD If the target is positive, it is considered a real moving target; otherwise, it is considered a non-moving target.

[0068] In step (5), after detecting all moving targets in all frames, the minimum bounding rectangle method and the neighborhood threshold discrimination method are designed to further analyze and judge the target trajectory, and the final target tracking result is obtained as follows:

[0069] (5.1) Based on the results of moving target detection in each frame obtained in step (4), the minimum bounding rectangle method is used to obtain the coordinates [x] of the predicted bounding box in that frame. p y p w p h p ], where (x p y p ) represents the coordinates of the top-left corner of the prediction box, w p and h p Let x and y represent the width and height of the predicted bounding box, respectively; then, use the center coordinates (x, y) of the predicted bounding box. c y c The coordinates of the predicted target are used as the coordinates of the target.

[0070]

[0071]

[0072] (5.2) Design a neighborhood discrimination method to determine whether a moving target detected in two consecutive frames is on the same trajectory. Calculate the Euclidean distance between the position coordinates of the two frames as the distance between the two predicted moving targets. If the distance between the predicted moving target's position in the current frame and the predicted moving target's position in the next frame is less than the neighborhood threshold β... n If the predicted moving target in a given frame differs from the predicted moving target in the two frames in the next frame by more than a neighborhood threshold β, then the two targets are considered to be on the same trajectory, and the target tracking is successful. Otherwise, if the position of a detected predicted moving target in a given frame differs from the positions of the detected predicted moving targets in the two frames before and after the next frame by more than a neighborhood threshold β, the target tracking is considered successful. n If the target is not tracked, the tracking of the moving object in that frame is considered to have failed.

[0073] Beneficial effects: The present invention, by adopting the above technical solution, has the following beneficial effects:

[0074] (1) A remote sensing image target tracking method simulating the compound eye vision system of flies is designed. It can track targets not only in simple backgrounds but also in complex motion backgrounds. In addition, this method can track not only pedestrian targets in remote sensing images but also other targets such as vehicles and ships, thus having good universality.

[0075] (2) Compared with methods such as deep learning, the method of the present invention has low hardware configuration requirements, is easy to implement, and can quickly and accurately capture key visual information, ensuring efficient and reliable target tracking.

[0076] (3) The side inhibition function of the fly-like visual system side inhibition mechanism can suppress background interference in remote sensing images, improve the contrast between the target and the background, and thus further improve the accuracy and success rate of target tracking. Attached Figure Description

[0077] Figure 1 This is a framework diagram of the method of the present invention.

[0078] Figure 2 This is a comparison chart of the tracking performance of the method of the present invention with other methods. Detailed Implementation

[0079] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0080] like Figure 1 As shown, the technical solution of the present invention is further described in detail below:

[0081] (1) In the retinal layer, a brightness receiver was constructed to extract image brightness information by simulating the microphthalmia of the compound eye of flies.

[0082] Let the output of the small eye located at pixel (x, y) be R(x, y), then:

[0083]

[0084] Where I(x, y) represents the brightness value at pixel (x, y) in the current image frame, and * represents the convolution operation. Then represents the Gaussian function, and its expression is:

[0085]

[0086] Where σ1 represents the standard deviation of the Gaussian function, and its value is σ1 = 1.

[0087] (2) In the thin-plate layer, the characteristics of thin-plate unipolar cells (LMCs) and the lateral inhibition mechanism between LMCs are simulated to detect changes in image brightness and enhance the contrast between the target and the background.

[0088] (2.1) Design a bandpass filter to simulate the characteristics of LMCs neurons. Let the output of the LMCs neuron be L. MC (x, y, t), then:

[0089] L MC (x,y,t)=R(x,y,t)*B(t)

[0090] Where R(x, y, t) represents the output of the small eye at time t in step (1.1) at pixel (x, y), * represents the convolution operation, and B(t) represents the bandpass filter. B(t) is implemented by the difference between two gamma kernel functions, as shown in the following formula:

[0091]

[0092] in, and Both refer to the gamma kernel function, which is defined as follows:

[0093]

[0094] Where, n i τ represents the order of the gamma kernel function. i This represents the time constant in the gamma kernel function. The value of i is i = 1, 2, n. i and τ i The values ​​of are: n1 = 2, τ1 = 3, n2 = 6, τ2 = 9;

[0095] (2.2) Design a lateral inhibition function to simulate the lateral inhibition mechanism between LMCs neurons. Lateral inhibition can increase the contrast between the target and the background and improve anti-interference performance, which is beneficial for target tracking. Let the output of the LMCs neuron after lateral inhibition be L. MCI (x, y, t), then:

[0096] L MCI (x, y, t) = L MC (x, y, t)*W LMC (x, y, t)

[0097] Among them, L MC (x, y, t) represents the output of the LMCs neurons in step (2.1), * represents the convolution operation, W LMC (x, y) represents the lateral suppression kernel function. LMC (x, y) is defined as follows:

[0098] W LMC (x, y) = W ss (x, y) + W sh (x, y)

[0099] Among them, W ss (x, y) represents the excitation part of the upper-side inhibition mechanism in the spatial domain, W sh (x, y) represents the spatial domain inhibition portion of the lateral inhibition mechanism. W ss (x, y) and W sh (x, y) are all defined by a LoG distribution function, which is:

[0100]

[0101] Where σ² represents the standard deviation in the LoG distribution function, and its value is σ² = 2. Based on the LoG distribution function, W... ss (x, y) and W sh (x, y) is defined as follows:

[0102]

[0103]

[0104] Where, [x] + This represents max(x, 0), [x] - This means min(x, 0).

[0105] (3) In the medulla, the ON-OFF channel characteristics of the medulla are simulated so that the brightness increase information and the brightness decrease information are processed separately.

[0106] (3.1) Using a half-wave rectification method to mimic the ON-OFF channel characteristics of flies, the visual information is divided into a brightness increase signal and a brightness decrease signal, which facilitates subsequent processing on the two parallel channels, the ON channel and the OFF channel, respectively. Let M ON (x, y, t) represents the brightness increase signal on the ON channel, M OFF (x, y, t) represents the brightness reduction signal on the OFF channel. The above process can be represented as:

[0107] M ON (x, y, t) = [L] MCI (x, y, t)] +

[0108] M OFF (x, y, t) = -[L MCI (x, y, t)] -

[0109] Among them, L MCI (x, y, t) represents the output of the LMCs neurons after lateral inhibition in step (2.2), [x] + This represents max(x, 0), [x] - This means min(x, 0);

[0110] (3.2) Simulate four motion-sensitive medullary neurons Tm1, Tm2, ​​Tm3, and Mi1. Since Tm2 and Tm3 respond immediately to increases and decreases in the brightness signal, the brightness signals on Tm2 and Tm3 are equated to the brightness signals on the OFF and ON channels, respectively. Let M... Tm2 (x, y, t) represents the output of Tm2 medullary neurons, M Tm3 (x, y, t) represents the output of the Tm3 medullary neuron, then:

[0111] M Tm 2(x, y, t) = M OFF (x, y, t)

[0112] M Tm3 (x, y, t) = M ON (x, y, t)

[0113] Among them, M OFF (x, y, t) represents the brightness reduction signal on the OFF channel in step (3.1), M ON (x, y, t) represents the brightness increase signal on the ON channel in step (3.1);

[0114] Furthermore, since the responses of Tm1 and Mi1 medullary neurons exhibit a certain delay, a design was devised to convolve them with a gamma kernel function to achieve a temporal delay. Let M... Tm1 (x, y, t) and M Mi1 (x, y, t) represent the outputs of Tm1 medullary neurons and Mi1 medullary neurons, respectively. Then:

[0115]

[0116]

[0117] in, and The gamma kernel function is defined as follows:

[0118]

[0119] Where, n i τ represents the order of the gamma kernel function. i This represents the time constant in the gamma kernel function. The value of i is i = 3, 4, n. i and τ i The values ​​of τ are: n3 = 5, τ3 = 25, n4 = 3, τ4 = 15.

[0120] (4) In the lobular layer, the characteristics of STMDs neurons and the lateral inhibition mechanism between STMDs neurons are simulated to detect motor information.

[0121] (4.1) Let S TMD (x, y, t) represents the output of the STMDs neuron, then:

[0122] S TMD (x, y, t) = M Tm1 (x, y, t) × M Tm3 (x, y, t)

[0123] Among them, M Tm1 (x, y, t), M Tm3 (x, y, t) represents the output of the Tm1 and Tm3 medullary neurons in step (3.2);

[0124] (4.2) Since there is a lateral inhibition mechanism in STMDs neurons, simulating this mechanism can suppress the response to large targets in the background, thereby reducing the false detection rate. Based on this, the lateral inhibition function is implemented by convolving the output of STMDs neurons with the lateral inhibition function. First, the lateral inhibition function W in STMDs neurons is defined. STMD (x, y) is:

[0125]

[0126] Where k and m are constants, taking values ​​of k = 1 and m = 3, [x] + and [x] - Let max(x, 0) and min(x, 0) represent these values ​​respectively. Then, a LoG distribution function is defined as follows:

[0127]

[0128] Where σ3 represents the LoG distribution function The standard deviation in the range is σ³ = 2.

[0129] Then use S TMDI Let (x, y, t) represent the output of STMDs neurons after lateral inhibition, then:

[0130] S TMDI (x, y, t) = S TMD (x, y, t)*W STMD (x, y)

[0131] Among them, S TMD (x, y, t) is the output of the STMDs neurons in step (4.1);

[0132] Finally, S is compared using a threshold judgment method. TMDI (x, y, t) and detection threshold βSTMD The magnitude of S determines whether the target detected at (x, y) at time t is a real target. TMDI (x, y, t) > β STMD If the target is not detected, it is considered a real moving target; otherwise, it is considered a non-moving target.

[0133] (5) Based on the completion of motion target detection in all frames, the minimum bounding rectangle method and the neighborhood threshold discrimination method are designed to further analyze and judge the target trajectory and obtain the final target tracking result.

[0134] (5.1) Based on the results of moving target detection in each frame obtained in step (4), the minimum bounding rectangle method is used to obtain the coordinates [x] of the predicted bounding box in that frame. p y p w p h p ], where (x p y p ) represents the coordinates of the top-left corner of the prediction box, w p and h p Let x and y represent the width and height of the predicted bounding box, respectively; then, use the center coordinates (x, y) of the predicted bounding box. c y c ) are used as the location coordinates of the predicted target, where,

[0135]

[0136]

[0137] (5.2) Design a neighborhood discrimination method to determine whether a moving target detected in two consecutive frames is on the same trajectory. Calculate the Euclidean distance between the position coordinates of the two frames as the distance between the two predicted moving targets. If the distance between the predicted moving target's position in the current frame and the predicted moving target's position in the next frame is less than the neighborhood threshold β... n If the predicted moving target in a given frame differs from the predicted moving target in the two frames in the next frame by more than a neighborhood threshold β, then the two targets are considered to be on the same trajectory, and the target tracking is successful. Otherwise, if the position of a detected predicted moving target in a given frame differs from the positions of the detected predicted moving targets in the two frames before and after the next frame by more than a neighborhood threshold β, the target tracking is considered successful. n If the target is not tracked, the tracking of the moving object in that frame is considered to have failed.

[0138] This invention compares two different target tracking methods with the proposed method. The comparison method used is:

[0139] The Elementary Small Target Motion Detector (ESTMD) model, proposed by Wiederman et al. in “A Model for the Detection of Moving Targets in VisualClutter Inspired by Insect Physiology[J]. PloS One, 2008.”, is abbreviated as ESTMD.

[0140] Wang et al. proposed a Directionally Selective Small Target Motion Detecting Visual Neural Network in Cluttered Backgrounds[J].IEEE Transactions on Cybernetics, 2020., which is abbreviated as DSTMD.

[0141] Figure 2 This paper compares the target tracking accuracy and success rate curves of three methods on the publicly available UAV-123 remote sensing target tracking dataset. These two curves are commonly used metrics for target tracking performance, and the larger the area under these curves (as shown by the numbers in the lower right corner of the figure), the better the performance. Therefore, it is evident that the method proposed in this invention achieves the best target tracking effect on remote sensing images.

Claims

1. A remote sensing image target tracking method simulating the compound eye visual system of flies, characterized in that, The method includes the following steps: (1) In the retinal layer, a brightness receiver was constructed to extract image brightness information by simulating the microphthalmia of the compound eye of flies. (2) In the thin-plate layer, the characteristics of thin-plate unipolar cells (LMCs) and the lateral inhibition mechanism between LMCs are simulated to detect changes in image brightness and enhance the contrast between the target and the background. (3) In the medulla, the ON-OFF channel characteristics of the medulla are simulated, and the brightness increase information and brightness decrease information are processed separately; (4) In the lobular layer, the characteristics of STMDs neurons and the lateral inhibition mechanism between STMDs neurons are simulated to detect motor information. (5) Based on the completion of motion target detection in all frames, the minimum bounding rectangle method and the neighborhood threshold discrimination method are designed to further analyze and judge the target trajectory, and the final target tracking result is obtained. In step (2), the following method is used to simulate the characteristics of thin-plate unipolar cells (LMCs) and the lateral inhibition mechanism between LMCs, detect changes in image brightness, and enhance the contrast between the target and the background: (2.1) Design a bandpass filter to simulate the characteristics of LMCs neurons. Let the output of the LMCs neuron be L. MC (x, y, t), then: L MC (x,y,t):R(x,y,t)*B(t) Where R(x, y, t) represents the output of the small eye at time t at pixel (x, y), * represents the convolution operation, and B(t) represents the bandpass filter; B(t) is implemented by the difference of two gamma kernel functions, as shown in the following formula: in, and The gamma kernel function is defined as follows: Where, n i τ represents the order of the gamma kernel function. i This represents the time constant in the gamma kernel function; (2.2) Design a lateral inhibition function to simulate the lateral inhibition mechanism between LMCs neurons; the lateral inhibition mechanism can increase the contrast between the target and the background and improve the anti-interference performance, which is beneficial to target tracking; let the output of the LMCs neuron after lateral inhibition be L MCI (x, y, t), then: L MCI (x,y,t)=L MC (x,y,t)*W LMC (x,y,t) Among them, L MC (x, y, t) represents the output of the LMCs neurons in step (2.1), * represents the convolution operation, W LMC (x, y) represents the lateral suppression kernel function; W LMC (x, y) is defined as follows: W LMC (x,y)=W ss (x,y)+W sh (x,y) Among them, W ss (x, y) represents the excitation part of the upper-side inhibition mechanism in the spatial domain, W sh (x, y) represents the spatial domain inhibition portion of the lateral inhibition mechanism, W ss (x, y) and W sh (x, y) are all defined by a LoG distribution function, which is: Where σ² represents the standard deviation in the LoG distribution function; based on the LoG distribution function, W is defined. ss (x, y) and W sh (x, y) are as follows: Where, [x] + This represents max(x, 0), [x] - This means min(x, 0).

2. The remote sensing image target tracking method according to claim 1, simulating the microphthalmia characteristic of fly compound eyes, constructing a brightness receiver, and extracting image brightness information in step (1) is as follows: (1.1) Let the output of the small eye at pixel (x, y) be R(x, y), then: in, I(x, y) represents the brightness value at pixel (x, y) in the current image frame, and * indicates a convolution operation. The Gaussian function is expressed as: Where σ1 represents the standard deviation of the Gaussian function.

3. The remote sensing image target tracking method for simulating the compound eye visual system of flies according to claim 1, in step (3), the method for simulating the ON-OFF channel characteristics of the medulla layer to process the brightness increase information and brightness decrease information separately is as follows: (3.1) Using a half-wave rectification method to mimic the ON-OFF channel characteristics of flies, the visual information is divided into a brightness increase signal and a brightness decrease signal, so as to facilitate subsequent processing on the two parallel channels, the ON channel and the OFF channel, respectively; let M ON (x, y, t) represents the brightness increase signal on the ON channel, M OFF (x, y, t) represents the brightness reduction signal on the OFF channel. The above process can be represented as: M ON (x,y,t)=[L MCI (x,y,t)] + M OFF (x,y,t)=-[L MCI (x,y,t)] - in, L MCI (x, y, t) represents the output of the LMCs neurons after lateral inhibition in step (2.2), [x] + This represents max(x, 0), [x] - This means min(x, 0); (3.2) Simulate four motion-sensitive medullary neurons Tm1, Tm2, ​​Tm3, and Mi1; among them, since Tm2 and Tm3 respond immediately to increases and decreases in the brightness signal, the brightness signals on Tm2 and Tm3 are equated to the brightness signals on the OFF and ON channels, respectively; let M Tm2 (x, y, t) represents the output of Tm2 medullary neurons, M Tm3 (x, y, t) represents the output of the Tm3 medullary neuron, then: M Tm2 (x,y,t)=M OFF (x,y,t) M Tm3 (x,y,t)=M ON (x,y,t) Among them, M OFF (x, y, t) represents the brightness reduction signal on the OFF channel in step (3.1), M ON (x, y, t) represents the brightness increase signal on the ON channel in step (3.1); Furthermore, since the responses of Tm1 and Mi1 medullary neurons exhibit a certain delay, a design was devised to convolve them with a gamma kernel function to achieve a temporal delay. Let M... Tm1 (x, y, t) and M Mi1 (x, y, t) represent the outputs of Tm1 medullary neurons and Mi1 medullary neurons, respectively. Then: in, and The gamma kernel function is defined as follows: Where, n i τ represents the order of the gamma kernel function. i This represents the time constant in the gamma kernel function.

4. The remote sensing image target tracking method for simulating the compound eye visual system of flies according to claim 3, in step (4), the method for detecting motion information by simulating the characteristics of STMDs neurons and the lateral inhibition mechanism between STMDs neurons is as follows: (4.1) Let S TMD (x, y, t) represents the output of the STMDs neuron, then: S TMD (x,y,t)=M Tm1 (x,y,t)×M Tm3 (x,y,t) in, M Tm1 (x, y, t), M Tm3 (x, y, t) represent the outputs of the Tm1 and Tm3 medullary neurons in step (3.2), respectively; (4.2) Since there is a lateral inhibition mechanism in STMDs neurons, simulating this mechanism can suppress the response to large targets in the background, thereby reducing the false detection rate. Based on this, the lateral inhibition function is implemented by convolving the output of STMDs neurons with the lateral inhibition function. First, the lateral inhibition function W in STMDs neurons is defined. STMD (x, y) is: Where k and m are constants, [x] + and [x] - Let max(x, 0) and min(x, 0) represent these values ​​respectively. Then, a LoG distribution function is defined as follows: Where σ3 represents the LoG distribution function The standard deviation in; Then use S TMDI Let (x, y, t) represent the output of the STMDs neuron after lateral inhibition, then: S TMDI (x,y,t)=S TMD (x,y,t)*W STMD (x,y) Among them, S TMD (x, y, t) is the output of the STMDs neuron in step (4.1); Finally, S is compared using a threshold judgment method. TMDI (x, y, t) and the detection threshold β STMD The magnitude of S determines whether the target detected at (x, y) at time t is a real target; if S TMDI (x, y, t) > β STMD If the target is positive, it is considered a real moving target; otherwise, it is considered a non-moving target.

5. The remote sensing image target tracking method according to claim 1, simulating the compound eye vision system of flies, in step (5), based on the completion of detection of moving targets in all frames, the minimum bounding rectangle method and the neighborhood threshold discrimination method are designed to further analyze and judge the target trajectory, and the final target tracking result is obtained as follows: (5.1) Based on the results of moving target detection in each frame obtained in step (4), the minimum bounding rectangle method is used to obtain the coordinates [x] of the predicted bounding box in that frame. p y p w p h p ],in, (x p y p ) represents the coordinates of the top-left corner of the prediction box, w p and h p Let x and y represent the width and height of the predicted bounding box, respectively; then, use the center coordinates (x, y) of the predicted bounding box. c y c The coordinates of the predicted target are used as the coordinates of the target. (5.2) Design a neighborhood discrimination method to determine whether the moving targets detected in two consecutive frames are on the same trajectory; calculate the Euclidean distance between the position coordinates of the two consecutive frames as the distance between the two predicted moving targets; if the distance between the position of the predicted moving target in the current frame and the position of the predicted moving target in the next frame is less than the neighborhood threshold β n If the predicted moving target in a given frame differs from the predicted moving target in the two frames in the next frame by more than a neighborhood threshold β, then the two targets are considered to be on the same trajectory, and the target tracking is successful. Otherwise, if the position of a detected predicted moving target in a given frame differs from the positions of the detected predicted moving targets in the two frames before and after the next frame by more than a neighborhood threshold β, the target tracking is considered successful. n If the target is not tracked, the tracking of the moving object in that frame is considered to have failed.

Images