Method for creating infrared sequence multi-target dataset

Abstract

The present invention belongs to the field of infrared small target detection technology, and in particular relates to a method for generating an infrared sequence multi-target dataset. The method comprises: S1: segmenting a large-scale infrared background image into at least two sub-images; S2: generating simulated targets, including point-like targets and spot-like targets; S3: adding the simulated targets to the sub-images of the large-scale infrared background image and fusing the simulated targets with the sub-images; S4: simulating the motion states of multiple targets based on a composite motion trajectory equation and a perspective switching and motion state inheritance mechanism. The present invention can efficiently generate large-scale, high-quality, and diverse infrared small target datasets, providing data support for the research and application of infrared small target detection algorithms.

Description

Technical Field

[0001] The present invention belongs to the technical field of infrared small target detection, and in particular relates to a method for producing an infrared sequence multi-target data set. Background Art

[0002] Infrared small target detection technology has important application value in both military and civilian fields. However, the high cost, long cycle time, and difficulty in labeling large-scale, high-quality infrared small target datasets in real scenes have seriously restricted the research and application of infrared small target detection. Some existing infrared small target dataset simulation methods have the following shortcomings:

[0003] 1. Single target model: Most of them use point targets or geometric shape models with fixed grayscale values, which makes it difficult to simulate the diversity and complexity of real targets.

[0004] 2. Insufficient background fidelity: The simulation of background noise and interference is not realistic enough, and there is no tracking target background switching state, which makes it difficult to reflect the target detection challenges in complex scenarios.

[0005] 3. Target motion state: The generated dataset is small in size and cannot meet the training requirements of deep learning, especially for deep neural network models that require a large amount of training data. Summary of the Invention

[0006] In view of this, the present invention aims to provide a method for producing an infrared sequence multi-target dataset to solve the problems of the existing infrared small target dataset production methods, such as single target model, insufficient background realism and small dataset size. The present invention can efficiently generate large-scale, high-quality and diversified infrared dense small target datasets, providing data support for the research and application of infrared small target detection algorithms.

[0007] To achieve the above object, the technical solution created by the present invention is implemented as follows:

[0008] A method for preparing an infrared sequence multi-target data set comprises the following steps:

[0009] S1: Split the large-size infrared background image into at least two sub-images;

[0010] S2: Create simulation targets, including point targets and spot targets;

[0011] S3: adding a simulation target to the sub-image and performing fusion processing on the simulation target and the sub-image;

[0012] S4: Realize multi-target motion state simulation based on composite motion trajectory equations and perspective switching and motion state inheritance mechanisms.

[0013] Furthermore, in step S1 , a cropping window is set, and the large-size infrared background image is cropped with a step size of 256 pixels to obtain no less than two sub-images, and an independent coordinate system is established for each sub-image.

[0014] Furthermore, step S2 specifically includes the following steps:

[0015] S21: Use the two-dimensional Gaussian kernel distribution model to simulate the shape and radiation characteristics of the point target to obtain the point target. The two-dimensional Gaussian kernel distribution model is:

[0016] ;

[0017] Where G(x, y) represents the grayscale value of the point target at the image coordinate (x, y), A is the amplitude of the two-dimensional Gaussian kernel distribution model; (x0, y0) is the center of mass position of the point target; σ x and σ y are the standard deviations of the Gaussian kernel in the x and y directions respectively;

[0018] S22: For speckled targets, generate a two-dimensional Gaussian kernel distribution model centered at the specified centroid and superimpose Gaussian kernels within a pixel range around the centroid. The number of Gaussian kernels is no less than two and no more than nine. The brightness of the Gaussian kernel is 0.6 to 0.8 times that of the main kernel, and the extension of the Gaussian kernel is 0.6 to 0.8 times that of the main kernel.

[0019] S23: Constructing a random deformation function, and using the random deformation function to randomly deform the point target and the spot target respectively, to complete the production of the simulation target.

[0020] Furthermore, in step S23, the random deformation function is:

[0021] ;

[0022] ;

[0023] ;

[0024] in, is the deformed image, is the input image, A random number uniformly distributed in [0, 1] is generated for the pixel point at the coordinate position (x, y) of the input image; The smoothed displacement field is obtained by applying a Gaussian smoothing function with a standard deviation of σ = M to random noise; is the deformation strength parameter; is the displacement field in the x direction after the input image is Gaussian smoothed and scaled; is the displacement field in the y direction of the input image after Gaussian smoothing and scaling.

[0025] Furthermore, step S3 specifically includes the following steps:

[0026] S31: Dynamically assign weights to the sub-image and the simulation target using the following formula to obtain the weights of the sub-image and the simulation target:

[0027] ;

[0028] ;

[0029] ;

[0030] in, S To control the coefficient of sensitivity to brightness difference, the weight of the simulation target is , the weight of the sub-image is , is the pixel value difference between the sub-image and the simulation target, For the simulation target, is a sub-image;

[0031] S32: Perform weighted fusion of the sub-image and the simulation target using the following formula:

[0032] ;

[0033] in, is the fused image;

[0034] S33: Perform brightness compensation on the fused image using the following formula:

[0035] ;

[0036] ;

[0037] in, is the brightness compensation factor, This is the image after brightness compensation.

[0038] Furthermore, in step S4, the composite motion trajectory equation is:

[0039] ;

[0040] in,( x 0， y 0) is the initial coordinate of the simulation target, Δv x ( k ) and Δv y ( k) are the velocity increment parameters in the x-direction and y-direction respectively, α k and are the weight coefficients in the x and y directions respectively, sdx and sdy are the migration deviations of the coordinate system in the x and y directions, respectively.

[0041] Furthermore, in step S4, the perspective switching and motion state inheritance mechanism is specifically as follows: an edge buffer is set for each sub-image, the size of the edge buffer being 5% of the sub-image size, and the next sub-image that the simulation target will reach is obtained based on the motion direction of the target mass center of the simulation target in the edge buffer of the current sub-image;

[0042] When the simulation target enters the new sub-image, the coordinate value of the target mass center of the simulation target in the composite motion trajectory equation is updated to the coordinate value in the new sub-image.

[0043] Furthermore, step S4 specifically includes the following steps:

[0044] S41: inputting all the simulation targets produced in step S2 into the composite motion trajectory equation in sequence, and randomly adjusting the weight coefficient of the composite motion trajectory equation to obtain a simulation target with a motion trajectory;

[0045] S42: randomly selecting a sub-image as the currently displayed sub-image, taking all simulated targets whose target centroids enter the edge buffer of the currently displayed sub-image as targets to be observed, and selecting one target to be observed from the set of targets to be observed as the main target;

[0046] S43: Mapping the target centroid coordinates of the main target to the coordinate system of the adjacent sub-images through the affine transformation matrix, and retaining the motion state parameters of all targets to be observed.

[0047] Furthermore, the infrared small target multi-target dataset is a collection of sub-images that the main target passes through during its movement.

[0048] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0049] (1) The infrared sequence multi-target dataset preparation method created by the present invention uses a real infrared background image to perform sub-image segmentation, and adds random noise and illumination changes to the sub-image to generate a complex background.

[0050] (2) The method for producing an infrared sequence multi-target dataset created by this invention includes: 1) using a two-dimensional Gaussian kernel model and superposition mechanism to generate point and patch targets; 2) designing a random deformation function to generate targets with different geometric shapes and grayscale values, thereby improving the realism and diversity of target simulation; and 3) performing adaptive weighted fusion of the targets and the background. Through these operations, diverse target generation and background fusion are achieved.

[0051] (3) The infrared sequence multi-target data set preparation method created by the present invention makes the small target have a real motion state by designing the motion trajectory equation. By designing functions such as speed change, trajectory smoothing and perspective switching, the authenticity of the target motion state is enhanced, and the multi-target motion state simulation is realized.

[0052] (4) The present invention creates a method for producing an infrared sequence multi-target dataset. The dense infrared small target dataset generated by the present invention can be used to train and evaluate various infrared small target detection algorithms, such as methods based on traditional image processing, methods based on machine learning, and methods based on deep learning, thereby promoting the development and application of infrared small target detection technology. In addition, the present invention can efficiently generate large-scale, high-quality, and diverse infrared small target datasets, providing data support for the research and application of infrared small target detection algorithms. The present invention has the advantages of efficiency, authenticity, diversity, and automation, and can be widely used in the fields of infrared small target detection, infrared imaging algorithm performance testing, and infrared target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings, which constitute part of the present invention, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0054] Figure 1 A flow chart of a method for producing an infrared sequence multi-target data set according to an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of the structure of a method for producing an infrared sequence multi-target data set according to an embodiment of the present invention;

[0056] Figure 3 A schematic diagram of segmentation of a large-size infrared background image according to an embodiment of the present invention;

[0057] Figure 4 A simulation diagram of a point-shaped target for creating an embodiment of the present invention;

[0058] Figure 5 A simulation diagram of a patch target for creating an embodiment of the present invention;

[0059] Figure 6A simulation diagram of the multi-target motion trajectory of an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0061] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0062] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second" and the like are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, features defined as "first", "second" and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0063] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art can understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0064] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0065] like Figure 1-Figure 2As shown, the infrared sequence multi-target data set production method proposed in the present invention specifically includes the following steps: S1: dividing a large-size infrared background image into no less than two sub-images; S2: producing simulated targets, which include point targets and spot targets; S3: adding simulated targets to sub-images, and fusing the simulated targets and sub-images; S4: realizing multi-target motion state simulation based on the composite motion trajectory equation and the perspective switching and motion state inheritance mechanism.

[0066] The present invention proposes a method for preparing an infrared sequence multi-target dataset. The method uses a real infrared background image for sub-image segmentation and adds random noise and illumination changes to the sub-image to generate the background. A two-dimensional Gaussian kernel model and a superposition mechanism are used to generate point targets and patch targets to generate targets. A random deformation function is designed to simulate targets of different sizes. The target and background are fused, and a motion trajectory equation is designed to achieve trajectory smoothing and perspective switching, thereby realizing multi-target motion state simulation.

[0067] In some embodiments, in step S1 , a cropping window is set, and the large-size infrared background image is cropped with a step size of 256 pixels to obtain no less than two sub-images, and an independent coordinate system is established for each sub-image.

[0068] like Figure 3 As shown in the figure, taking diagonal segmentation as an example, assuming a cropping window size of 512 pixels × 512 pixels, the large-scale infrared background image is cropped with a step size of 256 pixels, ultimately completing the cropping of the entire large-scale infrared background image. The resolution of the large-scale infrared background image is greater than 1920x1080. To construct an infrared background that resembles a realistic scene, background images are first read from a pre-collected library of large-scale infrared background images. These background images are derived from real-world infrared scenes. To facilitate subsequent processing and annotation, the large-scale infrared background image is segmented into multiple 512×512 pixel sub-images, and an independent coordinate system is established for each sub-image (a two-dimensional coordinate system is established with the upper left corner of the sub-image as the coordinate origin). This segmentation method facilitates subsequent target motion trajectory simulation and perspective switching.

[0069] After segmenting the background image, each sub-image needs to be further processed to simulate the lighting variations and random noise found in real scenes. For example, the brightness and contrast of the sub-images are adjusted according to a preset lighting model to simulate lighting variations at different times of day and under different weather conditions. Furthermore, random noise, such as Gaussian noise or salt-and-pepper noise, is added to the sub-images to simulate the noise interference generated during infrared imaging. These processing steps significantly improve the realism of the background image and provide training data for infrared small target detection algorithms that is closer to real scenes.

[0070] In some embodiments, step S2 specifically includes the following steps:

[0071] S21: Use the two-dimensional Gaussian kernel distribution model to simulate the shape and radiation characteristics of the point target to obtain the point target. The two-dimensional Gaussian kernel distribution model is:

[0072] ;

[0073] Where G(x, y) represents the grayscale value of the point target at the image coordinate (x, y), A is the amplitude of the two-dimensional Gaussian kernel distribution model; (x0, y0) is the center of mass position of the point target; σ x and σ y are the standard deviations of the Gaussian kernel in the x and y directions respectively;

[0074] S22: For a speckled target, generate a two-dimensional Gaussian kernel distribution model centered at a specified centroid, and superimpose Gaussian kernels within a pixel range around the centroid, wherein the number of Gaussian kernels is no less than two and no more than nine, the brightness of the Gaussian kernels is 0.6 to 0.8 times that of the main kernel, and the extent of the Gaussian kernels is 0.6 to 0.8 times that of the main kernel;

[0075] S23: Constructing a random deformation function, and using the random deformation function to randomly deform the point target and the spot target respectively, to complete the production of the simulation target.

[0076] Furthermore, in step S23, the random deformation function is:

[0077] ;

[0078] ;

[0079] ;

[0080] in, is the deformed image, is the input image, A random number uniformly distributed in [0, 1] is generated for the pixel point at the coordinate position (x, y) of the input image; The smoothed displacement field is obtained by applying a Gaussian smoothing function with a standard deviation of σ = M to random noise; is the deformation strength parameter; is the displacement field in the x direction after the input image is Gaussian smoothed and scaled; is the displacement field in the y direction of the input image after Gaussian smoothing and scaling.

[0081] It should be noted that in order to simulate small infrared targets in real scenes, a two-dimensional Gaussian kernel model and superposition mechanism are adopted, and a random deformation function is designed to generate point targets and patch targets of different sizes.

[0082] For point targets, the point targets are set to Gaussian kernel form, and a single two-dimensional Gaussian kernel distribution model is used to model the different forms and radiation characteristics of the point targets. The two-dimensional Gaussian kernel distribution model can effectively simulate the energy distribution of the point targets, making them present a grayscale distribution effect similar to that of the real targets in the infrared image.

[0083] For patch targets, a Gaussian kernel group centered on the centroid coordinates is constructed, that is, a Gaussian kernel model with a large extension degree centered on the specified centroid is generated, and multiple Gaussian kernels with lower brightness and smaller extension degree are superimposed around it. This construction method effectively reflects the brightness distribution characteristics of patch targets in imaging, increases the naturalness and complexity of the target, and simulates the shape changes and background clutter of the target in the real environment.

[0084] In order to make the generated target more natural, a random deformation function is designed to randomly deform the generated target, making the shape of the simulated target in the infrared image more natural and asymmetric. (i.e. point target or patch target), set the deformation strength to , the deformed image It can be expressed as:

[0085] ;

[0086] in:

[0087] ;

[0088] ;

[0089] The random deformation process is: for each pixel position of the simulation target Generate random displacement , the random displacement is processed by Gaussian smoothing and has spatial continuity; each pixel in the original image I (i.e., the simulation target) is moved to the new position ; at the new coordinates Perform linear interpolation sampling on the original image I to obtain the deformed image .

[0090] According to the actual application scenario, various parameters of the simulation target can be set, such as the size and intensity of the simulation target. In addition, the shape and radiation characteristics of the Gaussian kernel can be adjusted and different point targets (such as Figure 4 ) and patch targets (e.g. Figure 5)’s infrared signature.

[0091] In some embodiments, step S3 specifically includes the following steps:

[0092] S31: Dynamically assign weights to the sub-image and the simulation target using the following formula to obtain the weights of the sub-image and the simulation target:

[0093] ;

[0094] ;

[0095] ;

[0096] in, S To control the coefficient of sensitivity to brightness difference, the weight of the simulation target is , the weight of the sub-image is , is the pixel value difference between the sub-image and the simulation target, For the simulation target, is a sub-image;

[0097] S32: Perform weighted fusion of the sub-image and the simulation target using the following formula:

[0098] ;

[0099] in, is the fused image;

[0100] S33: Perform brightness compensation on the fused image using the following formula:

[0101] ;

[0102] ;

[0103] in, is the brightness compensation factor, This is the image after brightness compensation.

[0104] It should be noted that fusion means assuming that the sub-image occupies a pixel value of 256×256 and the simulation target occupies a pixel value of 9×9. After fusion, the sub-image discards 9×9 pixels to display the simulation target.

[0105] Furthermore, after the target (simulated target) and background (sub-image) are generated, they are dynamically weighted and fused. The fusion process is as follows: 1. Dynamic weight allocation: The sigmoid function is used for dynamic weight allocation. The weight of the target area is , the weight of the background area is ; 2. Fusion generation: superimpose the weighted target and background components; 3. Brightness compensation: To prevent the generated area from being darker than the original background image, multiply the fused image by the brightness compensation factor .

[0106] In some embodiments, in step S4, the composite motion trajectory equation is:

[0107] ;

[0108] in,( x 0， y 0) is the initial coordinate of the simulation target, Δv x ( k ) and Δv y ( k ) are the velocity increment parameters in the x-direction and y-direction respectively, α k and are the weight coefficients in the x and y directions respectively, sdx and sdy are the migration deviations of the coordinate system in the x and y directions, respectively.

[0109] It should be noted that the traditional method describes the target displacement law through predefined motion matrices (such as uniform velocity motion matrix and uniform acceleration motion matrix), but the fixed incremental parameters make the motion trajectory mechanically repetitive, making it difficult to simulate the maneuvering characteristics of the real target (such as speed change, turning, trajectory disturbance, etc.). To this end, the present invention proposes a dynamic random trajectory model, that is, by introducing a trajectory generation mechanism that combines random processes with physical constraints, a composite motion trajectory equation is constructed, such as Figure 6 As shown in the figure, the composite motion trajectory equation can create different motion trajectories for multiple simulation targets by introducing random weight coefficients and combining them with physical constraints to generate trajectories. This makes the trajectory of each simulation target unique and diverse, avoids the trajectories of different simulation targets being too similar, and makes the generated trajectories more natural and better able to simulate the actual situation.

[0110] In some embodiments, in step S4, the perspective switching and motion state inheritance mechanism is specifically as follows: setting an edge buffer for each sub-image, where the size of the edge buffer is 5% of the sub-image size, and obtaining the next sub-image that the simulated target will reach based on the motion direction of the target mass center of the simulated target in the edge buffer of the current sub-image;

[0111] When the simulation target enters the new sub-image, the coordinate value of the target mass center of the simulation target in the composite motion trajectory equation is updated to the coordinate value in the new sub-image.

[0112] It should be noted that in order to simulate the observation effect when the simulation target crosses different sub-images, the present invention designs a perspective switching and motion state inheritance mechanism:

[0113] 1. Edge trigger condition: When the target centroid enters the edge buffer of the sub-image, the perspective switching is initiated.

[0114] 2. Perspective migration: A grid coordinate system is constructed based on the two-dimensional rectangular coordinate system of each sub-image established during background segmentation. The spatial adjacency relationship between adjacent sub-images is established. According to the movement direction of the target center of mass when entering the image edge buffer, the movement direction of the target center of mass from the original coordinate system to the new coordinate system is obtained. For multiple simulation tasks, multi-directional perspective migration can be achieved.

[0115] 3. Motion parameter inheritance: After the target center of mass enters the new sub-image, the x and y in its motion equation are updated to the new coordinates of the coordinate system of the current sub-image to avoid trajectory distortion caused by sudden changes in the coordinate system.

[0116] In some embodiments, step S4 specifically includes the following steps: S41: input all simulation targets produced in step S2 into the composite motion trajectory equation in sequence, and randomly adjust the weight coefficient of the composite motion trajectory equation (randomly adjust α k and ), obtain a simulated target with a motion trajectory; S42: randomly select a sub-image as the currently displayed sub-image, take all simulated targets whose target mass centers enter the edge buffer of the currently displayed sub-image as targets to be observed, and select one target to be observed from the set of targets to be observed as the main target; S43: map the target mass center coordinates of the main target to the coordinate system of the adjacent sub-image through the affine transformation matrix, and retain the motion state parameters of all targets to be observed (the motion state parameters include the motion speed and motion direction of the main target).

[0117] It should be noted that when multiple simulated targets move in the image sequence, a priority queue management method is used to dynamically track the targets. The process is as follows:

[0118] 1. Edge proximity assessment: The Euclidean distance between each target's centroid and the image boundaries (four boundaries) is calculated in real time. The view switching condition is triggered when the target's centroid enters the image edge buffer (the position within 5% of the image width from the image boundary).

[0119] 2. Main target selection strategy: Among multiple simulation targets close to the edge, the main target is selected according to user needs.

[0120] 3. Coordinate system migration: When switching the perspective of the main target, the target center of mass coordinates are mapped to the coordinate system of the adjacent sub-image through the affine transformation matrix, while retaining the motion state parameters of all targets to be observed to ensure motion continuity.

[0121] In some embodiments, the infrared small target multi-target data set is a collection of sub-images that the main target passes through during its movement.

[0122] It should be noted that users can replace the main target according to actual needs to obtain different infrared small target multi-target data sets.

[0123] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved. This is not limited herein.

[0124] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A method for producing an infrared sequence multi-target dataset, characterized by: The specific steps include: S1: Split the large-size infrared background image into at least two sub-images; S2: creating a simulation target, wherein the simulation target includes a point target and a spot target; Step S2 specifically includes the following steps: S21: simulating the shape and radiation characteristics of the point target using a two-dimensional Gaussian kernel distribution model to obtain the point target. The two-dimensional Gaussian kernel distribution model is: ； Where G(x, y) represents the grayscale value of the point target at the image coordinate (x, y), A is the amplitude of the two-dimensional Gaussian kernel distribution model; (x0, y0) is the center of mass position of the point target; σ x and σ y are the standard deviations of the Gaussian kernel in the x and y directions respectively; S22: For a speckled target, generate a two-dimensional Gaussian kernel distribution model centered at a specified centroid, and superimpose Gaussian kernels within a pixel range around the centroid, wherein the number of Gaussian kernels is no less than two and no more than nine, the brightness of the Gaussian kernels is 0.6 to 0.8 times that of the main kernel, and the extent of the Gaussian kernels is 0.6 to 0.8 times that of the main kernel; S23: constructing a random deformation function, and using the random deformation function to randomly deform the point target and the spot target respectively, to complete the production of the simulation target; In step S23, the random deformation function is: ； ； ； in, is the deformed image, is the input image, A random number uniformly distributed in [0, 1] is generated for the pixel point at the coordinate position (x, y) of the input image; The smoothed displacement field is obtained by applying a Gaussian smoothing function with a standard deviation of σ = M to random noise; is the deformation strength parameter; is the displacement field in the x direction after the input image is Gaussian smoothed and scaled; is the displacement field in the y direction of the input image after Gaussian smoothing and scaling; S3: adding a simulation target to the sub-image, and performing fusion processing on the simulation target and the sub-image; S4: Realize multi-target motion state simulation based on composite motion trajectory equations and perspective switching and motion state inheritance mechanism; In step S4, the composite motion trajectory equation is: ； in,( x 0， y 0) is the initial coordinate of the simulation target, Δv x ( k ) and Δv y ( k ) are the velocity increment parameters in the x-direction and y-direction respectively, α k and are the weight coefficients in the x and y directions respectively, sdx and sdy are the migration deviations of the coordinate system in the x and y directions respectively, and t represents the t-th frame image; In step S4, the perspective switching and motion state inheritance mechanism is specifically as follows: setting an edge buffer for each sub-image, wherein the size of the edge buffer is 5% of the sub-image size, and obtaining the next sub-image that the simulation target will reach based on the motion direction of the target mass center of the simulation target in the edge buffer of the current sub-image; When the simulation target enters the new sub-image, the coordinate value of the target mass center of the simulation target in the composite motion trajectory equation is updated to the coordinate value in the new sub-image.

2. The method for producing an infrared sequence multi-target dataset according to claim 1, characterized in that: In step S1 , a cropping window is set, and the large-size infrared background image is cropped with a step size of 256 pixels to obtain no less than two sub-images, and an independent coordinate system is established for each sub-image.

3. The method for producing an infrared sequence multi-target dataset according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31: Dynamically assign weights to the sub-image and the simulation target using the following formula to obtain the weights of the sub-image and the simulation target: ； ； ； in, S To control the coefficient of sensitivity to brightness difference, the weight of the simulation target is , the weight of the sub-image is , is the pixel value difference between the sub-image and the simulation target, For the simulation target, is a sub-image; S32: Perform weighted fusion of the sub-image and the simulation target using the following formula: ； in, is the fused image; S33: Perform brightness compensation on the fused image using the following formula: ； ； in, is the brightness compensation factor, This is the image after brightness compensation.

4. The method for producing an infrared sequence multi-target dataset according to claim 1, characterized in that: Step S4 specifically includes the following steps: S41: inputting all the simulation targets produced in step S2 into the composite motion trajectory equation in sequence, and randomly adjusting the weight coefficient of the composite motion trajectory equation to obtain a simulation target with a motion trajectory; S42: randomly selecting a sub-image as the currently displayed sub-image, taking all simulated targets whose target centroids enter the edge buffer of the currently displayed sub-image as targets to be observed, and selecting one target to be observed from the set of targets to be observed as the main target; S43: Mapping the target centroid coordinates of the main target to the coordinate system of the adjacent sub-images through the affine transformation matrix, and retaining the motion state parameters of all targets to be observed.

5. The method for producing an infrared sequence multi-target dataset according to claim 4, characterized in that: The infrared small target multi-target data set is a collection of sub-images that the main target passes through during its movement.

Images