A low signal-to-clutter ratio infrared small target extraction method under a motion platform

Abstract

The application discloses a kind of low signal-to-clutter ratio infrared small target extraction methods under motion platform, comprising: based on the camera posture information obtained by inertial measurement unit, rotation compensation is carried out to infrared image sequence, and phase correlation method is estimated in combination with interframe translation, and the image sequence is realized image stabilization registration;Target detection is carried out to the infrared image sequence after image stabilization, and target track is established by multi-frame association;When target enters low signal-to-clutter ratio environment, with track estimation speed as center, construct multi-speed hypothesis set, carry out speed compensation and time domain accumulation to historical image sequence, and generate target enhancement graph by fusion, and carry out background suppression and sparse response generation;Based on the last frame track state, the position of target in current frame is predicted, the enhancement graph is re-detected in prediction area, and the detection threshold is adaptively adjusted according to track stability, to realize the continuous update and output of target track.The application effectively improves the detection sensitivity, continuity and system reliability of infrared small target under motion platform and low signal-to-clutter ratio condition.

Description

Technical Field

[0001] This invention belongs to the field of infrared target detection and image processing technology, specifically a method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform. Background Technology

[0002] Infrared small targets are characterized by their tiny size, weak signal, sparse texture, and complex motion. They often exhibit extremely low signal-to-noise ratios in long-distance imaging and complex dynamic background conditions, making single-frame detection difficult to achieve.

[0003] Existing detection methods are mainly divided into two categories: spatial domain and temporal domain. Spatial domain methods often enhance the target through local contrast or filtering, but they are prone to false detection in complex texture backgrounds. Temporal domain methods rely on the accumulation of multiple frames to enhance the signal, but they are difficult to guarantee alignment accuracy under platform jitter and parallax changes, often leading to energy diffusion and clutter amplification.

[0004] While inertial measurement units (IMUs) can be used for image stabilization, residual translation can still disrupt temporal consistency due to attitude noise and calibration errors. Deep learning methods, while adaptive, rely on large amounts of data and have insufficient generalization ability under low signal-to-noise ratio and cross-scene conditions.

[0005] Therefore, there is an urgent need for an image stabilization mechanism that integrates inertial constraints and phase coordination, combining multi-directional energy response detection, temporal consistency enhancement, and track-driven re-detection to achieve high-precision, stable, and continuous extraction of small infrared targets under a moving platform. Summary of the Invention

[0006] The purpose of this invention is to provide a method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform, which achieves image stabilization, enhancement, and continuous detection of small targets under low signal-to-clutter conditions while taking into account both algorithm stability and real-time performance.

[0007] The technical solution to achieve the objective of this invention is: a method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform, comprising the following steps:

[0008] Step 1: Based on the camera attitude information obtained by the inertial measurement unit, perform rotation compensation on the infrared image sequence, and combine the phase correlation method to estimate the inter-frame translation to achieve stable image registration of the image sequence.

[0009] Step 2: Detect targets in the stabilized infrared image sequence and establish target tracks by associating multiple frames;

[0010] Step 3: When the target enters a low signal-to-clutter environment, construct a set of multiple velocity assumptions centered on the trajectory estimated velocity, perform velocity compensation and temporal accumulation on historical image sequences, and fuse them according to the energy consistency factor and velocity weight under different velocity assumptions to generate a target enhancement map, and perform background suppression and sparse response generation.

[0011] Step 4: Based on the track status of the previous frame, predict the position of the target in the current frame, re-detect the enhanced map within the prediction area, and adaptively adjust the detection threshold according to the track stability to achieve continuous updating and output of the target track.

[0012] Compared with the prior art, the present invention has the following significant advantages:

[0013] 1. Inertial-phase co-stabilization to suppress plateau interference:

[0014] By integrating inertial attitude information with phase residual compensation, dual-layer image stabilization of rotation and translation is achieved, effectively suppressing background drift caused by jitter, rolling shutter, and attitude errors.

[0015] 2. Multi-directional energy response detection enhances local separability:

[0016] By using a detection model that integrates the second derivative in four directions with local brightness contrast, the response difference between small targets and complex textured backgrounds is significantly enhanced, reducing the false detection rate.

[0017] 3. Enhanced multi-speed consistency, improving signal convergence:

[0018] By introducing a velocity neighborhood and consistency weighting mechanism, energy fusion and adaptive noise suppression under multiple assumptions are achieved, ensuring that the target energy is concentrated and superimposed while the background clutter is dispersed and canceled.

[0019] 4. Track-driven closed-loop detection to maintain output stability:

[0020] By employing a trajectory prediction and threshold adaptation mechanism, the detection threshold is dynamically adjusted over time to ensure continuous acquisition and robust output in low signal-to-noise ratio environments. Attached Figure Description

[0021] Figure 1 Overall flowchart of the method.

[0022] Figure 2 The original infrared image after image stabilization.

[0023] Figure 3 This is a magnified view of a small infrared target after image stabilization.

[0024] Figure 4 The results and candidate peaks (including NMS) are obtained by multi-directional energy response and comparative weighting.

[0025] Figure 5 This is a schematic diagram of candidate peak correlation and track formation.

[0026] Figure 6 A local map showing the target entering a low signal-to-noise ratio environment.

[0027] Figure 7 Predicting the spatiotemporal location of a target under different velocity assumptions.

[0028] Figure 8 This is a comparison image of the enhanced infrared small target and the original small target. Detailed Implementation

[0029] like Figure 1 As shown, this invention proposes a method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform. This method is based on an inertial constraint and phase-coordinated image stabilization framework, constructing a closed-loop processing flow consisting of image stabilization, detection, enhancement, and re-detection. By combining inertial compensation, temporal consistency enhancement, and track-driven detection, high-precision extraction of small infrared targets under low signal-to-clutter conditions is achieved. The method comprises four steps: inertial constraint background registration, target detection and track establishment, temporal consistency enhancement, and track-driven re-detection, which are specifically as follows:

[0030] Step 1, Inertial Constraint Background Registration: The image sequence is rotated using attitude information acquired by the inertial measurement unit (IMU), and inter-frame translation is estimated using a phase correlation algorithm to achieve accurate registration of the infrared image sequence, thereby eliminating the influence of platform jitter and attitude changes on the image sequence. The specific steps are as follows:

[0031] (11) Attitude acquisition and despin compensation

[0032] The first digit is obtained using the gyroscope and accelerometer in the inertial measurement unit (IMU). The camera's three-axis angular velocity and linear acceleration information at each frame is used, and the above sensor data is fused and estimated using an extended Kalman filter (EKF) to obtain the first... Euler angles of pose for the frame camera:

[0033]

[0034] in Indicates circling The rotation angle of the axis; Indicates circling The rotation angle; Indicates circling The rotation angle.

[0035] Calculate the first based on the Euler angles. The rotation matrix corresponding to the frame:

[0036]

[0037] in , , They represent circumference respectively. , , The rotation matrix of the axis is expressed as:

[0038]

[0039] Using rotation matrix For the first The infrared image frame is subjected to reverse rotation compensation, and the image coordinates are transformed to the reference attitude coordinate system, thereby eliminating the rotational motion component caused by the camera attitude change and realizing derotation registration of the image sequence.

[0040] Among them, the Frame infrared image denoted as , Represents the pixel coordinates of the image.

[0041] (12) Homography initialization and phase residual alignment

[0042] After rotation compensation is completed, the residual translation between adjacent frames is further estimated to achieve more accurate inter-frame registration.

[0043] First, calculate the inter-frame rotation relationship based on the rotation matrices of two adjacent frames:

[0044]

[0045] in, Indicates the first Frame rotation matrix; Indicates the first Frame rotation matrix.

[0046] Combined with camera intrinsic matrix Construct the rotation homography matrix:

[0047]

[0048] After initial alignment of two adjacent image frames using the rotation homography matrix, the residual translation between the two frames is estimated using the phase correlation method. Let the Fourier transforms of the two image frames be:

[0049] ,

[0050] The residual translation is calculated using the following formula:

[0051]

[0052]

[0053] in, Indicates complex conjugation; Indicates the inverse Fourier transform; These represent the translation amounts of the image in the horizontal and vertical directions, respectively.

[0054] Construct a translation matrix based on the translation amount:

[0055]

[0056] Furthermore, the cooperative homography matrix integrating rotation and translation is obtained:

[0057]

[0058] Using the aforementioned cooperative homography matrix The first Frame image projected to the first In the frame coordinate system, rotation and translation compensation between adjacent image frames are achieved, thereby completing high-precision image stabilization and registration of infrared image sequences.

[0059] Step 2, Target Detection and Track Establishment: After image sequence registration, target detection is performed on the stabilized infrared image sequence, and a target track is established through multi-frame correlation. Specifically, firstly, the target response map of the image is calculated using local multi-directional energy response and brightness contrast weighting, from which candidate peaks that may correspond to small infrared targets are extracted; then, non-maximum suppression is performed on the candidate response map to obtain sparse candidate target points; finally, a continuous and stable target track is established through correlation of candidate points between adjacent frames and velocity estimation. (See...) Figure 2 The original infrared image after image stabilization; Figure 3 A magnified view of a small infrared target after image stabilization; Figure 4 The detection results and candidate peaks (including non-maximum suppression) are obtained by multi-directional energy response and comparative weighting. Figure 5 This diagram illustrates the correlation between candidate peaks and track formation. The specific steps are as follows:

[0060] (21) Calculation of multi-directional energy response

[0061] The first image after stabilization Frame infrared image A second-order polynomial is fitted to the local neighborhood of the image to obtain local curvature information.

[0062] Let the coefficients of the second-order polynomial obtained by local fitting be... At the direction angle Calculate the second-order directional derivative response of this pixel:

[0063]

[0064] in: Represents pixels In direction The second-order directional derivative response; These are the curvature coefficients obtained by fitting a second-order polynomial. This is the direction angle.

[0065] In this embodiment, four main directions are selected:

[0066]

[0067] The above calculations yield response maps in four directions, which are used to describe the curvature variation characteristics of the local structure of the image in different directions.

[0068] (22) Local contrast calculation and response fusion

[0069] After obtaining the directional response, the brightness difference between the target area and the background area is further calculated to enhance the separability between the small target and the surrounding background.

[0070] For each pixel location in the image Within its neighborhood, a three-layer window structure is constructed, including: the target core area, the transition area, and the background area.

[0071] Calculate the luminance statistics for each of the three regions, and then calculate the local contrast ratio based on the statistical results:

[0072]

[0073] in, Represents pixels Local contrast; Indicates the average brightness of the target core region This represents the average brightness of the transition region; This represents the standard deviation of brightness in the background region. Subsequently, the second-order derivative responses in each direction are weighted and fused with the local contrast to obtain the target response map:

[0074]

[0075] in: Indicates the first The target response value of the frame; Represents the weighting coefficients for different directions; This indicates that the responses from multiple directions are accumulated and fused. Through the above fusion process, the local response of small targets can be enhanced, while suppressing interference from complex textured backgrounds.

[0076] (23) Candidate peak extraction

[0077] After obtaining the target response map Then, non-maximum suppression processing is applied to extract local peak values ​​that may correspond to small targets.

[0078] Specifically, a neighborhood window is constructed centered on each pixel in the response map, and it is determined whether the response value of that pixel is the maximum value within the neighborhood. If the response value of a pixel is greater than that of other pixels in its neighborhood, that pixel is retained as a candidate target point; otherwise, it is suppressed.

[0079] The nonmaximum suppression operation can be expressed as:

[0080]

[0081] in, Indicates the first The set of candidate peaks for a frame; Indicates nonmaximum operation; This represents the radius of the neighborhood window. This step yields a sparse set of candidate target locations.

[0082] (24) Multi-frame association and track establishment

[0083] After obtaining the set of candidate target points for each frame, the candidate targets between adjacent frames are correlated to establish a continuous track.

[0084] First, based on the existing track state in the previous frame, the constant velocity model is used to predict the target's position in the current frame:

[0085]

[0086] in, Indicates the target position in the previous frame; This indicates the target velocity estimated in the previous frame; Indicates the inter-frame time interval; This indicates the predicted target location.

[0087] A gated search region is constructed around the predicted location, and the distance between the candidate target point and the predicted location is calculated. When the distance meets the gating condition, the candidate point is selected as a matching candidate.

[0088] When multiple candidate matches exist, the optimal matching relationship is selected through the matching cost function, and the global optimal matching is solved using the Hungarian algorithm, thereby completing the track association.

[0089] By matching consecutive frames, an initial set of infrared small targets can be gradually established, providing motion prior information for subsequent temporal enhancement and re-detection.

[0090] Step 3: When the target enters a low signal-to-clutter environment, to enhance the target signal and suppress background interference, a multi-velocity hypothesis set is constructed centered on the trajectory-estimated velocity. Velocity compensation and temporal accumulation are performed on historical image sequences, and weighted fusion is performed based on the energy consistency factor and velocity weights under different velocity hypotheses to generate a target enhancement map. (See...) Figure 6 A local map showing the target entering a low signal-to-noise ratio environment; Figure 7 A schematic diagram illustrating the prediction of the spatiotemporal position of a target under different velocity assumptions; Figure 8 (This is a comparison of the enhanced infrared small target with the original small target.)

[0091] (31) Multi-velocity assumption and forward alignment

[0092] Based on the track information obtained in step 2, let the current frame be the [number]. The frame, whose estimated target velocity is

[0093]

[0094] in, The velocity is in the horizontal direction; The velocity is in the vertical direction.

[0095] Using this estimated velocity as the center, construct a set of multiple velocity hypotheses within its neighborhood:

[0096]

[0097] in, Indicates the first A speed assumption; Indicates the velocity disturbance; This indicates the number of velocity assumptions.

[0098] For any historical frame Calculate its predicted displacement in the current frame based on the velocity assumption:

[0099]

[0100] in, Indicates from frame to frame The displacement; Number the current frame; Assign historical frame numbers. Based on the aforementioned displacement, perform forward mapping on the historical frame image to align it with the coordinates of the current frame:

[0101]

[0102] in, Indicates the first Frame image; Indicating the velocity assumption Align the pixel values ​​down to the current frame. Through the above operations, a sequence of historical images aligned under different speed assumptions can be obtained.

[0103] (32) Time-domain energy accumulation

[0104] After completing multi-frame alignment, historical frame images are accumulated in the temporal domain to enhance the target signal.

[0105] For the The velocity assumption has a time-domain cumulative energy defined as:

[0106]

[0107] in, Indicating the velocity assumption Accumulated energy below; This represents the corresponding background estimate; This represents the square operation. By accumulating the energy of multiple historical frames, the target signal along the actual motion trajectory can be gradually enhanced, while the background noise tends to become average.

[0108] (33) Energy consistency and velocity weighted fusion

[0109] Since different velocity assumptions correspond to different cumulative effects, it is necessary to perform weighted fusion of the energies of each velocity assumption. First, calculate the energy consistency factor for each velocity assumption:

[0110]

[0111] in, Indicates the first Energy consistency of the velocity assumption; This represents the maximum cumulative energy among all velocity assumptions. Velocity weights are also calculated based on the difference between the actual velocity and the trajectory-estimated velocity.

[0112]

[0113] in, For the first The weight of each velocity assumption; This is a speed weighting adjustment parameter; For summation index variables.

[0114] Finally, the cumulative energy of each velocity assumption is weighted and fused to obtain the enhanced map:

[0115]

[0116] in, This represents the enhanced response map after fusion. Through the above fusion process, the velocity assumption, consistent with the actual target motion, receives higher weight, thereby further enhancing the target energy and suppressing background clutter.

[0117] (34) Background suppression and sparse response generation

[0118] To further suppress background interference, the local background standard deviation is introduced as a suppression term in the enhancement image:

[0119]

[0120] in, Indicates the local background standard deviation; This represents the background suppression coefficient.

[0121] Then, background subtraction is performed on the enhanced image:

[0122]

[0123] Non-maximum suppression (NMS) was then applied to the results to obtain the enhanced sparse response map:

[0124]

[0125] in, This represents the enhanced set of candidate target points; The non-maximum suppression window radius is defined. This sparse response map is used for subsequent track-driven re-detection and state updates.

[0126] Step 4, Track-driven re-detection: Based on the track state of the previous frame, predict the position of the target in the current frame, and construct a search window centered on the predicted position. Perform local re-detection and confidence update in the enhanced map; at the same time, adaptively adjust the detection threshold according to the track stability index to maintain the continuity and stability of target detection in low signal-to-clutter environment.

[0127] (41) Track prediction and search window generation

[0128] Let the previous frame be the first The status of the track is position. and speed Predict the target in the current frame using a constant velocity model Position in the middle:

[0129]

[0130] in, Indicates the first The track is in Target position of the frame; Indicates the corresponding target speed; Indicates the inter-frame time interval; This indicates the predicted target location. (The predicted location is used as the reference.) Centered on the local search window in the enhanced response graph:

[0131]

[0132] in Indicates the first The track is in The search area in the frame.

[0133] (42) Response detection within the window

[0134] In the search window Local maximum response in the internally computed augmented response plot:

[0135]

[0136] in, This represents the enhanced response map obtained in step 3; Indicates the first The maximum response value of the track in the current frame.

[0137] when

[0138]

[0139] When a target is detected at a given location, the detection result is taken as a new observation point.

[0140] in This is the detection threshold.

[0141] (43) Track status update

[0142] When the detection is successful, the track status is updated based on the observed location.

[0143] Let the detected target location be Calculate the observation residuals:

[0144]

[0145] Update the target position and velocity using residuals:

[0146]

[0147]

[0148] in, Indicates the updated target location; Indicates the updated target speed; Update the gain for the location; The speed gain is updated. This update improves the accuracy of track estimation.

[0149] (44) Track stability assessment and threshold adaptation

[0150] To maintain track continuity in low signal-to-clutter environments, the stability of the track is assessed, and the detection threshold is dynamically adjusted based on the stability.

[0151] Within a time window length of Calculate the track stability index within the range:

[0152]

[0153] in, Indicates the first Stability of the flight path; Indicates the first The track is in The response value of the frame.

[0154] The detection threshold is adaptively adjusted based on stability.

[0155]

[0156] in, This is the initial detection threshold; This is the threshold adjustment coefficient.

[0157] If the confidence level of a certain track is lower than a preset threshold in multiple consecutive frames, the track is terminated; the remaining tracks are output as stable targets.

Claims

1. A method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform, characterized in that, Includes the following steps: Step 1: Based on the camera attitude information obtained by the inertial measurement unit, perform rotation compensation on the infrared image sequence, and combine the phase correlation method to estimate the inter-frame translation to achieve stable image registration of the image sequence. Step 2: Detect targets in the stabilized infrared image sequence and establish target tracks by associating multiple frames; Step 3: When the target enters a low signal-to-clutter environment, construct a set of multiple velocity assumptions centered on the trajectory estimated velocity, perform velocity compensation and temporal accumulation on historical image sequences, and fuse them according to the energy consistency factor and velocity weight under different velocity assumptions to generate a target enhancement map, and perform background suppression and sparse response generation. Step 4: Based on the track status of the previous frame, predict the position of the target in the current frame, re-detect the enhanced map within the prediction area, and adaptively adjust the detection threshold according to the track stability to achieve continuous updating and output of the target track.

2. The method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform according to claim 1, characterized in that, The specific methods for achieving image stabilization and registration of infrared image sequences by performing rotation compensation based on camera attitude information acquired by the inertial measurement unit and combining it with phase correlation methods to estimate inter-frame translation include: The camera's angular velocity and linear acceleration information are acquired using an inertial measurement unit (IMU), and then fused and estimated using an extended Kalman filter to obtain the camera's angular velocity and linear acceleration at the [missing information - likely a specific point in time]. Frame pose Euler angles And calculate the rotation matrix based on the Euler angles. The rotation matrix is ​​used to perform reverse rotation compensation on the infrared image to eliminate rotational motion caused by changes in camera attitude. The residual translation between adjacent image frames is calculated using the phase correlation method, and a translation matrix is ​​constructed based on the residual translation. The translation matrix and the rotation matrix are combined to form a homography matrix, and joint rotation and translation compensation are performed on the image frames to achieve stable registration of the infrared image sequence.

3. The method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform according to claim 2, characterized in that, The specific process of calculating the residual translation between adjacent image frames using the phase correlation method is as follows: Calculate the inter-frame rotation relationship based on the rotation matrices of two adjacent frames; in, Indicates the first Frame rotation matrix; Indicates the first Frame rotation matrix; Combined with camera intrinsic matrix Construct the rotation homography matrix: Initial alignment of two adjacent frames is performed using a rotation homography matrix, and the residual translation between the two frames is estimated using a phase correlation method. The Fourier transforms of the two frames are as follows: ， The residual translation amount is as follows: in, Indicates complex conjugation; Indicates the inverse Fourier transform; These represent the translation amounts of the image in the horizontal and vertical directions, respectively.

4. The method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform according to claim 1, characterized in that, The specific method for target detection in the stabilized infrared image sequence and for establishing the target trajectory through multi-frame correlation is as follows: Step 21: Multidirectional energy response calculation: The first image after stabilization Frame infrared image A second-order polynomial is fitted to the local neighborhood to obtain response maps in four directions; Step 22: Local contrast calculation and response fusion: The position of each pixel in the stabilized infrared image A three-layer window structure is constructed within the neighborhood, including: target core area, transition area, and background area; Calculate the luminance statistics for each of the three regions, and then calculate the local contrast ratio based on the statistical results: in, Represents pixels Local contrast; Indicates the average brightness of the target core region This represents the average brightness of the transition region; This represents the standard deviation of brightness in the background area; The target response map is obtained by weighted and fused with the second derivative responses in each direction and the local contrast. in: Indicates the first The target response value of the frame; Represents the weighting coefficients for different directions; This indicates that responses from multiple directions are accumulated and fused. Represents pixels In direction The second-order directional derivative response; Step 23: Candidate Peak Extraction: Analyze the target response map Non-maximum suppression is performed to extract local peaks that may correspond to small targets, thus obtaining a sparse set of candidate target locations. Step 24: Multi-frame association and track establishment: Associate candidate targets between adjacent frames to establish continuous tracks.

5. The method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform according to claim 4, characterized in that, The specific method for associating candidate targets between adjacent frames and establishing continuous tracks is as follows: Based on the existing track status in the previous frame, the constant velocity model is used to predict the target's position in the current frame: in, Indicates the target position in the previous frame; This indicates the target velocity estimated in the previous frame; Indicates the inter-frame time interval; Indicates the predicted target location; A gated search region is constructed around the predicted location, and the distance between the candidate target point and the predicted location is calculated. When the distance meets the gating condition, the candidate point is used as a matching candidate. When multiple matching candidates exist, the optimal matching relationship is selected through the matching cost function, and the global optimal matching is solved using the Hungarian algorithm, thereby completing the track association.

6. The method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform according to claim 1, characterized in that, When the target enters a low signal-to-clutter environment, a set of multiple velocity assumptions is constructed centered on the trajectory-estimated velocity. Velocity compensation and temporal accumulation are performed on historical image sequences. The target enhancement map is then generated by fusing these assumptions based on the energy consistency factor and velocity weights under different velocity assumptions. Step 31: Based on the track information obtained in Step 2, determine the... The estimated target velocity of the frame image is in, The velocity is in the horizontal direction; The velocity is in the vertical direction; Centered on the estimated target velocity, construct a set of multiple velocity hypotheses within its neighborhood: in, Indicates the first A speed assumption; Indicates the velocity disturbance; Indicates the number of velocity assumptions; For any historical frame Calculate its predicted displacement in the current frame based on the velocity assumption: in, Indicates from frame to frame The displacement; Number the current frame; Number the historical frames; Based on the predicted displacement, the historical frame images are forward-mapped to align with the coordinates of the current frame, resulting in a sequence of aligned historical images under different velocity assumptions: in, Indicates the first Frame image; Indicating the velocity assumption Bottom-aligned to the pixel value of the current frame; Step 32: Calculate the first... The temporal cumulative energy of a velocity hypothesis is accumulated by accumulating the energy of multiple historical frames, so that the target signal along the direction of the real motion trajectory is gradually enhanced, while the background noise tends to be averaged. Step 33: Calculate the energy consistency factor for each velocity assumption: in, Indicates the first Energy consistency of the velocity assumption; This represents the maximum cumulative energy among all velocity assumptions; Simultaneously, speed weights are calculated based on the difference between the speed and the estimated speed from the trajectory; The cumulative energy of each velocity assumption is weighted and fused to obtain the enhanced map: in, This represents the enhanced response map after fusion. Step 34: Introduce the local background standard deviation as a suppression term into the enhancement plot: in, Indicates the local background standard deviation; Background suppression coefficient; Background subtraction is performed on the enhanced image: Nonmaximum suppression is then applied to the results to obtain the enhanced sparse response map: in, This represents the enhanced set of candidate target points; The radius of the non-maximum suppression window.

7. The method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform according to claim 6, characterized in that, The calculation of the first The time-domain cumulative energy of each velocity assumption is specifically as follows: in, Indicating the velocity assumption The accumulated energy below; This represents the corresponding background estimate; This represents the squaring operation.

8. The method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform according to claim 6, characterized in that, The speed weight is calculated based on the difference between the actual speed and the estimated speed from the track, specifically as follows: in, For the first The weights of each velocity assumption; This is a speed weighting adjustment parameter; For summation index variables.

9. The method for extracting small infrared targets with low signal-to-clutter ratio under a moving platform according to claim 1, characterized in that, The method for continuously updating and outputting the target trajectory is as follows: Based on the trajectory state of the previous frame, the target's position in the current frame is predicted; the enhanced map is re-detected within the prediction area; and the detection threshold is adaptively adjusted according to the trajectory stability. Step 41: Track prediction and search window generation: Let the previous frame be the first The status of the track is the position. and speed Predict the target in the current frame using a constant velocity model Position in the middle: in, Indicates the first The track is in Target position of the frame; Indicates the corresponding target speed; Indicates the inter-frame time interval; Indicates the predicted target location; To predict location A local search window is constructed in the enhanced response graph, centered on the target area. ,in Indicates the first The track is in The search area in the frame; Step 42: Response detection within the window: In the search window Local maximum response in the internally computed augmented response plot: in, This represents the enhanced response map obtained in step 3; Indicates the first The maximum response value of the track in the current frame; when When a target is detected at a given location, the detection result is taken as a new observation point. in, The detection threshold; Step 43: Track Status Update: When the detection is successful, update the track status based on the observed position; Let the detected target location be Calculate the observation residuals: Update the target position and velocity using residuals: in, Indicates the updated target location; Indicates the updated target speed; Update the gain for the location; Update the gain for speed.

Images