Target detection tracking method for resisting infrared decoy interference based on multi-dimensional features

Abstract

The invention relates to a multi-dimensional feature-based anti-infrared decoy interference target detection tracking method, and relates to the technical field of computer vision and target tracking, and the method comprises the steps: constructing a detection tracking network through a data set generated through simulation, and obtaining track-level time series data; extracting a radiation energy feature, a geometric contour feature and a motion vector feature of the target from the trajectory-level time sequence data, and fusing the features into a multi-dimensional dynamic vector; constructing a Trans-SRU cascade time sequence discrimination network, inputting a multi-dimensional dynamic vector, focusing a key radiation attenuation and maneuvering change stage in combination with parallel feature processing and a time attention mechanism, and outputting a classification result of a target and a bait; the defect that a traditional circulation network is insufficient in capturing the cross-cycle physical process is overcome, and the stable anti-interference capacity is provided for an infrared guidance system in the complex environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and target tracking, and particularly relates to a target detection and tracking method based on multi-dimensional feature anti-infrared decoy jamming. BACKGROUND

[0002] With the continuous upgrading of infrared guided weapons, as a key countermeasure, the jamming effectiveness of infrared decoy bombs directly determines the survivability of the combat platform. The decoy bomb simulates the thermal radiation characteristics and motion trajectory of the target vehicle accurately, induces the target recognition deviation of the infrared seeker, and forms a tactical level defense barrier. However, in the real combat environment, the trajectories of the target and multiple decoys are intertwined, the radiation characteristics are similar, and the complex background interference causes the traditional tracking system to face the serious challenges of a sharp increase in misjudgment rate and trajectory breakage, and it is urgent to develop intelligent anti-jamming technology with physical law cognition ability.

[0003] The traditional anti-jamming method is mainly based on motion feature analysis or radiation feature recognition; the former uses Kalman or particle filtering to rely on motion difference for discrimination, but it is easy to fail when the target is low-speed or the decoys are dense; the latter classifies through instantaneous infrared statistics, but ignores the time-varying evolution law of thermal radiation, and the modeling of thermal decay rate difference is insufficient, and the robustness is low. In recent years, although the deep learning such as YOLO series detector performs efficiently in the positioning task, it does not utilize the radiation homology of target-decoy, and lacks the modeling ability of time evolution, resulting in high false detection rate; the tracking algorithm such as DeepSORT depends on appearance matching, and the identity frequently switches when the texture is similar.

[0004] In order to break through the above bottleneck, the front-line research begins to explore a multi-modal fusion scheme: in the spatial dimension, the attention mechanism is introduced to enhance the feature discrimination, such as the Transformer module which models the global dependence to improve the detection robustness in the occlusion scene; in the time dimension, a bidirectional recurrent network (such as Bi-LSTM) is used to analyze the radiation energy curve, trying to capture the difference of thermal decay trend; however, the existing scheme still has defects: the spatial detection network is insufficient in extracting the geometric deformation features of the target, resulting in limited scale adaptability; the feature extraction excessively depends on a single radiation channel; the time modeling network has weak learning ability for long-range physical evolution law, and it is difficult to meet the precise discrimination demand of high dynamic environment. SUMMARY

[0005] The purpose of the present application is to provide a target detection and tracking method based on multi-dimensional feature anti-infrared decoy jamming, which effectively solves the defect of insufficient modeling of cross-cycle evolution law by traditional recurrent network.

[0006] To solve the above technical problems, the present application adopts the following technical scheme: a target detection and tracking method based on multi-dimensional feature anti-infrared decoy jamming, comprising the following steps:

[0007] S1, generating a dataset by simulation, the dataset comprising "dynamic heat source" category data and corresponding basic annotation data containing a bounding box, a millisecond timestamp, and kinematic parameters;

[0008] S2, constructing a detection and tracking network to obtain trajectory-level timing data based on the dataset;

[0009] S3, extracting radiation energy features, geometric profile features, and motion vector features of the target from the trajectory-level timing data, and after alignment and normalization processing of the radiation energy features, the geometric profile features, and the motion vector features, fusing them into a multi-dimensional dynamic vector;

[0010] S4, constructing a Trans-SRU cascaded timing discrimination network, comprising:

[0011] a simplified recurrent unit that models frame-level parallelism of the sequence through a gating mechanism to quickly capture local dynamic features;

[0012] a temporal attention mechanism module that corrects time difference information based on embedding of relative position encoding, so that multi-head self-attention can focus on key phase turning points of radiation decay;

[0013] an output layer that uses a gated linear unit to fuse local and global timing dependencies to obtain a decision vector;

[0014] S5, inputting the multi-dimensional dynamic vector to the Trans-SRU cascaded timing discrimination network to obtain a category discrimination result of the target / decoy through Sigmoid activation.

[0015] Preferably, in S1, a dynamic confrontation sequence of the target and the decoy is generated by a simulation platform to simulate thermal radiation characteristics and motion trajectories in a complex environment, the thermal radiation characteristics are modeled by introducing material specific emissivity and surface temperature field gradients; the target and the decoy are uniformly labeled as "dynamic heat source" category.

[0016] Further preferably, the dataset construction stage is divided into independent sequence segments according to spatiotemporal continuity; for each target's continuous frame data, the three types of initial features are standardized by Z-score to eliminate dimensional differences:

[0017] ;

[0018] wherein x represents an original feature value (such as wavelet energy entropy, Hu moment curvature, etc.), represents the arithmetic mean of the feature on the training set, is the standard deviation of the corresponding feature, and z is the standardized feature value, which has the physical meaning of the standard deviation multiple of the original data from the mean;

[0019] Finally, the training set, validation set and test set are divided in the ratio of 7:2:1, and the sample diversity is expanded through trajectory interpolation enhancement technology, synthetic trajectories that meet the aerodynamic constraints are inserted in the original motion path, and the generalization ability of the model to unknown maneuver modes is improved.

[0020] Preferably, in S2, the detection tracking network comprises:

[0021] The U-Net-XFeat detection framework is mainly formed by adaptively weighting fusion of a U-Net module and an XFeat module through a gated attention mechanism, and is used for extracting a dynamic heat source candidate region; and

[0022] The U-Net-XFeat detection framework is trained by introducing a multi-task joint optimization strategy, and the objective function of the multi-task joint optimization strategy is:

[0023] ;

[0024] Wherein, IoU is the intersection over union of the predicted box and the real box, represents the Euclidean distance between the center point b of the predicted box and the center point of the real box, c is the diagonal length of the minimum bounding rectangle of the predicted box and the real box, respectively represent the width of the predicted box and the real box, corresponding to the height of the other, C w , C h is the width and height of the minimum bounding rectangle;

[0025] The improved DeepSORT framework: spatiotemporal correlation of target motion trajectory, generating a unique trajectory ID of the target in consecutive frames; at the same time, recording the dynamic information such as gray scale, spatial position, speed and displacement of the target, obtaining the trajectory-level time series data that can be used for feature analysis.

[0026] Preferably, the XFeat module is connected to the output node of the third stage of the U-Net encoder in a cascading manner, shares the convolutional weights of the first three layers of the U-Net encoder, and adopts a depth separable convolution to process shared features at the detection head; and

[0027] The XFeat module descriptor D is channel spliced with the 512-dimensional semantic features Fsem of the third jump connection of the U-Net encoder, and is input into the gated attention module after being reduced to 256 dimensions by 1x1 convolution; and a spatial weight mask M∈[0,1]80x64 is generated through three 3x3 convolutions, and finally the features are fused to realize adaptive weighting.

[0028] Further preferably, the output layer of the U-Net-XFeat detection framework adopts a decoupled double-branch design: the positioning branch is reduced to 64 channels through a 3x3 convolution, and outputs 4-dimensional bounding box parameters; the classification branch adopts the same structure to output a "dynamic heat source" confidence score pobj∈[0,1]; the output layer generates an initial response heat map through Softmax activation, and after screening candidate points through non-maximum suppression (NMS), sub-pixel accuracy positioning is performed.

[0029] Preferably, in S3, the extraction includes:

[0030] The target gray sequence is decomposed into three layers using the Daubechies wavelet basis, the energy entropy and decay rate of each sub-band are calculated, the mean, variance, skewness and kurtosis are extracted, and a 4-dimensional radiation energy feature is constructed;

[0031] Based on Canny edge detection and spline fitting, a sub-pixel level contour is obtained, based on the Hu invariant moment of the contour of consecutive frames, the curvature change rate and area shrinkage gradient are calculated, and a 6-dimensional morphological evolution track is formed as a geometric contour feature;

[0032] Through Harris corner detection combined with optical flow acceleration spectrum analysis, the amplitude-frequency feature of the motion mutation point is extracted, and an 8-dimensional motion vector feature is generated.

[0033] Further preferably, the three types of feature vectors are spliced into an 18-dimensional joint feature, the first appearance of the target is taken as the time origin, and the three spline interpolation is used to uniformly resample to 40 ms intervals; the resampled features are corrected for outliers (Tukey's Fence method, IQR coefficient 1.5) and Min-Max normalized (mapped to the [-1, 1] interval); finally, a structured dataset is constructed, each record containing: metadata, feature vector, label. The dataset is stored in columnar format to optimize IO efficiency, and SHA-256 checksum is used to ensure data integrity.

[0034] Preferably, in S4, a preprocessing layer is set before the simplified recurrent unit, which is used to map and expand the 18-dimensional multi-dimensional dynamic vector to a 256-dimensional feature space through full connection.

[0035] Preferably, in S4, relative position encoding is introduced in the time attention mechanism module , so that the attention mechanism can perceive the order of frames; based on the standard self-attention score, a timestamp difference correction term is added to generate physically perceptible attention weights, enhancing the sensitivity to key stages of motion mutation and heat decay; four-head attention layers are introduced to focus on different evolution stages: heads 1 and 2 capture the change in radiation decay slope, heads 3 and 4 detect motion mutation events, and through a feedforward network and a gated linear unit, global temporal dependence features are obtained.

[0036] Preferably, the revised attention weight formula is:

[0037] ;

[0038] wherein is a query and key vector, is a scaling factor, is a final attention weight.

[0039] Preferably, in S4, the output layer is designed as a spatial confidence fusion module: the local dynamic feature is spliced with the global time sequence dependent feature :

[0040] ;

[0041] wherein, is a fused feature vector, and the GLU automatically selects the local or global information weight through the gate to realize adaptive fusion.

[0042] Preferably, a staged optimization strategy is designed to optimize the Trans-SRU cascaded time sequence discrimination network, including: freezing the T-Transformer layer, pre-training the SRU unit (learning rate 1×10⁻³) with mean square error loss; unfreezing the whole network for joint fine-tuning with binary cross entropy loss; introducing an improved Focal Loss to use adaptive class weights, dynamically calculating the balance factor according to the actual proportion of positive and negative samples in each batch, to alleviate the influence of batch distribution fluctuations and solve the sample imbalance; batch processing of continuous frame sequences, with a gradient clipping threshold of 2.0, and an early stopping strategy (patience value 20 rounds) to prevent overfitting.

[0043] Beneficial effects: the application integrates the physical evolution law of targets and decoys into the deep learning framework to form a mechanism-cognizant anti-interference tracking, effectively solving the misjudgment problem caused by radiation similarity in infrared countermeasures; the spatial detection layer adopts a U-Net-XFeat cascaded architecture to realize sub-pixel level high-precision positioning, and through the adaptive fusion of features by the gate attention, it still maintains high recall rate and boundary accuracy under the interference of occlusion and heat diffusion.

[0044] The application breaks through the limitation of single radiation sequence through a multi-modal dynamic feature fusion framework, and initiates a joint modeling paradigm of radiation energy fluctuation, geometric profile evolution and motion dynamics characteristics; through wavelet energy entropy quantization of thermal decay anisotropy, Hu moment differential capture of deformation continuous law, and optical flow acceleration spectrum revealing of maneuvering mutation characteristics, a physically interpretable discrimination basis is constructed; the feature space essentially distinguishes the thermodynamic behavior difference between targets and decoys, providing a decision basis with strong generalization ability for the classifier.

[0045] The application realizes a major breakthrough through a Trans-SRU cascade network designed by a timing discrimination layer: the parallelization preprocessing of a simplified cycle unit (SRU) greatly improves the calculation efficiency, and the time attention mechanism (T-Transformer) dynamically focuses on the key evolution stage through relative position coding; both of them fuse local dynamic features and global stage dependencies through a gating linear unit, significantly enhance the modeling accuracy of long-range radiation attenuation rules, and solve the defects of traditional cycle networks in capturing cross-cycle physical processes. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 It is a schematic diagram of the overall structure of the method of the application.

[0047] Figure 2 It is a schematic diagram of the U-Net-XFeat network of the application.

[0048] Figure 3 It is a schematic diagram of the Trans-SRU network structure of the application.

[0049] Figure 4 It is an example diagram of the network result of the application.

[0050] Figure 5 It is a PR curve diagram of the application. DETAILED DESCRIPTION

[0051] In order to make the purpose and advantages of the application clearer and more apparent, the application will be specifically described below in combination with embodiments. It should be understood that the following text is only used to describe one kind of target detection and tracking method based on multi-dimensional feature anti-infrared decoy jamming or several specific embodiments of the application, and does not strictly limit the specific protection scope requested by the application.

[0052] The technical scheme adopted by the application is as shown in Figure 1 A target detection and tracking method based on multi-dimensional feature anti-infrared decoy jamming, comprising:

[0053] S1, simulation of infrared image sequence generation: a dynamic confrontation scene is constructed through a Unity3D engine, and a Planck blackbody radiation law and Newton-Euler kinematics equation are coupled to model; a sky background is set in the simulation environment, and the motion trajectory of a jet fighter target and an MK-234 decoy bomb is dynamically generated;

[0054] In this embodiment, the target is set to perform tactical maneuvers such as snake maneuver and barrel roll, and the decoy follows a parabolic motion with random perturbations superimposed; the thermal radiation characteristics modeling introduces material emissivity ε (target: 0.85-0.93, decoy: 0.78-0.82) and surface temperature field gradient (target: 650-1100K, decoy: 800-1400K), and the atmospheric transmission attenuation and background clutter interference are simulated by the Monte Carlo method to finally generate a dynamic infrared sequence with 1280x1024 resolution and continuous frames;

[0055] The target and decoy thermal radiation regions are uniformly labeled as "dynamic heat source" category; the annotation file contains four-dimensional bounding box coordinates, kinematic parameters (three-axis velocity vector, angular acceleration) and millisecond-level timestamps;

[0056] The dataset construction stage is divided into independent sequence segments according to spatiotemporal continuity; for each target's continuous frame data, the three types of initial features are standardized by Z-score to eliminate dimensional differences:

[0057] ;

[0058] where x represents the original feature value (such as wavelet energy entropy, Hu moment curvature, etc.), represents the arithmetic mean of the feature on the training set, is the standard deviation of the corresponding feature, and z is the output normalized feature value, which has the physical meaning of the standard deviation multiple of the original data from the mean;

[0059] Finally, the training set, validation set and test set are divided in the ratio of 7:2:1, and the sample diversity is expanded through trajectory interpolation enhancement technology, which inserts synthetic trajectories that meet the aerodynamic constraints in the original motion path, improving the model's generalization ability to unknown maneuver modes.

[0060] S2, construct a detection and tracking network based on the trajectory-level time series data obtained from the dataset; the detection and tracking network includes a U-Net-XFeat detection framework and an improved DeepSORT framework;

[0061] The U-Net-XFeat detection framework is formed by adaptive weighted fusion of U-Net module and XFeat module through gated attention mechanism, and is used to extract dynamic heat source candidate regions, specifically:

[0062] Reference Figure 2As shown: the U-Net encoder based on the ResNet-34 skeleton outputs a 512-dimensional feature map with a resolution of 1 / 8 in the third stage, the lightweight XFeat matching network is connected to the output node of the third stage in a cascading manner, and the first three layers of the U-Net convolution weights (channel number [64, 128, 256]) are shared to reduce the computational load; the detection head uses a depth separable convolution to process shared features;

[0063] The feature fusion stage concatenates the XFeat descriptor D with the 512-dimensional semantic feature Fsem of the third jump connection of the U-Net (640 dimensions), and inputs it into the gate attention module after dimension reduction to 256 dimensions; this module generates a spatial weight mask M ∈ [0, 1]80×64 through three 3×3 convolutions (channel number [128, 64, 1]), and finally fuses the features to achieve adaptive weighting;

[0064] The output layer adopts a decoupled double-branch design: the positioning branch is reduced to 64 channels through a 3×3 convolution, and outputs 4-dimensional bounding box parameters; the classification branch uses the same structure to output the "dynamic heat source" confidence score pobj ∈ [0, 1]; the output layer generates an initial response heat map through Softmax activation, and then filters candidate points through non-maximum suppression (NMS) and performs sub-pixel accuracy positioning;

[0065] The training process adopts a multi-task joint optimization strategy, which combines the improved EIoU positioning loss and the binary cross-entropy classification loss through weighted fusion to form the overall objective function; this strategy enhances the regression accuracy and convergence speed of the bounding box through the EIoU loss, and promotes the cooperation between the positioning and classification tasks through joint optimization, ensuring the effectiveness of the network design and significantly improving the target capture and identification ability in complex occlusion environments. The training process jointly optimizes the improved EIoU positioning loss:

[0066] ;

[0067] where IoU is the intersection over union of the predicted box and the real box represents the Euclidean distance between the center points b and of the predicted box and the real box, c is the diagonal length of the smallest enclosing rectangle of the predicted box and the real box, represent the width of the predicted box and the real box, corresponding to the height of the other, C w , C h is the width and height of the smallest enclosing rectangle.

[0068] The time-space correlation of the target motion trajectory is realized through the improved DeepSORT framework after detection, and the core innovation lies in the construction of the Kalman filter model with Newton-Euler constraints. It is a Kalman filter predictor that explicitly incorporates acceleration into the state vector and uses the Euler kinematics formula based on the assumption of uniform acceleration to construct the state transition matrix. The target motion state is defined as an 8-dimensional vector wherein represents the center coordinates of the bounding box, is the width and height dimension, is the planar velocity component, is the acceleration component; the state transition matrix introduces the rigid body motion constraint formula:

[0069] ;

[0070] wherein, is the inter-frame time interval, is an n-dimensional identity matrix. The observation matrix is designed as , only tracking position and size changes.

[0071] In the target association stage, the adaptive Mahalanobis distance threshold is combined with the Hungarian algorithm to complete the matching, and a unique trajectory ID is assigned to each target and the motion trajectory is updated; a trajectory management module is set to synchronously record the spatial position, velocity vector, heading angle, inter-frame displacement and acceleration mutation information, and the infrared ROI, bounding box metadata and motion state of consecutive frames are stored in the ring buffer, thereby providing a time-space aligned data basis for subsequent multi-dimensional dynamic feature extraction.

[0072] S3, from the trajectory-level time sequence data, implement three-channel feature collaborative extraction and fusion, extract the radiation energy feature, geometric contour feature and motion vector feature of the target, and after aligning and normalizing the radiation energy feature, geometric contour feature and motion vector feature, fuse them into a multi-dimensional dynamic vector; including:

[0073] In the radiation energy channel, the Daubechies wavelet basis is used to perform three-layer decomposition on the target gray sequence, calculate the energy entropy and decay rate of each sub-band, extract the mean, variance, skewness and kurtosis, and construct a 4-dimensional radiation energy feature; specifically:

[0074] First, the infrared ROI region is divided into non-overlapping 16x16 blocks, and three-layer Daubechies9 wavelet packet decomposition (WPD) is performed on each block:

[0075] ;

[0076] wherein, represents the decomposition level, represents the 16 sub-band coefficients of the i,j-th layer decomposition, represents the 16 sub-band coefficients of the i,j-th layer decomposition, represents the pixel region of the i,j-th block in the t-th frame;

[0077] The energy entropy feature in the sliding window (window length 5 frames, step length 1 frame) is calculated, and finally a 4-dimensional feature vector (mean, variance, skewness, kurtosis) is generated to quantify the non-stationary characteristics of thermal radiation fluctuations; The energy entropy of each sub-band is calculated:

[0078] ;

[0079] wherein, is the energy entropy of the i,j-th layer block, is the energy entropy of the i,j-th layer block, is the energy entropy of the i,j-th layer block, is the energy of the k-th sub-band; The high-order statistics in the sliding window are counted:

[0080]

[0081] ; Mean

[0082] :

[0083] ; Variance

[0084] :

[0085] ; Skewness

[0086] :

[0087] Kurtosis

[0088] :

[0089] ; wherein, L is the length of the sliding window,

[0090] is the energy entropy of the t-th frame in the window. Geometric evolution channel, based on Canny edge detection and spline fitting, to obtain sub-pixel level contour

[0091] , based on the Hu invariant moment of the continuous frame contour, the curvature change rate and the area shrinkage gradient are calculated, forming a 6-dimensional morphological evolution track as the geometric contour feature, specifically: The 7th order Hu invariant moment is calculated for each frame contour:

[0092]

[0093] ​ ;

[0094] ;

[0095] where, is the jth Hu moment of the image, is the normalized central moment, is the central moment of the image, is the ROI barycentric coordinate, is the ROI pixel value;

[0096] Evolution trajectory is constructed by time differentiation, forming a 6-dimensional geometric evolution feature vector fgeo, which is used to describe the morphological evolution law of the target in continuous frames:

[0097] ;

[0098] Motion dynamics channel, through Harris corner detection combined with optical flow acceleration spectrum analysis, extracts amplitude-frequency features of motion mutation points, generating an 8-dimensional motion vector feature; Specifically:

[0099] Using the extracted Harris corner trajectory , the velocity and acceleration are calculated:

[0100] ;

[0101] where, is the velocity of the ith corner point in the tth frame, is the acceleration, is the inter-frame time interval; Where is the set of corner points in the tth frame, is the coordinate of the ith corner point in the tth frame, and N is the number of corner points;

[0102] Smooth the acceleration sequence by Savitzky-Golay filter (window length 7, order 3):

[0103] ;

[0104] where, is the filtered acceleration sequence, window is the filter window length, and order is the polynomial fitting order;

[0105] Perform short-time Fourier transform on the smoothed acceleration spectrum:

[0106] ;

[0107] where, is the frequency domain feature of the ith corner point at frequency f and time t, filtered acceleration for the t+nth frame, is a window function, N is the Fourier transform window length, and j is the imaginary unit;

[0108] Through frequency domain analysis, key features are extracted to form an 8-dimensional motion vector feature, which is used to describe the target maneuvering behavior and acceleration mutation law.

[0109] Three types of features are time domain alignment processing: the three types of feature vectors are spliced into 18-dimensional joint features, the first appearance of the target is taken as the time origin, and the three times spline interpolation is used to uniformly resample to 40ms interval; Outlier correction (Tukey's Fence method, IQR coefficient 1.5) and Min-Max normalization (mapped to [-1, 1] interval) are performed on the resampled features; Finally, a structured dataset is constructed, each record contains: metadata, feature vector, label. The dataset uses columnar storage to optimize IO efficiency, and SHA-256 checksum is used to ensure data integrity.

[0110] S4, a Trans-SRU cascaded time sequence discrimination network is constructed, which is used to discriminate the extracted multi-dimensional dynamic feature sequence to realize accurate classification of targets and decoys, as shown in Figure 3 , which includes:

[0111] Preprocessing layer: the obtained 18-dimensional multi-modal joint feature vector is mapped and expanded to a 256-dimensional feature space through full connection, to enhance the feature expression ability and facilitate subsequent time sequence modeling;

[0112] Simplified recurrent unit (SRU): the input is the 256-dimensional multi-modal feature vector of each target at time t, which is calculated in parallel with the hidden state at the previous time; Through the gating delay mechanism, linear transformation and nonlinear activation are separated, and cache optimization is used to realize frame-level parallel processing, so as to efficiently model local dynamic changes and output time sequence enhanced features;

[0113] Among them: the gating delay mechanism includes: the forget gate and the output gate are generated through independent full connection layers; the gating signals are calculated in parallel, and the state transmission path is separated, realizing frame-level parallel processing of 256-dimensional features, and the processing speed is significantly improved compared with the standard;

[0114] ;

[0115] Among them is the hidden state of the previous frame, is the current input feature after linear mapping, , is the gating vector, which is used to control the information retention and output ratio, The output represents the local dynamic features; ⊙ represents the Hadamard product, i.e., element-wise multiplication.

[0116] SRU achieves high-speed parallel processing at the frame level by parallel computing of gating signals and state transmission paths, while retaining local dynamic change information of the target, such as sudden acceleration, trajectory curvature or thermal decay.

[0117] The temporal attention module (T-Transformer) receives local features from the SRU output and uses it to learn long-term temporal dependencies and focus attention on key time points; specifically:

[0118] Introducing relative position encoding into self-attention computation This enables the attention mechanism to perceive the order between frames; a timestamp difference correction term is added to the standard self-attention score. This generates attention weights for physical perception, enhancing sensitivity to critical phases such as motion abrupt changes and thermal decay; a quad-head attention layer (d) is introduced. k =d v =64) Focusing on different evolutionary stages respectively: Head 1 and Head 2 capture changes in the radiation attenuation slope, Head 3 and Head 4 detect abrupt motion events;

[0119] The revised attention weight formula is as follows:

[0120] ;

[0121] in, For query and key vectors, Scaling factor For the final attention weight, This represents the time difference between frames.

[0122] T-Transformer enhances nonlinear representation capabilities through feedforward networks and gated linear units (GLUs), achieving deep fusion of local dynamic features and global dependency features;

[0123] Output layer: The local dynamic features of SRU and the global dependency features of T-Transformer are fused by gated linear units (GLU) to obtain a discrimination vector, which is then mapped to the class discrimination probability of target or decoy by the sigmoid activation function;

[0124] Among them, the output layer design includes a spatiotemporal confidence fusion module: which integrates the local dynamic features output by the SRU. Global temporal dependency features of T-Transformer Concatenate into a 320-dimensional vector:

[0125] ;

[0126] wherein, is the fused feature vector, GLU automatically selects local or global information weights through the gate to realize adaptive fusion; the fusion vector is reduced to 128 dimensions through full connection, and then adaptive weights are generated through the gate fusion unit g to generate the final fusion vector; the binary classification probability is output through Sigmoid activation, and the physical meaning thereof represents the confidence of the target at time t.

[0127] A phased optimization strategy is designed to optimize the Trans-SRU cascaded time sequence discrimination network. First, the T-Transformer layer is frozen, and the SRU unit is pre-trained with mean square error loss (learning rate 1×10⁻³); then the whole network is unfrozen, and is fine-tuned with binary cross entropy loss; the improved Focal Loss is introduced to solve the sample imbalance, and adaptive class weights are used to dynamically calculate the balance factor according to the actual proportion of positive and negative samples in each batch, so as to alleviate the influence of batch distribution fluctuation and improve the discrimination robustness of the model under extreme imbalance conditions. Batch processing is performed on continuous frame sequences, the gradient clipping threshold is set to 2.0, and the early stopping strategy (patience value 20 rounds) is used to prevent overfitting.

[0128] Reference Figure 4 As shown in the result example diagram, the complete processing flow of the target detection, tracking and discrimination in a complex infrared interference environment is shown; Figure 4 The upper left corner of the figure is the original infrared image, and multiple high-light candidate regions can be seen, including real flying targets and multiple infrared decoys released by the targets; under the characteristics of infrared imaging, the targets and decoys are all presented as high-light spots in a single frame image, and it is difficult to distinguish them only with single frame information;

[0129] Figure 4 The upper right corner of the figure is the output result of the U-Net-XFeat double-flow detection network; the network can simultaneously realize high-precision positioning and preliminary discrimination of all potential targets in the scene, and output detection frames with confidence. The detection stage fully fuses the multi-scale semantic features of U-Net and the rotation-invariant description ability of XFeat, ensures that the real targets will not be missed, and provides reliable input for subsequent tracking and discrimination. However, since the class distinction is not introduced in this stage, the decoys and the aircraft are still in the same label, and cannot be directly used for identification.

[0130] Figure 4The lower left corner figure shows the multi-target tracking result based on the DeepSORT framework; the module predicts the target motion trajectory by introducing the Kalman filter with Newton-Euler constraints, and uses adaptive Mahalanobis distance for data association, assigns a unique ID to each target and maintains identity consistency across frames. As shown in the figure, based on the detection result, a unique ID is assigned to each candidate target, and the ID is kept consistent during the cross-frame process; in this way, not only the spatial position of a single frame is obtained, but also the time sequence trajectory information across frames is extracted, effectively constructing the time sequence evolution sequence of each target, providing structured spatio-temporal data for the subsequent discrimination module.

[0131] Figure 4 The lower right corner figure is the final output of the Trans-SRU time sequence discrimination network; based on the multi-modal physical characteristics of the target, the module analyzes the radiation attenuation law and motion behavior characteristics of the target through cascaded time sequence modeling. Compared with decoys, the infrared energy of real aircraft changes more stably, and the motion trajectory also presents continuity and consistency. The model can accurately distinguish real targets from interference sources by learning these features; the target with ID = 1 in the figure is judged to be a real aircraft (green box, confidence 0.99), and the remaining decoy targets are correctly excluded, indicating that the method has excellent discrimination ability and reliability in anti-decoy interference.

[0132] In summary, the present application not only effectively improves the stability of target capture and tracking in complex scenes, but also significantly reduces the misjudgment risk caused by decoy interference through a discrimination mechanism driven by physical characteristics, and the system performance meets the needs of actual engineering applications.

[0133] S5, a comprehensive performance evaluation system is established, and core indicators such as target classification accuracy, recall rate and F1 score are calculated on the test set; the distribution characteristics of misjudgment samples are analyzed through the confusion matrix, and the robustness of the model is evaluated combined with the PR curve; the track interruption rate and identity switching frequency are synchronized to provide quantitative basis for model optimization, specifically:

[0134] The core indicators of target / decoy classification on the test set are calculated: accuracy reflects the overall discrimination ability, recall rate quantifies the real target capture rate, and F1 score comprehensively evaluates the balance between accuracy and recall rate. The distribution law of misjudgment samples is analyzed through the confusion matrix, and special attention is paid to two key error modes: misjudging high-maneuvering targets as decoys and misjudging decoys with slow thermal decay as targets, and the proportion of them in different signal-to-noise ratio intervals is counted. Robustness evaluation uses PR curve to analyze the stability of the model when the discrimination threshold is dynamically adjusted, and calculates the area under the curve (AUC) as a comprehensive generalization ability indicator.

[0135] As shown in Figure 5 , PR curves of different categories are given; from Figure 5As can be seen, the PR curve of the decoy is smoother as a whole, and the AUC reaches 0.981, indicating that the detection and discrimination performance of this type of target is the most stable; the AUC of the target is 0.952, and there is a certain fluctuation in the high recall rate area, but the overall precision remains high; the average precision (mAP@0.5) of the two types reaches 0.966, indicating that the overall discrimination model has strong robustness under different thresholds.

[0136] The application significantly improves the target capture accuracy in complex environments by the sub-pixel positioning technology of the XFeat lightweight matching network and the rotation invariant descriptor, and the U-Net deformation robust feature; innovatively constructs a three-modal feature fusion framework of radiation energy entropy, Hu moment differential and optical flow acceleration spectrum, quantifies the essential differences of targets and decoys in thermal decay anisotropy, contour evolution inertia and maneuvering mutation frequency, and forms a structured interpretable feature sequence; designs a Trans-SRU cascaded time sequence discrimination network: uses a simplified recurrent unit to model local dynamics in parallel, combines a time attention mechanism to focus on key phases, and realizes the deep fusion of local short-term changes and global long-term dependencies; while maintaining efficient processing and real-time performance, the problem of target trajectory breakage and identity switching is effectively solved, and the anti-interference technology support with physical interpretability is provided for the infrared guidance system.

[0137] The application has general migration ability and can be extended to various infrared countermeasure scenes. According to the infrared feature differences of different combat platforms, the material specific radiance parameters and kinematic constraints of the infrared physical engine are adjusted to regenerate the countermeasure sequence data; the U-Net-XFeat feature fusion mechanism and the Trans-SRU network architecture remain unchanged, and the three-modal feature weight distribution is optimized according to the thermal decay characteristics of the new target, so that an anti-interference tracking model suitable for a specific environment can be quickly constructed, which has significant military application value and equipment adaptation potential.

[0138] The embodiments of the application are described in detail above in combination with the embodiments, but the application is not limited to the above-described embodiments. For ordinary skilled persons in the art, after learning the content described in the application, without departing from the principles of the application, a number of equivalent transformations and substitutions can be made, which should also be considered as belonging to the protection scope of the application.

Claims

1. A target detection and tracking method based on multi-dimensional features to resist infrared decoy interference, characterized in that: Includes the following steps: S1. A dataset generated through simulation, the dataset including "dynamic heat source" category data and corresponding basic annotation data containing bounding boxes, millisecond-level timestamps and kinematic parameters; S2. Construct a detection and tracking network to obtain trajectory-level time-series data based on the dataset; S3. Extract the target's radiation energy features, geometric contour features, and motion vector features from the trajectory-level time series data, and then merge the radiation energy features, geometric contour features, and motion vector features into a multi-dimensional dynamic vector after alignment and normalization. S4. Construct a Trans-SRU cascaded temporal discrimination network, including: Simplify the loop unit and use a gating mechanism to model the sequence in frame-level parallelism, quickly capturing local dynamic features; The time attention mechanism module, based on the embedded relative position encoding, corrects the time difference information, enabling the multi-head self-attention to focus on the key phase inflection point of radiative decay. The output layer uses gated linear units to fuse local and global temporal dependencies to obtain a decision vector; S5. Input a multidimensional dynamic vector into the Trans-SRU cascaded temporal discriminant network, and obtain the target / decoy classification result after Sigmoid activation.

2. The target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 1, characterized in that: In S1, a dynamic confrontation sequence between the target and the decoy is generated through a simulation platform to simulate the thermal radiation characteristics and motion trajectory under complex environments. The thermal radiation characteristics modeling introduces the material emissivity and surface temperature field gradient; the target and the decoy are uniformly labeled as "dynamic heat source".

3. The target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 1, characterized in that: In S2, the detection and tracking network includes: The U-Net-XFeat detection framework is primarily formed by adaptive weighted fusion of the U-Net and XFeat modules through a gated attention mechanism. This framework is used to extract dynamic heat source candidate regions. The U-Net-XFeat detection framework introduces a multi-task joint optimization strategy for training, and the objective function of the multi-task joint optimization strategy is: ； Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. This indicates the relationship between the center point b of the predicted bounding box and the center point of the ground truth bounding box. The Euclidean distance is given by c, where c is the diagonal length of the smallest bounding rectangle between the predicted and ground truth boxes. These represent the widths of the predicted bounding box and the ground truth bounding box, respectively. Corresponding to the heights of the two, C w C h It represents the width and height of the smallest bounding rectangle; The improved DeepSORT framework: It correlates the target's motion trajectory in time and space, generating a unique trajectory ID for the target in consecutive frames; at the same time, it records dynamic information such as the target's grayscale, spatial position, velocity, and displacement, obtaining trajectory-level temporal data that can be used for feature analysis.

4. The target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 1, characterized in that: The XFeat module is cascaded to the output node of the third stage of the U-Net encoder, sharing the convolutional weights of the first three layers of the U-Net encoder, and employs depthwise separable convolution to process the shared features at the detection head; and The XFeat module descriptor D and the 512-dimensional semantic feature Fsem of the third hop connection of the U-Net encoder are concatenated by channels, and then the dimension is reduced to 256 by 1×1 convolution before being input into the gated attention module. A spatial weight mask M∈[0,1]80×64 is generated through three layers of 3×3 convolution, and finally the features are fused to achieve adaptive weighting.

5. The target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 4, characterized in that: In the improved DeepSORT framework, acceleration is explicitly incorporated into the state vector, and a Kalman filter predictor for the state transition matrix is ​​constructed using Euler's kinematic formula based on the uniform acceleration assumption; the state transition matrix incorporates rigid body motion constraint formulas. ； in, This is the inter-frame time interval. It is an n-dimensional identity matrix; the observation matrix is ​​designed as follows: It only tracks changes in location and scale; and The spatiotemporal association uses an adaptive Mahalanobis distance threshold combined with a Hungarian algorithm to complete the matching, assigning a unique trajectory ID to each target and updating the motion trajectory.

6. The target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 1, characterized in that: Extracting from S3 includes: The target grayscale sequence is decomposed into three levels using the Daubechies wavelet basis. The energy entropy and decay rate of each sub-band are calculated, and the mean, variance, skewness and kurtosis are extracted to form a 4-dimensional radiative energy feature. Based on Canny edge detection and spline fitting, sub-pixel level contours are obtained. Based on the Hu invariant moments of the continuous frame contours, the rate of curvature change and area shrinkage gradient are calculated to form a 6-dimensional morphological evolution trajectory as geometric contour features. By combining Harris corner detection with optical flow acceleration spectrum analysis, the amplitude-frequency features of motion abrupt change points are extracted to generate 8-dimensional motion vector features.

7. A target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 6, characterized in that: In S4: A preprocessing layer is set before the simplified loop unit, the preprocessing layer is used to extend the 18-dimensional multidimensional dynamic vector to a 256-dimensional feature space through a fully connected mapping; Introducing relative position encoding into the time attention mechanism module This enables the attention mechanism to perceive the order between frames; a timestamp difference correction term is added to the standard self-attention score. The system generates attention weights for physical perception, enhancing sensitivity to key stages of motion mutation and thermal decay. It introduces a four-head attention layer to focus on different evolutionary stages: head 1 and head 2 capture changes in the radiation decay slope, and head 3 and head 4 detect motion mutation events. Through a feedforward network and gated linear units, global temporal dependency features are obtained.

8. The target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 7, characterized in that: The revised attention weight formula is as follows: ； in For query and key vectors, Scaling factor This represents the final attention weight.

9. A target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 7, characterized in that: In S4, the output layer design incorporates a spatiotemporal confidence fusion module: integrating local dynamic features. Global temporal dependency features splicing: ； in, For the fused feature vector, GLU automatically selects local or global information weights through gating to achieve adaptive fusion.

10. A target detection and tracking method based on multi-dimensional features to resist infrared decoy interference according to claim 1, characterized in that: A phased optimization strategy is designed to optimize the Trans-SRU cascaded temporal discriminant network, including: The frozen-time attention mechanism module uses mean squared error loss to pre-train simplified recurrent units; Unfreeze the entire network and fine-tune it jointly using binary cross-entropy loss; An improved Focal Loss is introduced, which uses adaptive class weights to dynamically calculate the balance factor based on the actual proportion of positive and negative samples in each batch, thereby mitigating the impact of distribution fluctuations between batches and solving sample imbalance.

Images