A target detection and tracking method based on multi-dimensional feature anti-infrared decoy jamming

Abstract

The present application relates to a kind of target detection tracking methods based on multi-dimensional feature anti-infrared decoy interference, it is related to computer vision and target tracking technical field, including the data set generated by simulation, construct detection tracking network and obtain trajectory level timing data;Radiant energy feature, geometric profile feature and motion vector feature of target are extracted from trajectory level timing data, and are fused into multidimensional dynamic vector;Trans-SRU cascaded timing discrimination network is constructed, multidimensional dynamic vector is input, and classification result of target and decoy is output by combining parallelization feature processing and time attention mechanism focusing key radiation attenuation and maneuvering change stage;The defect that traditional cyclic network is insufficient to capture across cycle physical process is solved, and robust anti-interference capability is provided for infrared guidance system in complex environment.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target tracking technology, specifically to a target detection and tracking method based on multi-dimensional features to resist infrared decoy interference. Background Technology

[0002] With the continuous upgrading of infrared-guided weapons, infrared decoy flares, as a key countermeasure, directly determine the survivability of combat platforms through their jamming effectiveness. Decoy flares induce target identification deviations in infrared seekers by precisely simulating the thermal radiation characteristics and trajectory of target aircraft, forming a tactical-level defensive barrier. However, in real combat environments, the intertwining trajectories of targets and multiple decoys, the convergence of radiation characteristics, and complex background interference lead to severe challenges for traditional tracking systems, such as a surge in misjudgment rates and trajectory breakage. Therefore, there is an urgent need to develop intelligent anti-jamming technologies with the ability to understand physical laws.

[0003] Traditional anti-jamming methods are mainly based on motion feature analysis or radiation feature recognition. The former uses Kalman or particle filtering to discriminate based on motion differences, but it is prone to failure when the target is moving at low speed or when decoys are densely packed. The latter classifies based on instantaneous infrared statistics, but ignores the time-varying evolution of thermal radiation, and has insufficient modeling of differences in thermal decay rates, resulting in low robustness. In recent years, deep learning such as the YOLO series detectors have performed efficiently in localization tasks, but they do not utilize the radiation homology between targets and decoys and lack the ability to model temporal evolution, leading to a high false detection rate. Tracking algorithms such as DeepSORT rely on appearance matching, resulting in frequent identity switching when textures converge.

[0004] To overcome these bottlenecks, cutting-edge research has begun to explore multimodal fusion solutions: introducing attention mechanisms in the spatial dimension to enhance feature discrimination, such as the Transformer module improving the robustness of occluded scene detection through global dependency modeling; and using bidirectional recurrent networks (such as Bi-LSTM) in the temporal dimension to analyze radiation energy curves in an attempt to capture differences in thermal decay trends. However, existing solutions still have shortcomings: spatial detection networks are insufficient in extracting target geometric deformation features, resulting in limited scale adaptability; feature extraction relies excessively on a single radiation channel; and temporal modeling networks have weak learning ability for long-range physical evolution laws, making it difficult to meet the accurate discrimination requirements of highly dynamic environments. Summary of the Invention

[0005] The purpose of this invention is to provide a target detection and tracking method based on multi-dimensional features to resist infrared decoy interference, which effectively solves the shortcomings of traditional cyclic networks in modeling cross-period evolution laws.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a target detection and tracking method based on multi-dimensional features to resist infrared decoy interference, comprising the following steps:

[0007] 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;

[0008] S2. Construct a detection and tracking network to obtain trajectory-level time-series data based on the dataset;

[0009] 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.

[0010] S4. Construct a Trans-SRU cascaded temporal discrimination network, including:

[0011] Simplify the loop unit and use a gating mechanism to model the sequence in frame-level parallelism, quickly capturing local dynamic features;

[0012] 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.

[0013] The output layer uses gated linear units to fuse local and global temporal dependencies to obtain a decision vector;

[0014] S5. Input a multidimensional dynamic vector into the Trans-SRU cascaded temporal discriminant network, and obtain the target / decoy classification result after Sigmoid activation.

[0015] Preferably, 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 conditions. 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".

[0016] More preferably, during the dataset construction phase, the dataset is divided into independent sequence segments based on spatiotemporal continuity; for each target's continuous frame data, its three initial features are standardized using Z-scores to eliminate dimensional differences.

[0017] ;

[0018] Where x represents the original feature value (such as wavelet energy entropy, Hu moment curvature, etc.). This represents the arithmetic mean of the feature on the training set. Let z be the standard deviation of the corresponding feature, and z be the standardized feature value of the output. Its physical meaning is the standard deviation multiple of the original data from the mean.

[0019] Finally, the training set, validation set, and test set were divided into three sets in a 7:2:1 ratio. The sample diversity was expanded by using trajectory interpolation enhancement technology. Synthetic trajectories that conform to aerodynamic constraints were inserted into the original motion path to improve the model's generalization ability to unknown maneuvering modes.

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

[0021] 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.

[0022] 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:

[0023] ;

[0024] 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. C represents the height of the predicted bounding box and the ground truth bounding box, respectively. w C h It represents the width and height of the smallest bounding rectangle;

[0025] 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.

[0026] Preferably, the XFeat module is cascaded to the output node of the third stage of the U-Net encoder, shares the convolutional weights of the first three layers of the U-Net encoder, and uses depthwise separable convolution to process the shared features at the detection head; and

[0027] 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, reduced to 256 dimensions by 1×1 convolution, and then input into the gated attention module; a spatial weight mask M∈[0,1] is generated by three layers of 3×3 convolution. 80×64 Ultimately, features are merged to achieve adaptive weighting.

[0028] More preferably, the output layer of the U-Net-XFeat detection framework adopts a decoupled dual-branch design: the localization branch reduces the dimensionality to 64 channels through 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 heatmap through Softmax activation, and performs sub-pixel precision localization after screening candidate points through non-maximum suppression (NMS).

[0029] Preferably, the extraction in S3 includes:

[0030] 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.

[0031] 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.

[0032] 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.

[0033] More preferably, the three types of feature vectors are concatenated into an 18-dimensional joint feature. Using the first appearance of the target as the time origin, cubic spline interpolation is used to uniformly resample to a 40ms interval. Outlier correction (Tukey's Fence method, IQR coefficient 1.5) and Min-Max normalization (mapping to the [-1,1] interval) are applied to the resampled features. Finally, a structured dataset is constructed, with each record containing: metadata, feature vector, and label. The dataset uses columnar storage to optimize IO efficiency and SHA-256 verification to ensure data integrity.

[0034] Preferably, in S4, a preprocessing layer is provided before the simplified loop unit, the preprocessing layer being used to extend the 18-dimensional multidimensional dynamic vector to a 256-dimensional feature space through a fully connected mapping.

[0035] Preferably, in S4, relative position encoding is introduced into the temporal 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.

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

[0037] ;

[0038] in For query and key vectors, Scaling factor This represents the final attention weight.

[0039] Preferably, in S4, the output layer design includes a spatiotemporal confidence fusion module: which integrates local dynamic features. Global temporal dependency features splicing:

[0040] ;

[0041] in, For the fused feature vector, GLU automatically selects local or global information weights through gating to achieve adaptive fusion.

[0042] Preferably, a phased optimization strategy is designed to optimize the Trans-SRU cascaded temporal discriminant network, including: freezing the T-Transformer layers and pre-training SRU units with mean squared error loss (learning rate 1×10⁻³); unfreezing the entire network and jointly fine-tuning it with binary cross-entropy loss; introducing an improved Focal Loss with adaptive class weights, dynamically calculating the balance factor based on the actual proportion of positive and negative samples in each batch to mitigate the impact of distribution fluctuations between batches and solve sample imbalance; batch processing of continuous frame sequences, setting the gradient clipping threshold to 2.0, and preventing overfitting through an early stopping strategy (patience value of 20 rounds).

[0043] Beneficial effects: This invention integrates the physical evolution of targets and decoys into a deep learning framework to form an anti-interference tracking system with mechanistic understanding, effectively solving the problem of misjudgment caused by radiation similarity in infrared countermeasures; the spatial detection layer adopts a U-Net-XFeat cascade architecture to achieve sub-pixel-level high-precision positioning, and maintains high recall and boundary accuracy under occlusion and thermal diffusion interference through gated attention adaptive feature fusion.

[0044] This invention breaks through the limitations of a single radiation sequence by using a multimodal dynamic feature fusion framework, pioneering a joint modeling paradigm for radiation energy fluctuations, geometric contour evolution, and motion dynamics characteristics. It quantifies thermal attenuation anisotropy through wavelet energy entropy, captures the continuous deformation law through Hu moment differentiation, and reveals the abrupt changes in maneuvering characteristics through optical flow acceleration spectrum, constructing a physically interpretable discrimination criterion. This feature space essentially distinguishes the differences in thermodynamic behavior between targets and decoys, providing a decision basis with strong generalization ability for the classifier.

[0045] This invention achieves a major breakthrough through a Trans-SRU cascaded network designed with a time-series discriminant layer: the parallel preprocessing of the simplified recurrent unit (SRU) significantly improves computational efficiency, while the time attention mechanism (T-Transformer) dynamically focuses on key evolution stages through relative position encoding; both enhance the modeling accuracy of long-range radiation decay laws by fusing local dynamic features and global stage dependencies through gated linear units, thus solving the deficiency of traditional recurrent networks in capturing cross-period physical processes. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall structure of the method of the present invention.

[0047] Figure 2 This is a schematic diagram of the U-Net-XFeat network of the present invention.

[0048] Figure 3 This is a schematic diagram of the Trans-SRU network structure of the present invention.

[0049] Figure 4 This is an example diagram of the network results of the present invention.

[0050] Figure 5 This is a PR curve diagram of the present invention. Detailed Implementation

[0051] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe a target detection and tracking method based on multi-dimensional features to resist infrared decoy interference, or several specific implementation methods of this invention, and does not strictly limit the scope of protection specifically claimed by this invention.

[0052] The technical solution adopted in this invention is as follows: Figure 1 As shown, a target detection and tracking method based on multi-dimensional features to resist infrared decoy interference includes:

[0053] S1. Simulated Infrared Image Sequence Generation: A dynamic combat scenario is constructed using the Unity3D engine, and Planck's blackbody radiation law and Newton-Euler kinematic equations are coupled for modeling; a sky background is set in the simulation environment, and the movement trajectories of jet fighter targets and MK-234 decoy missiles are dynamically generated;

[0054] In this embodiment, the target is set to perform tactical maneuvers such as serpentine maneuvers and barrel rolls, while the decoy flares follow parabolic motion and are superimposed with random perturbations. The thermal radiation characteristics modeling introduces the material emissivity ε (target: 0.85-0.93, decoy: 0.78-0.82) and the surface temperature field gradient (target: 650-1100K, decoy: 800-1400K). The Monte Carlo method is used to simulate atmospheric transmission attenuation and background clutter interference, and finally a dynamic infrared sequence with a resolution of 1280×1024 and continuous frames is generated.

[0055] The thermal radiation areas of both the target and the decoy are uniformly labeled as "dynamic heat source"; the labeling file includes four-dimensional bounding box coordinates, kinematic parameters (three-axis velocity vectors, angular acceleration) and millisecond-level timestamps;

[0056] During the dataset construction phase, the dataset is divided into independent sequence segments based on spatiotemporal continuity. For each target's continuous frame data, its three initial features are standardized using Z-scores to eliminate dimensional differences.

[0057] ;

[0058] Where x represents the original feature value (such as wavelet energy entropy, Hu moment curvature, etc.). This represents the arithmetic mean of the feature on the training set. Let z be the standard deviation of the corresponding feature, and z be the standardized feature value of the output. Its physical meaning is the standard deviation multiple of the original data from the mean.

[0059] Finally, the training set, validation set, and test set were divided into three sets in a 7:2:1 ratio. The sample diversity was expanded by using trajectory interpolation enhancement technology. Synthetic trajectories that conform to aerodynamic constraints were inserted into the original motion path to improve the model's generalization ability to unknown maneuvering modes.

[0060] S2. Construct a detection and tracking network to obtain trajectory-level time-series data based on the dataset; the detection and tracking network includes the U-Net-XFeat detection framework and the improved DeepSORT framework;

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

[0062] refer to Figure 2As shown: The U-Net encoder based on the ResNet-34 skeleton outputs a 512-dimensional feature map at 1 / 8 resolution in the third stage. The lightweight XFeat matching network is connected to the output node of the third stage in a cascaded manner, sharing the convolutional weights of the first three layers of U-Net (number of channels [64,128,256]), reducing the computational load through parameter sharing; the detection head uses depthwise separable convolution to process the shared features.

[0063] In the feature fusion stage, the XFeat descriptor D and the 512-dimensional semantic feature Fsem of the third hop connection of U-Net are concatenated (640-dimensional), and then the dimensionality is reduced to 256-dimensional by 1×1 convolution before being input into the gated attention module. This module generates a spatial weight mask M∈[0,1] through three layers of 3×3 convolutions (number of channels [128,64,1]). 80×64 Ultimately, features are fused to achieve adaptive weighting;

[0064] The output layer adopts a decoupled dual-branch design: the localization branch reduces the dimension to 64 channels through 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 heatmap through Softmax activation, and performs sub-pixel precision localization after screening candidate points through non-maximum suppression (NMS).

[0065] The training process employs a multi-task joint optimization strategy, using a weighted fusion of improved EIoU localization loss and binary cross-entropy classification loss to form the overall objective function. This strategy enhances the regression accuracy and convergence speed of bounding boxes through EIoU loss, and promotes synergy between localization and classification tasks through joint optimization, ensuring the effectiveness of the network design and ultimately significantly improving the target capture and discrimination capabilities in complex occlusion environments. The improved EIoU localization loss is obtained through joint optimization during training.

[0066] ;

[0067] Wherein, IoU is the intersection-over-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. C represents the height of the predicted bounding box and the ground truth bounding box, respectively. w C h It represents the width and height of the smallest bounding rectangle.

[0068] After detection, the spatiotemporal correlation of the target's motion trajectory is achieved through an improved DeepSORT framework. The core innovation lies in constructing a Newton-Euler-constrained Kalman filter model, which is a Kalman filter predictor that explicitly incorporates acceleration into the state vector and uses Euler's kinematic 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. ,in Indicates the coordinates of the bounding box center. These are the width and height dimensions. It is a planar velocity component. For acceleration components; the state transition matrix introduces rigid body motion constraint formulas:

[0069] ;

[0070] 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.

[0071] In the target association stage, an adaptive Mahalanobis distance threshold combined with the Hungarian algorithm is used to complete the matching, assign a unique trajectory ID to each target and update the motion trajectory; a trajectory management module is set up to synchronously record spatial position, velocity vector, heading angle, inter-frame displacement and acceleration change information, and store the infrared ROI, bounding box metadata and motion state of consecutive frames in a circular buffer, providing a spatiotemporally aligned data foundation for subsequent multidimensional dynamic feature extraction.

[0072] S3. From trajectory-level time-series data, perform collaborative extraction and fusion of three-channel features to extract the target's radiation energy features, geometric contour features, and motion vector features. After aligning and normalizing the radiation energy features, geometric contour features, and motion vector features, they are fused into a multi-dimensional dynamic vector; including:

[0073] In the radiant energy channel, 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 radiant energy feature. Specifically:

[0074] First, the infrared ROI region is analyzed. Perform non-overlapping 16×16 block decomposition, and perform three-level Daubechies9 wavelet packet decomposition (WPD) on each block:

[0075] ;

[0076] in, Indicates the number of decomposition layers. Indicates the first The 16 sub-band coefficients of the layer decomposition This represents the pixel region of the i-th,j-th block in the t-th frame;

[0077] Calculate the energy entropy features within a sliding window (window length 5 frames, step size 1 frame), ultimately generating a 4-dimensional feature vector (mean, variance, skewness, kurtosis) to quantify the non-stationary characteristics of thermal radiation fluctuations; calculate the energy entropy of each sub-band:

[0078] ;

[0079] in, For the first Energy entropy of layered blocks For the first Energy entropy of layered blocks The energy of the k-th subband;

[0080] Calculate higher-order statistics within a sliding window:

[0081] ;

[0082] mean :

[0083] ;

[0084] variance :

[0085] ;

[0086] Skewness :

[0087]

[0088] Kudo :

[0089] ;

[0090] Where L is the length of the sliding window. Let be the energy entropy of the t-th frame within the window.

[0091] Geometric evolution channel, based on Canny edge detection and spline fitting, obtains sub-pixel level contours. 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, which serves as the geometric contour feature, specifically:

[0092] Calculate the 7th order Hu invariant moments for each frame's contour:

[0093] ;

[0094] ;

[0095] in, For the j-th Hu invariant moment, To normalize the central moments, The central moment of the image, The coordinates of the ROI centroid ROI pixel value;

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

[0097] ;

[0098] The motion dynamics channel extracts the amplitude-frequency features of abrupt motion points through Harris corner detection combined with optical flow acceleration spectrum analysis, generating 8-dimensional motion vector features; specifically:

[0099] Using the extracted Harris corner trajectory Calculate velocity and acceleration:

[0100] ;

[0101] in, Let be the velocity of the i-th corner point in frame t. For acceleration, The time interval between frames; where Let be the set of corner points in frame t. Let N be the coordinates of the i-th corner point in frame t, and N be the number of corner points.

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

[0103] ;

[0104] in, The sequence is the filtered acceleration sequence, window is the length of the filtering window, and order is the order of the polynomial fitting.

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

[0106] ;

[0107] in, Let f be the frequency domain characteristic of the i-th corner point at frequency f and time t. The acceleration after filtering in the (t+n)th frame. Here, N is the window function, N is the Fourier transform window length, and j is the imaginary unit;

[0108] Key features are extracted through frequency domain analysis to form an 8-dimensional motion vector feature, which is used to describe the target's maneuvering behavior and acceleration abrupt change patterns.

[0109] The three types of features are aligned in the temporal domain: the three feature vectors are concatenated into an 18-dimensional joint feature, with the first appearance of the target as the time origin, and uniformly resampled to a 40ms interval through cubic spline interpolation; outlier correction (Tukey's Fence method, IQR coefficient 1.5) and Min-Max normalization (mapped to the [-1,1] interval) are performed on the resampled features; finally, a structured dataset is constructed, with each record containing: metadata, feature vector, and label. The dataset uses columnar storage to optimize IO efficiency and SHA-256 verification to ensure data integrity.

[0110] S4. Construct a Trans-SRU cascaded temporal discriminant network to discriminate the extracted multidimensional dynamic feature sequences, thereby achieving accurate classification of targets and decoys. (Refer to...) Figure 3 As shown, it includes:

[0111] Preprocessing layer: The obtained 18-dimensional multimodal joint feature vector is extended to a 256-dimensional feature space through a fully connected mapping to enhance the feature expressive power and facilitate subsequent temporal modeling;

[0112] Simplified Recurrent Unit (SRU): The input is a 256-dimensional multimodal feature vector of each target at time t, which is computed in parallel with the hidden state of the previous time step; the linear transformation and nonlinear activation are separated by a gating delay mechanism, and frame-level parallel processing is achieved by using high-speed cache optimization, thereby efficiently modeling local dynamic changes and outputting temporal enhancement features;

[0113] Among them, the gating delay mechanism includes: forget gate. and output gate It is generated through an independent fully connected layer; by parallelizing the calculation of the gated signal and separating the state transmission path, it achieves frame-level parallel processing of 256-dimensional features, and the processing speed is significantly improved compared to the standard.

[0114] ;

[0115] in This is the hidden state from the previous frame. The current input features are obtained after linear mapping. , This is a gating vector 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] in, The fused feature vector is obtained by GLU automatically selecting local or global information weights through gating to achieve adaptive fusion. After the fused vector is reduced to 128 dimensions by a fully connected layer, adaptive weights are generated by the gated fusion unit g, resulting in the final fused vector. The binary classification probability is output by sigmoid activation, and its physical meaning represents the confidence of the target at time t.

[0127] A phased optimization strategy was designed to optimize the Trans-SRU cascaded temporal discriminant network. First, the T-Transformer layers were frozen, and SRU units were pre-trained using mean squared error loss (learning rate 1×10⁻³). Then, the entire network was unfrozen, and joint fine-tuning was performed using binary cross-entropy loss. An improved Focal Loss was introduced to address sample imbalance, employing adaptive class weights. The balance factor was dynamically calculated based on the actual ratio of positive to negative samples in each batch, mitigating the impact of batch-to-batch distribution fluctuations and improving the model's robustness under extreme imbalance conditions. Continuous frame sequences were batch-processed, with a gradient clipping threshold set to 2.0. An early stopping strategy (with a patience value of 20 rounds) was used to prevent overfitting.

[0128] refer to Figure 4 The figure shown is an example of the results, illustrating the complete processing flow of the method of the present invention for the detection, tracking and discrimination of targets in a complex infrared interference environment; Figure 4 The top left image is the original infrared image, which shows multiple bright candidate areas, including the real flying target and multiple infrared decoys released by it. Under the characteristics of infrared imaging, the target and the decoys appear as bright patches in a single frame image, making it difficult to distinguish them based on a single frame of information.

[0129] Figure 4 The top-right image shows the output of the U-Net-XFeat dual-stream detection network. This network can simultaneously achieve high-precision localization and preliminary identification of all potential targets in the scene, outputting detection boxes with confidence scores. The detection stage fully integrates the multi-scale semantic features of U-Net and the rotation-invariant descriptive capabilities of XFeat, ensuring that no real targets are missed and providing reliable input for subsequent tracking and discrimination. However, since no category distinction is introduced in this stage, decoys and aircraft are still under the same label and cannot be directly used for identification.

[0130] Figure 4The lower left figure shows the multi-target tracking results based on the DeepSORT framework. This module predicts target trajectories using Kalman filtering with Newton-Euler constraints and employs adaptive Mahalanobis distance for data association, assigning a unique ID to each target and maintaining consistency across frames. As shown in the figure, a unique ID is assigned to each candidate target based on the detection results, and the ID remains consistent across frames. In this way, not only is the spatial position of a single frame obtained, but the temporal trajectory information across frames can also be extracted, effectively constructing the temporal evolution sequence of each target and providing structured spatiotemporal data for the subsequent discrimination module.

[0131] Figure 4 The bottom right figure shows the final output of the Trans-SRU temporal discriminant network. This module, based on the multimodal physical characteristics of the target, analyzes the target's radiation attenuation law and motion behavior characteristics through cascaded temporal modeling. Compared with decoys, the infrared energy changes of real aircraft are more stable, and their motion trajectories also exhibit continuity and consistency. By learning these characteristics, the model can accurately distinguish between real targets and interference sources. In the figure, the target with ID=1 is identified as a real aircraft (green box, confidence level 0.99), and all other decoy targets are correctly excluded, indicating that the method of this invention has excellent discrimination ability and reliability in resisting decoy interference.

[0132] In summary, this invention not only effectively improves the stability of target acquisition and tracking in complex scenarios, but also significantly reduces the risk of misjudgment caused by decoy interference through a physical feature-driven discrimination mechanism, and the system performance meets the needs of practical engineering applications.

[0133] S5. Establish a comprehensive performance evaluation system, calculate core indicators such as target classification accuracy, recall, and F1 score on the test set; analyze the distribution characteristics of misclassified samples through confusion matrix analysis, and evaluate the robustness of the model in conjunction with PR curves; simultaneously track trajectory interruption rate and identity switching frequency to provide quantitative basis for model optimization, specifically:

[0134] The core metrics for target / decoy classification are calculated on the test set: precision reflects overall discrimination ability, recall quantifies the capture rate of true targets, and F1 score comprehensively evaluates the balance between precision and recall. The distribution patterns of misclassified samples are analyzed using a confusion matrix, with particular attention paid to two key error patterns—misclassifying highly maneuverable targets as decoys and misclassifying decoys with slow thermal decay as targets—and their proportions are statistically analyzed across different signal-to-noise ratio ranges. Robustness is evaluated using the PR curve to analyze the model's stability when the discrimination threshold is dynamically adjusted, and the area under the curve (AUC) is calculated as a comprehensive generalization ability indicator.

[0135] like Figure 5 As shown, PR curves for different categories are presented; from Figure 5As can be seen, the PR curve of the decoy is smoother overall, with an AUC of 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, which fluctuates in the high recall region, but maintains a high precision overall. The average precision (mAP@0.5) of the two classes reaches 0.966, indicating that the overall discrimination model has strong robustness under different thresholds.

[0136] This invention significantly improves target acquisition accuracy in complex environments by combining the sub-pixel localization technology of the XFeat lightweight matching network with rotation-invariant descriptors and the deformation robustness features of U-Net. It innovatively constructs a three-modal feature fusion framework based on radiative energy entropy, Hu moment differential, and optical flow acceleration spectrum to quantify the essential differences between targets and decoys in thermal decay anisotropy, contour evolution inertia, and maneuver mutation frequency, forming a structured and interpretable feature sequence. A Trans-SRU cascaded temporal discrimination network is designed: it utilizes simplified recurrent units to model local dynamics in parallel and combines a temporal attention mechanism to focus on key phases, achieving a deep fusion of local short-term changes and global long-term dependencies. While maintaining high efficiency and real-time performance, it effectively solves the problems of target trajectory breakage and identity switching, providing a physically interpretable anti-interference technology guarantee for infrared guidance systems.

[0137] This invention possesses universal transfer capabilities and can be extended to various infrared countermeasure scenarios. Addressing the differences in infrared characteristics among different combat platforms, adversarial sequence data is regenerated by adjusting the material emissivity parameters and kinematic constraints of the infrared physics engine. Maintaining the U-Net-XFeat feature fusion mechanism and Trans-SRU network architecture, and optimizing the trimodal feature weight allocation based on the thermal attenuation characteristics of novel targets, an anti-jamming tracking model adapted to specific environments can be quickly constructed, demonstrating significant military application value and equipment adaptability potential.

[0138] The embodiments of the present invention have been described in detail above with reference to the examples. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

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; Both targets and decoys are categorized under "Dynamic Heat Source". 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 simplified recurrent unit input is a 256-dimensional multimodal feature vector of each target at time t, which is computed in parallel with the hidden state of the previous time step; the linear transformation and nonlinear activation are separated by a gating mechanism, and frame-level parallel processing is achieved by using cache optimization; 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; The output layer fuses the local dynamic features of the simplified recurrent unit with the global dependency features of the temporal attention mechanism module using a gated linear unit to obtain a discrimination vector, which is then mapped to the class discrimination probability of the target or decoy through a sigmoid activation function. S5. Input the multidimensional dynamic vector into the Trans-SRU cascaded temporal discriminant network to obtain the target / decoy category discrimination result.

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.

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. C represents the height of the predicted bounding box and the ground truth bounding box, respectively. 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 including the target's grayscale, spatial position, velocity and displacement, to obtain 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 3, 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] is generated by three layers of 3×3 convolution. 80×64 Ultimately, features are merged 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. The 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