Infrared unmanned aerial vehicle tracking method and system based on space-frequency cooperation and spectrum focusing

Abstract

The application provides an infrared unmanned aerial vehicle tracking method and system based on space-frequency cooperation and spectrum focusing. The method first acquires an infrared video stream, generates a two-dimensional space target prior heat map by using a bidirectional weighted feature pyramid embedded with coordinate attention. Then, the feature map and the heat map are mapped to the frequency domain, the space prior dynamic modulation frequency domain gating weight is used to reconstruct the space-frequency cooperative feature map that suppresses the background thermal clutter, and the high-frequency energy proportion factor is extracted. Finally, in the prediction stage, the factor is used to construct a non-monotonic focusing coefficient, the boundary box regression loss is adaptively corrected, and the low-quality anchor box gradient is dynamically suppressed. The application gives the system environmental adaptability through the spectrum energy feedback mechanism, effectively solves the scale collapse and trajectory divergence problems of the target in the thermal cross blind area or low signal-to-noise ratio, and realizes high-precision tracking.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and photoelectric target detection technology, and in particular to an infrared UAV tracking method and system based on space-frequency coordination and spectral focusing. Background Technology

[0002] With the exponential growth of unmanned aerial vehicles (UAVs), the asymmetric security threats posed by these "low, slow, and small" drones are becoming increasingly severe, making the construction of all-weather anti-drone tracking systems a major need. In traditional air-to-air detection, radar is easily overwhelmed by ground clutter, and visible light cameras are limited by ambient light. Infrared thermal imaging technology, with its advantages of all-weather passive detection and strong resistance to light interference, has become the mainstream technology.

[0003] In recent years, Siamese Networks in deep learning have been widely used in single-object tracking, such as Siamese Fully Convolutional Network (SiamFC), Siamese Region Proposal Network++ (SiamRPN++), and the Siamese Bisemantic Tracking Network (SiamDT) algorithm specifically for drone scenarios. In addition, Transformer model architectures (such as the tracking algorithm TransT and the tracking algorithm SwinTrack) have also been introduced into tracking tasks.

[0004] Existing depth tracking algorithms are mostly designed for general visible light targets, and when directly transferred to infrared target tracking, they have the following significant drawbacks: Feature diffusion and spatial prior loss: Infrared micro-targets typically occupy only a few pixels and lack texture. After multiple layers of downsampling, existing deep convolutional networks or conventional transformers can easily smooth or even erase the weak spatial features of the target, leading to severe feature diffusion.

[0005] Frequency domain filter failure caused by "thermal crossover": When the temperature of the environmental background (such as clouds, buildings) and the UAV target are similar, a severe thermal crossover phenomenon occurs, causing a sharp drop in spatial contrast. Some existing technologies attempt to introduce simple frequency domain filtering to separate high and low frequencies, but due to the lack of precise spatial coordinates of the target, the frequency domain filter is prone to blind searching at low signal-to-noise ratios, confusing the high-frequency edges of the real target with background thermal noise, leading to filter failure.

[0006] Traditional loss functions are ill-suited to thermal diffusion: Existing dynamic focusing loss functions calculate outliers entirely based on the intersection-union ratio (IUU) in the spatial domain. When severe thermal crossover leads to extreme blurring of the target boundary, the geometric error between the predicted and ground truth boxes is amplified abnormally, causing severe gradient oscillations during backpropagation and ultimately leading to irreversible drift of the model's tracking trajectory.

[0007] This invention aims to break the existing technology's disconnect between spatial domain extraction and frequency domain filtering, and to solve the technical problems of feature extraction failure of infrared micro UAVs under thermal crosstalk and strong background clutter interference, as well as the regression loss function being easily misled by low-quality blurred anchor boxes. Summary of the Invention

[0008] Purpose of the invention: The technical problem to be solved by the present invention is to provide an infrared UAV tracking method and system based on space-frequency coordination and spectrum focusing, which addresses the shortcomings of the existing technology.

[0009] The method is applied to a twin network architecture and includes the following steps: Step S1, Multi-scale spatial feature and target coordinate prior extraction: Acquire infrared video stream, extract the initial frame as template frame, and extract the current frame to be detected as search frame; input the template frame and search frame synchronously into the backbone feature extraction network with shared weights to obtain hierarchical original features; then, input the original features into a bidirectional weighted feature pyramid for multi-scale feature fusion. In the lateral connection stage of feature fusion, the embedded coordinate attention mechanism is used to decouple and encode the sub-pixel level coordinate responses in the horizontal and vertical directions to generate a two-dimensional spatial target prior heatmap and output the fused multi-scale spatial feature map; Step S2, Space-guided frequency-domain joint gating filtering and energy assessment: Two-dimensional discrete Fourier transforms are performed on the channel data of the multi-scale spatial feature map and the prior heat map of the spatial target, respectively, to map the spatial signal to the frequency domain; in the frequency domain feature space, the spectral response of the prior heat map of the spatial target is used as a prior structural constraint to dynamically and adaptively modulate the learnable complex weight matrix in the frequency domain gating module, and space-prior-guided joint gating filtering is performed; subsequently, a space-frequency cooperative feature map that suppresses background thermal clutter and preserves the high-frequency edge of the target is reconstructed through two-dimensional discrete Fourier inverse transform, and the high-frequency energy proportion factor of the current filtered feature is extracted simultaneously; Step S3, Non-monotonic Dynamic Focusing Regression and Position Prediction Based on Thermal Cross-Sensing: The spatial-frequency collaborative feature map is input into a multi-task prediction head containing classification and regression branches; during the network parameter training and update phase of the regression branch, based on the high-frequency energy proportion factor, the basic outlier parameter in the bounding box regression loss function is adaptively penalized and corrected to construct a non-monotonic focusing coefficient based on spectral energy sensing; the non-monotonic focusing coefficient is used to dynamically suppress the gradient of low-quality anchor boxes to complete network optimization, and the final predicted bounding box and motion trajectory of the infrared UAV target are output during the inference phase.

[0010] Step S1 includes: Step S1-1, Direction-aware pooling encoding: Obtain the input feature map at the lateral connections of the bidirectional weighted feature pyramid. ,in For the number of channels, For height, For width, For real numbers; use dimensions of [size] along the horizontal and vertical directions respectively. and One-dimensional global average pooling kernel paired feature map Each channel is encoded independently, and the vertical feature tensor is calculated. and horizontal feature tensor , of which Each channel is at a height Output at and width Output at The expressions are as follows: , , in, Indicates the first The first channel in the vertical direction The output feature components of each pixel location after one-dimensional global average pooling; Indicates the first The first channel in the width direction The output feature components after pooling at each pixel location; and These represent the input feature map in the corresponding channel and spatial coordinates, respectively. Pixel values ​​at the location Pixel value at location; horizontal coordinate variable and vertical coordinate variables These are the spatial indices for the integral summation; For channel indexing; Step S1-2, Cross-dimensional feature reduction and fusion: and Perform transpose and concatenation operations along the spatial dimension to obtain the fused feature map. Input the fused feature map to The convolutional layer performs channel dimensionality reduction with a compression ratio of r, followed by a batch normalization layer and a non-linear activation function. Generate intermediate feature maps containing cross-dimensional interaction information. The calculation formula is: , in, This represents an intermediate feature tensor that contains cross-dimensional interaction information; This represents the feature tensor obtained by concatenating the feature vectors in the vertical and horizontal directions along the spatial dimensions. express Convolution operation; This indicates a batch normalization operation; Represents a nonlinear activation function; Steps S1-3, Spatial orientation decoupling and attention activation: The intermediate feature map... Re-separate along the spatial dimension, decouple into dimensionality-reduced features in the vertical direction. Dimensionality reduction features in the horizontal direction ; Utilizing independent Convolutional layers respectively and Perform channel dimensionality upgrade to restore the original number of channels. And respectively mapped to the Sigmoid activation function. Interval, obtain the vertical attention weight tensor With horizontal attention weight tensor ; Steps S1-4, orthogonal aggregation of multi-channel spatial heatmaps: To extract a single spatial distribution suitable for frequency domain prior guidance, attention weight tensors are used in the vertical direction. With horizontal attention weight tensor The process performs channel-level orthogonal outer product operations and mean pooling aggregation along the channel dimensions to generate a two-dimensional spatial target prior heatmap. Space target prior heatmap G spatial coordinates Pixel response value at The expression is: , in, This represents the generated two-dimensional spatial target prior heatmap without channel dimensions in physical space coordinates. The scalar response value at that point, i.e., the aggregated pixel value; and They represent the first The vertical coordinates of each channel Attention weight value and horizontal coordinate at the location Attention weight values ​​at the location; the two-dimensional spatial target prior heatmap It aggregates the cross probability distributions of all feature channels on the vertical and horizontal coordinates to explicitly extract and lock the sub-pixel center geometry of infrared point targets. Steps S1-5, Feature Space Attention Weighting: Utilizing the vertical attention weight tensor With horizontal attention weight tensor For the input original feature map Perform channel-wise and pixel-wise multiplicative transform modulation to output attention-enhanced multi-scale spatial feature maps. Multi-scale spatial feature map In the passage Spatial coordinates Mathematical expression at the location : , in, This indicates that the attention-enhanced multi-scale spatial feature map is at the 1st... Each channel, spatial coordinates Precise pixel-level feature values ​​at the location; This represents the initial pixel-level feature response value of the original input feature map at the corresponding position; the multi-scale spatial feature map E serves as the input spatial feature flow for subsequent frequency domain analysis and feature matching.

[0011] In step S2, the spatially guided frequency domain joint gated filtering specifically includes the following steps: Step S2a-1, Dual-path spatial frequency forward mapping: Obtain the multi-scale spatial feature map output from step S1. and two-dimensional spatial target prior heat map ; Multi-scale spatial feature maps along the spatial dimension Prior heatmaps of each channel and space target Performing a two-dimensional discrete Fourier transform maps the multi-scale spatial feature map E and the prior heatmap of spatial targets G from the spatial domain to the complex frequency domain, obtaining the feature spectrum tensors respectively. and heat map spectrum In frequency domain coordinates The positive mapping expression at is: , , in, The imaginary unit; It is the characteristic spectrum tensor X in the frequency domain coordinates Positive mapping at the location, It is a heat map spectrum In frequency domain coordinates A positive mapping at; e is the natural constant; Step S2a-2, Prior extraction of frequency domain structure: Extracting the spectrum from the heatmap Extract the response amplitude information from low to high frequencies in space to construct a frequency domain structure prior that serves as a spatial geometric constraint. : , in and These represent the real and imaginary parts of a complex number, respectively. Step S2a-3, Spatially Guided Joint Gated Modulation: Construction and Characteristic Spectral Tensor Complex weight tensors of uniform size that are globally learnable during network backpropagation ; Prioritize the frequency domain structure After cross-channel broadcasting, element-wise joint gating modulation is performed on the characteristic spectrum tensor X to obtain the enhanced modulation spectrum tensor. ; Step S2a-4, Inverse reconstruction of space-frequency cooperative features: For the modulation spectrum tensor Perform a two-dimensional discrete Fourier inverse transform, extract the real part, and reconstruct the space-frequency co-location feature map mapped back to the spatial domain. .

[0012] In step S2a-3, the following calculation formula is used: , in, The enhanced spectral tensor output after joint gate modulation is represented in the frequency domain coordinates. Complex components at the location; These are the complex components of the original feature spectrum obtained by forward mapping; This represents the dynamic weight components of the globally learnable complex weight matrix at the corresponding channel and frequency domain positions; Representing the spatial prior heatmap in the frequency domain coordinates The structural prior amplitude at the location; It represents the Hadamardi (or Hadama) stack; An adaptive balance coefficient is used to control the spatial prior intervention intensity. This modulation process uses the coordinate response amplitude captured in the spatial domain. As a structural prior, the gain of the corresponding infrared micro-target frequency band component in the learnable weights is forcibly amplified.

[0013] In step S2a-4, the following calculation formula is used: , in, This indicates that the space-frequency cooperative feature map reconstructed back to the spatial domain through inverse transformation is in the th... Each channel, spatial coordinates Pixel-level feature values ​​at the location; This represents the mathematical operation of extracting the real part of a complex number. These are the modulated characteristic spectral components; The complex exponential basis functions are two-dimensional discrete Fourier inverse transforms; the space-frequency cooperative feature map It achieves global frequency domain suppression of background clutter with extremely low signal-to-noise ratio in infrared thermal crossover scenarios, while preserving the fine spatial edges of small targets.

[0014] In step S2, the high-frequency energy proportion factor is calculated using the following method: Step S2b-1, Spectrum Zero-Frequency Shift and Centering: Obtain the Enhanced Modulation Spectrum Tensor To correctly separate high and low frequency components, the modulation spectrum tensor... Each channel performs a two-dimensional zero-frequency shift operation in the spatial frequency plane, translating the zero-frequency DC component at the origin to the geometric center of the frequency spectrum plane. At this point, the centered spectral tensor is obtained. ; Step S2b-2, High-frequency energy boundary threshold division: based on the frequency domain center point Using the origin as the reference point, and based on the preset low-pass cutoff frequency threshold... In the two-dimensional frequency domain plane, a set of high-frequency region masks is delineated to represent the sharp edge response of the target. The set of high-frequency regions The elements satisfy the following two-dimensional Euclidean distance condition: , in, These are the two-dimensional pixel coordinates in the frequency domain after zero-frequency shift;

[0015] Step S2b-3, Cross-channel three-dimensional spectral energy aggregation: Based on centralized spectral tensor Calculate in the high frequency region respectively Sum of three-dimensional spectral energy across all channels And the total spectral energy across the entire frequency band. The spectral energy value is calculated from the square of the complex amplitude: , , in, Represents the mask in the high-frequency region. Within the range, the sum of the integrals of the squares of the complex magnitudes of the centered spectral tensors spanning all C channels is the high-frequency three-dimensional spectral energy of the target; It represents the sum of the three-dimensional spectral energy across all channels throughout the entire frequency band; Represents the coordinates of the centered spectral tensor in the c-th channel after shifting. Complex frequency response at; This indicates the calculation of the square of the magnitude of a complex number; Step S2b-4, Quantification of thermal crossover environmental factors: Based on the sum of the three-dimensional spectral energy, the high-frequency energy proportion factor is calculated. : , In infrared physics detection scenarios, when a target and background experience "thermal crossover," the target's thermal radiation edge becomes blurred and diffused, leading to a sharp attenuation of its corresponding high-frequency spatial structure characteristics; the high-frequency energy proportion factor... The positive correlation between the edge sharpness of the UAV target in the current input frame and the thermal cross signal-to-noise ratio attenuation is used as a scene environment parameter to provide real-time quantized state priors for the dynamic penalty loss function in subsequent steps.

[0016] Step S3 includes: Step S3-1, Basic distance metric for bounding boxes: Obtain the predicted bounding boxes output by the regression branch of the object detection prediction head and the ground truth bounding boxes; calculate the intersection-over-union (IoU) ratio between the predicted and ground truth bounding boxes, and convert the geometric distance error between the predicted and ground truth bounding boxes into the basic regression loss using existing techniques. The specific formula for calculating the basic regression loss is as follows: Simultaneously, the basic outlier metric parameters of the Wise-IoU mechanism are extracted. The basic outlier metric parameter Defined as the ratio of the basic regression loss of the current bounding box to the historical momentum moving average loss of the basic regression loss of all predicted bounding boxes in the current batch, it is used to initially characterize the relative regression quality of the current predicted bounding box; Step S3-2, Cross-domain penalty of spectral energy and outlier reconstruction: This addresses the fundamental outlier metric parameters caused by thermal dispersion at the target edge in infrared scenes. The abnormally amplified technical defect will affect the high-frequency energy ratio factor. Introduced as a cross-domain feedback parameter into the outlier calculation, the thermal cross-sensing outlier is reconstructed. : , in, The parameter is adjusted exponentially; when severe thermal crossover occurs in the infrared detection scene, causing attenuation of the target's high-frequency features, When it decreases, the exponential penalty term Synchronous scaling down adaptively forces the suppression of outlier artifacts caused by blurred target boundaries; Step S3-3, Generate non-monotonic dynamic focusing coefficients: This involves reconstructing the thermal cross-sensing outlier coefficients. Substituting into the non-monotonic focusing distribution function, the non-monotonic focusing coefficient based on spectral energy sensing is calculated. : , in, To control the shape hyperparameter of the peak position of the focus distribution curve; To control the shape hyperparameter of the peak width of the focused distribution curve; The non-monotonic focusing coefficient This ensures that the network prioritizes the allocation of optimization gradients to ordinary quality anchor boxes with moderate outliers in extremely low signal-to-noise ratio environments, thus preventing the network's backpropagation direction from being dominated by severely distorted extremely low quality anchor boxes. Step S3-4, Joint loss weighting and network iterative optimization: utilizing non-monotonic focusing coefficients Basic regression loss Dynamic weighting is applied to construct the final spectral energy-sensing regression loss. The spectrum energy perception regression loss and classification prediction loss are combined as the overall optimization objective of the multi-task prediction head. The backpropagation algorithm is executed to calculate the gradient of the parameters of each layer of the network and update the weights. The converged network model is optimized iteratively, and the precise positioning coordinates and continuous tracking trajectory of the infrared UAV target are output during the inference stage.

[0017] The present invention also provides an infrared UAV tracking system based on space-frequency cooperation and spectrum focusing for implementing the method, the system being applied to a twin network architecture, comprising: The multi-scale spatial feature and target coordinate prior extraction module is used to acquire template frames and search frames from the infrared video stream, and synchronously input the template frames and search frames into a backbone feature extraction network with shared weights to extract hierarchical original features and input them into a bidirectional weighted feature pyramid. In the lateral connection stage of the bidirectional weighted feature pyramid, the embedded coordinate attention mechanism is used to sequentially perform feature concatenation, channel dimensionality reduction, directional decoupling, and channel dimensionality increase processing on the horizontal and vertical pooled encoded features to obtain the horizontal and vertical attention weight tensors, respectively. Channel-level outer product and channel aggregation processing are performed on the horizontal and vertical attention weight tensors to generate a two-dimensional spatial target prior heatmap for characterizing the spatial position of infrared micro-targets. At the same time, the horizontal and vertical attention weight tensors are used to perform channel-wise and pixel-wise multiplicative modulation on the original features to output the attention-enhanced multi-scale spatial feature map. The spatially guided frequency-domain joint gating filtering and energy assessment module is used to synchronously map the multi-scale spatial feature map and the spatial target prior heat map to the complex frequency domain space through two-dimensional discrete Fourier transform; extract the spectral amplitude of the heat map to construct the frequency domain structure prior, and use the prior to perform spatially guided joint gating modulation on the globally learnable complex weight matrix in the frequency domain gating network; subsequently, extract the real part through two-dimensional discrete Fourier transform to reconstruct the space-frequency cooperative feature map; in addition, the spatially guided frequency-domain joint gating filtering and energy assessment module is also used to perform two-dimensional zero-frequency shift on the modulated feature spectrum, and calculate the ratio of the three-dimensional spectral energy of the high-frequency region across all feature channels to the total spectral energy of the entire frequency band, outputting the high-frequency energy proportion factor as a quantization parameter of the thermal crossover environment; The non-monotonic dynamic focusing regression and prediction module for heat-cross sensing is used to input the space-frequency collaborative feature map into the multi-task prediction head, extract the basic intersection-union ratio loss and basic outlier parameters of the predicted bounding box and the ground truth bounding box; during the backpropagation stage of network training, the high-frequency energy proportion factor is introduced as a cross-domain feedback parameter into the outlier calculation, an exponential penalty term is constructed to reconstruct the heat-cross sensing outlier, and the heat-cross sensing outlier is substituted into the non-monotonic focusing distribution function to generate dynamic focusing coefficients; finally, the dynamic focusing coefficients are used to adaptively weight the basic regression loss of the bounding box to forcibly suppress the low-quality anchor box gradient interference in severe heat-crossing and edge blurring scenarios, complete the iterative optimization of network parameters, and output the accurate UAV position and tracking trajectory during the inference stage.

[0018] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0019] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0020] This invention innovatively proposes a fundamental mathematical architecture based on space-frequency cooperative sensing and dynamic spectral energy focusing (SE-WIoU): (1) Spatial-frequency cooperative sensing and adaptive gating filtering mechanism guided by spatial prior knowledge Decoupling and Locking of Sub-pixel-Level Spatial Coordinates: In the lateral connectivity of the bidirectional weighted feature pyramid, this method utilizes one-dimensional global average pooling to perform direction-aware encoding of features in the horizontal and vertical directions. Unlike conventional attention mechanisms that only output one-dimensional weights, this invention uniquely performs channel-level orthogonal outer product operations and mean aggregation between the horizontal and vertical attention weights, generating a two-dimensional spatial target prior heatmap without channel dimensions. This heatmap explicitly and uniquely locks the sub-pixel-level geometric center of the infrared micro-target in the spatial domain in a physically meaningful way.

[0021] Forced physical navigation with frequency domain gating: When faced with low signal-to-noise ratio environments caused by thermal crossover, traditional frequency domain global filtering often falls into blind search due to a lack of target location information, mistakenly retaining high-frequency background clutter as the target. This scheme directly maps the aforementioned spatial target prior heatmap into a "frequency domain structure prior" through discrete Fourier transform and deeply embeds it into the calculation formula of frequency domain gating.

[0022] The essence of the above formula is to dynamically and adaptively excite and amplify the gain of the learnable weight matrix in the frequency domain on the corresponding target frequency band by utilizing the target coordinate response amplitude captured in the spatial domain. This enables the filter not only to globally perceive the high and low frequency distribution, but also to be forced to perform high-frequency edge enhancement around the physical coordinates of the real target, completely solving the target feature dispersion problem under severe thermal crossover backgrounds.

[0023] (2) Non-monotonic dynamic focusing regression mechanism of cross-domain environmental feedback spectrum energy (SE-WIoU) To address the technical challenge of trajectory drift in traditional target tracking networks under harsh infrared conditions, this invention innovatively proposes a non-monotonic loss function architecture that directly controls spatial regression using frequency domain energy.

[0024] Quantitative Extraction of Thermal Crossing Environmental Factor: This invention points out that when severe thermal crossover occurs between an infrared target and the background, its physical manifestation is the blurring and dispersion of the target's thermal radiation edges, which directly maps to a sharp attenuation of high-frequency structural features in the frequency domain. Therefore, this scheme strictly performs two-dimensional zero-frequency shift after frequency domain modulation and accurately calculates the ratio of the three-dimensional spectral energy across all channels in the high-frequency region to the total energy of the entire frequency band, extracting the "high-frequency energy ratio factor." This factor quantifies the edge sharpness and thermal crossover severity of the current input frame in real time.

[0025] Cross-domain exponential penalty for outliers: Existing Wise-IoU loss functions rely solely on the geometric intersection-union ratio of spatial predicted boxes to calculate outliers. When target boundaries become extremely blurred due to hot crossovers, this purely spatial distance metric becomes completely ineffective, leading to amplified outliers in low-quality anchor boxes and causing severe gradient oscillations. This proposal uses the calculated high-frequency energy proportion factor as the cross-domain environment feedback state, constructs an exponential penalty term, and multiplies it into the calculation of the basic outlier.

[0026] The brilliance of this cross-domain interaction lies in the fact that when severe hot crossover occurs in the tracking scene, the target's high-frequency energy drops sharply, and the exponential penalty term adaptively shrinks accordingly, thereby forcibly suppressing and shielding outliers caused by blurred boundaries. It guides the network to remain calm under extremely low signal-to-noise ratios, smoothly distributing the optimization gradient to ordinary anchor boxes of medium quality, avoiding being skewed in the backpropagation direction by distorted, extremely low-quality anchor boxes, and ensuring the anti-drift and high robustness of the tracking trajectory from the lowest level of the loss function mechanism.

[0027] The present invention has the following beneficial effects: (1) High accuracy and high success rate: On the highly challenging Anti-UAV410 large-scale anti-drone benchmark dataset, the accuracy and success rate of this invention reached 70.5% and 60.35% respectively, which is a significant improvement of 6.15% compared with the benchmark SiamDT algorithm. In the field test data, the accuracy rate is as high as 90.35%.

[0028] (2) Completely solve the trajectory drift problem: The spectrum energy feedback mechanism enables the system to have environmental adaptability. When the target enters the hot cross blind zone, the network can rely on cross-domain priors to maintain high confidence center locking, avoiding scale collapse and trajectory divergence in traditional methods under low signal-to-noise ratio. Attached Figure Description

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0030] Figure 1 This is a schematic diagram of the overall architecture of the infrared drone tracking method of the present invention.

[0031] Figure 2 This is a schematic diagram of the multi-scale spatial feature and target coordinate prior extraction module of the present invention.

[0032] Figure 3 This is a schematic diagram of the frequency domain joint gating filter and high-frequency energy assessment process guided by the spatial prior of this invention.

[0033] Figure 4This is a comparison chart of the tracking performance of the method of the present invention and existing representative tracking algorithms in a thermal cross-infrared scenario. Detailed Implementation

[0034] This invention provides an infrared UAV tracking method and system based on space-frequency coordination and spectrum focusing. Some concepts involved in this invention are explained below:

[0035] Thermal crossover phenomenon: This refers to the phenomenon in infrared photoelectric detection scenarios where, when the thermal radiation temperature of a UAV target approaches the same as the temperature of the background environment (such as clouds or buildings), the contrast of the target in the infrared image is drastically reduced. In the frequency domain, this phenomenon manifests as severe dispersion of the target's high-frequency spatial structure features (i.e., sharp edges).

[0036] Prior heat map of space targets ( This invention features a unique intermediate feature representation. It is a two-dimensional spatial matrix generated by extracting horizontal and vertical attention weights through a coordinate attention mechanism and performing "orthogonal outer product operation" and "mean pooling aggregation". Its physical meaning is to explicitly lock and calibrate the geometric center coordinates of infrared point targets at the sub-pixel level.

[0037] Space-frequency cooperative feature map ( This refers to the process of using the spectral amplitude of the aforementioned "space target prior heatmap" as physical navigation within the frequency domain feature space. This involves dynamically modulating frequency-domain gated learnable weights and reconstructing the feature map back to the spatial domain using a two-dimensional discrete Fourier transform. This feature map successfully suppresses global thermal clutter at extremely low signal-to-noise ratios and directionally preserves the high-frequency edges of the target.

[0038] High-frequency energy proportion factor ( The environmental quantification parameters constructed in this invention are as follows: After centering the zero frequency by performing a two-dimensional zero-frequency shift on the frequency domain features, the ratio of the three-dimensional spectral energy across all feature channels in the high-frequency region to the total energy of the entire frequency band is calculated. This factor (between 0 and 1) is positively correlated with the severity of thermal crosstalk in the current scene and the clarity of the target edge.

[0039] The non-monotonic focusing coefficient of spectral energy sensing ( In the network regression branch, a cross-domain exponential penalty term is constructed using the aforementioned "high-frequency energy proportion factor." This term adaptively corrects the amplified base outlier caused by boundary ambiguity, resulting in a novel gradient weighting coefficient. This coefficient can forcibly suppress the interference of low-quality anchor frames on network backpropagation in severely hot crossover scenarios.

[0040] To ensure that the mathematical formulas and parameters described in this method have excellent computational convergence and generalization ability, the hyperparameters in the formula can preferentially adopt the following recommended value ranges in general implementation scenarios: (1) In the cross-dimensional feature dimensionality reduction step, the dimensionality reduction compression ratio The typical value is or To achieve a balance between feature preservation and computational overhead; (2) In the frequency domain modulation process, adaptive balance coefficients The typical value is to (3) When dividing the high-frequency energy mask, the low-pass cutoff frequency threshold is used to control the intensity of the a priori intervention frequency domain of the spatial heat map; The value is typically taken as the width of the input feature map scale. arrive For example, for common local feature map sizes, the possible values ​​are... (4) When calculating the outlier degree of heat cross-sensing, the exponential smoothing adjustment parameter used to control the feedback intensity. The typical value is or (5) When generating non-monotonic dynamic focusing coefficients, the shape hyperparameter The typical value is Shape hyperparameters The typical value is The aforementioned shape parameters ensure that the non-monotonic curve exhibits the optimal gradient gain distribution for moderately difficult samples.

[0041] The method mainly consists of three progressive stages: (1) First stage: Multi-scale spatial features and sub-pixel center locking Acquire continuously transmitted video frames from the infrared thermal imager. Initialize the initial frame as the template frame and extract the current frame to be detected as the search frame. Simultaneously input the two frames of data into the backbone network to obtain hierarchical features, and then feed them into the bidirectional weighted feature pyramid.

[0042] At this stage, we embed an improved coordinate attention mechanism at the lateral connections of the bidirectional weighted feature pyramid. The network performs global average pooling along both height and width, and then... Convolutional dimensionality reduction, activation, and dimensionality upscaling, outputting a weight vector with decoupled orientation. and Instead of directly using one-dimensional weights, cross-channel weighting is performed. Orthogonal outer product and channel compression are used to generate a pure two-dimensional spatial target prior heatmap. This step is like providing a precise "spatial anchor" for subsequent frequency domain operations amidst a sea of ​​clutter. Simultaneously, the feature flow, after attention weighting, outputs a multi-scale spatial feature map. .

[0043] (2) Second stage: frequency domain forward mapping and spatial prior-guided gated filtering In complex air-ground contexts, direct convolution in the airspace can easily lead to the dispersion of minute features. The system then processes the feature map... Each channel and heat map Perform two-dimensional fast Fourier transforms on each component, mapping them to the complex frequency domain to obtain the characteristic spectrum. and prior spectrum .

[0044] Calculate the amplitude of the prior spectrum Substituting into the core modulation formula of this invention: , This step completely changes the blindness of traditional learnable frequency domain gating, using spatial coordinate amplitude to forcibly increase the filtering weight of the frequency band where the real target is located. Then, a two-dimensional discrete Fourier inverse transform is performed and the real part is extracted to reconstruct a pure space-frequency cooperative feature map.

[0045] Simultaneously, the system performs a strict two-dimensional zero-frequency shift operation on the modulated characteristic spectrum to separate the low- and high-frequency boundaries and calculate the high-frequency energy integral across the channels. Energy integral across the entire frequency band The high-frequency energy proportion factor of the environmental quantitative state parameter is obtained. .

[0046] (3) Third stage: Dynamic regression of cross-domain penalty of spectral energy The reconstructed high-quality features are fed into the multi-task prediction head. During bounding box regression training, the target edges become extremely blurry due to the hot crossover phenomenon. Traditional Wise-IoU (Intelligent Cross-Union Ratio) outlier calculations based on distance and IoU are rendered ineffective. It will cause severe distortion and amplification.

[0047] The system directly calls the high-frequency energy ratio factor extracted in real time during the second stage. Construct an exponential penalty term to calculate a novel heat-crossing-aware outlier: When encountering severe heat crossover, A sharp decrease, with the index penalty item forcibly suppressing the abnormal. To prevent gradient avalanche, non-monotonic focusing coefficients are generated based on the corrected outlier, smoothly updating network parameters. In actual inference deployment, this mechanism ensures that the tracking box can still firmly grasp the target center even in low signal-to-noise ratio and adverse weather conditions, stably outputting the drone trajectory.

[0048] To verify the substantial progress of the implementation scheme of this invention in solving practical engineering pain points, a multi-dimensional evaluation was carried out based on the anti-drone benchmark dataset (Anti-UAV410).

[0049] Experimental data shows that, due to the precise navigation of frequency domain filtering by spatial priors and the cross-domain feedback control of frequency domain energy on spatial regression loss achieved by this invention at the underlying architecture, the robustness of the model to low, slow, and small targets is greatly improved. In test sequences containing extreme hot crossovers and complex cloud camouflage, the tracking accuracy of this invention reached 70.5%. Most importantly, because this mechanism completely eliminates the phenomenon of tracking box scale collapse under low signal-to-noise ratio, the success rate of this invention achieves an absolute increase of up to 6.15% compared to traditional benchmark algorithms. This groundbreaking quantitative data fully demonstrates that this technical solution not only possesses extremely high theoretical innovation but also has immediate application value in the practical deployment of anti-drone systems.

[0050] To fully demonstrate that the proposed space-frequency coordination mechanism and spectral energy feedback mechanism have outstanding substantive features and significant progress, the applicant conducted rigorous ablation experiments on the Anti-UAV410 complex infrared dataset (as shown in Table 1).

[0051] Table 1 Ablation Experiment Results of the Improved Module

[0052] (1) Control experiment setup (see Table 1 for details): Control Group A (Validation of Spatial Prior Locking Effectiveness): A BiFPN network with embedded coordinate attention was introduced on top of the baseline model. For extremely small infrared targets smaller than 5 pixels, traditional unidirectional feature pyramids are prone to feature diffusion due to the lack of bottom-up texture supplementation. Configuration A successfully locked the sub-pixel level geometric center of the tiny target by using the CA mechanism to decouple and generate a "two-dimensional spatial target prior heatmap".

[0053] Control Group B (Verification of Space-Frequency Co-modulation Effect): Based on Configuration A, the traditional standard spatial domain self-attention module is abandoned, and the space-frequency cooperative modulation unique to this invention is used instead. The prior heatmap of spatial targets extracted by Configuration A is mapped to the frequency domain and used as a structural prior to guide frequency domain filtering.

[0054] The complete set of this invention (cross-domain penalty mechanism verification): Based on configuration B, the traditional SmoothL1 or standard WIoU regression loss is replaced with the thermal cross-sensing spectral energy dynamic focusing loss (SE-WIoU) proposed in this invention.

[0055] (2) Comparison of key parameters and results:

[0056] Overcoming feature diffusion, success rate leaps for the first time: As shown in Table 1, after introducing Swin-CA-BiFPN, the model success rate improved significantly by 3.65% compared to the baseline. This strongly demonstrates that multi-channel orthogonal converged heatmaps play a decisive role in maintaining the spatial topology of infrared micro-targets and mitigating downsampled feature diffusion.

[0057] Overcoming the thermal crossover blind zone, a significant breakthrough in accuracy was achieved: After introducing the FDCA module guided by spatial priors, the accuracy was directly improved by 0.95%. This verifies the synergistic effect of the present invention, namely, using spatial coordinate amplitude to forcibly increase the weight of the real target frequency band, which can more thoroughly suppress low-frequency background thermal clutter and prevent the target from merging with the background compared to pure spatial attention.

[0058] Non-monotonic dynamic focusing leads to extreme convergence capability: Ultimately, after applying the SE-WIoU mechanism, the model success rate reached 60.35%, a total increase of 6.15% compared to the baseline model. This fully demonstrates the scientific validity of this invention's use of the "high-frequency energy proportion factor" to penalize outliers across domains. When faced with annotation box blurring and noise caused by severe hot crossover, this mechanism successfully suppressed the gradient distortion of low-quality anchor boxes, exhibiting robustness and trajectory smoothing capabilities unattainable by traditional loss functions, achieving unexpectedly outstanding technical results.

[0059] The data and structures in the accompanying drawings are explained in detail.

[0060] like Figure 1 As shown, the overall processing flow of the infrared UAV tracking method of this invention includes a template branch and a search branch. First, template frames and search frames are acquired from the infrared video stream, and then input into a backbone feature extraction network with shared weights to extract raw features at different levels. Subsequently, the raw features are input into a coordinate attention-based bidirectional weighted feature pyramid for multi-scale feature fusion, and spatial-frequency collaborative features are obtained through frequency domain joint gating. Candidate proposal boxes are generated based on these spatial-frequency collaborative features, and candidate region features are extracted through region-of-interest alignment. Then, a matching and bounding box prediction head is used to classify and regress the candidate proposal boxes, selecting those with higher scores. Simultaneously, a dynamic geometric regression branch further refines the position and scale parameters of the candidate boxes, ultimately outputting the predicted bounding box of the infrared UAV target.

[0061] like Figure 2As shown, the multi-scale spatial feature and target coordinate prior extraction module includes a backbone feature extraction network, a bidirectional weighted feature pyramid, and a coordinate attention module. The backbone feature extraction network outputs multiple levels of raw features and performs cross-scale fusion through the top-down and bottom-up paths of the bidirectional weighted feature pyramid. At the lateral connection positions of the bidirectional weighted feature pyramid, the input features are averaged along the horizontal and vertical directions respectively to obtain horizontal pooled encoded features and vertical pooled encoded features. Subsequently, feature concatenation, two-dimensional convolution, batch normalization, and nonlinear activation are performed sequentially, and the processed intermediate features are split along the spatial direction to obtain horizontal and vertical dimensionality-reduced features. Channel upscaling and nonlinear mapping are performed on the horizontal and vertical dimensionality-reduced features respectively to obtain horizontal and vertical attention weight tensors. The original features are multiplicatively modulated channel-wise and pixel-wise using attention weight tensors in two directions to output attention-enhanced multi-scale spatial features. At the same time, channel-level outer product and channel aggregation are performed on the attention weight tensors in two directions to generate a two-dimensional spatial target prior heatmap.

[0062] like Figure 3 As shown, the spatial prior-guided frequency domain joint gated filtering and high-frequency energy assessment process includes spatial domain feature input, frequency domain mapping, frequency domain gated modulation, spatial domain feature reconstruction, and high-frequency energy assessment. First, a two-dimensional discrete Fourier transform is performed on the multi-scale spatial feature map and the two-dimensional spatial target prior heatmap to obtain the feature spectrum tensor and the heatmap spectrum, respectively. Then, the amplitude information of the heatmap spectrum is extracted to construct a frequency domain structure prior, and the learnable complex weight tensor is modulated using this frequency domain structure prior. The modulated weights are then subjected to element-wise operations with the feature spectrum tensor to obtain the enhanced modulated spectrum tensor. A two-dimensional inverse discrete Fourier transform is performed on the modulated spectrum tensor, and the real part is extracted to reconstruct the spatial-frequency co-factor feature map. Simultaneously, a zero-frequency shift is performed on the modulated spectrum tensor, and the high-frequency region spectral energy and the total spectral energy of the entire frequency band are calculated. The high-frequency energy proportion factor is obtained based on the ratio of the two.

[0063] like Figure 4As shown, in infrared video sequences with target occlusion, background thermal clutter, and thermal crossover, the method of this invention can maintain continuous localization of infrared UAV targets. As the temperature difference between the target and the background decreases, the edge response of the target in the infrared image gradually weakens. Some existing tracking methods exhibit predicted bounding box scale changes, positional shifts, or target loss. In contrast, the method of this invention utilizes spatial target priors to guide frequency domain gated filtering and dynamically corrects the bounding box regression loss based on a high-frequency energy proportion factor, thereby reducing the impact of background thermal clutter and edge blurring on the target localization results. The actual bounding boxes in the figure represent the actual position of the UAV target, while the other bounding boxes with different labels represent the prediction results of various comparative tracking methods and the method of this invention. In the figure, Algorithm 1 is the KCFCF algorithm, Algorithm 2 is the KCFD algorithm, Algorithm 3 is the STCGIF algorithm, Algorithm 4 is the SiamRPN algorithm, Algorithm 5 is the SiamCAR algorithm, and Algorithm 6 is the SiamDT algorithm. The English explanations are as follows: KCFCF: Kernelized Correlation Filter with Curvature Filter; KCFD: Kernelized Correlation Filter with Detection; STCGIF: Spatio-Temporal Context and Guide Image Filter; SiamRPN: Siamese Region Proposal Network. SiamCAR: Siamese Fully Convolutional Classification and Regression. SiamDT: Siamese Detection and Tracking.

[0064] This invention provides an infrared UAV tracking method and system based on space-frequency coordination and spectrum focusing. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. An infrared UAV tracking method based on space-frequency coordination and spectrum focusing, characterized in that, The method is applied to a twin network architecture and includes the following steps: Step S1, Multi-scale spatial feature and target coordinate prior extraction: Acquire infrared video stream, extract the initial frame as template frame, and extract the current frame to be detected as search frame; input the template frame and search frame synchronously into the backbone feature extraction network with shared weights to obtain hierarchical original features; then, input the original features into a bidirectional weighted feature pyramid for multi-scale feature fusion. In the lateral connection stage of feature fusion, the embedded coordinate attention mechanism is used to decouple and encode the sub-pixel level coordinate responses in the horizontal and vertical directions to generate a two-dimensional spatial target prior heatmap and output the fused multi-scale spatial feature map; Step S2, Space-guided frequency-domain joint gating filtering and energy assessment: Two-dimensional discrete Fourier transforms are performed on the channel data of the multi-scale spatial feature map and the prior heat map of the spatial target, respectively, to map the spatial signal to the frequency domain; in the frequency domain feature space, the spectral response of the prior heat map of the spatial target is used as a prior structural constraint to dynamically and adaptively modulate the learnable complex weight matrix in the frequency domain gating module, and space-prior-guided joint gating filtering is performed; subsequently, a space-frequency cooperative feature map that suppresses background thermal clutter and preserves the high-frequency edge of the target is reconstructed through two-dimensional discrete Fourier inverse transform, and the high-frequency energy proportion factor of the current filtered feature is extracted simultaneously; Step S3, Non-monotonic Dynamic Focusing Regression and Position Prediction Based on Thermal Cross-Sensing: The spatial-frequency collaborative feature map is input into a multi-task prediction head containing classification and regression branches; during the network parameter training and update phase of the regression branch, based on the high-frequency energy proportion factor, the basic outlier parameter in the bounding box regression loss function is adaptively penalized and corrected to construct a non-monotonic focusing coefficient based on spectral energy sensing; the non-monotonic focusing coefficient is used to dynamically suppress the gradient of low-quality anchor boxes to complete network optimization, and the final predicted bounding box and motion trajectory of the infrared UAV target are output during the inference phase.

2. The method according to claim 1, characterized in that, Step S1 includes: Step S1-1, Direction-aware pooling encoding: Obtain the input feature map at the lateral connections of the bidirectional weighted feature pyramid. ,in For the number of channels, For height, For width, For real numbers; use dimensions of [size] along the horizontal and vertical directions respectively. and One-dimensional global average pooling kernel paired feature map Each channel is encoded independently, and the vertical feature tensor is calculated. and horizontal feature tensor , of which Each channel is at a height Output at and width Output at The expressions are as follows: ， ， in, Indicates the first The first channel in the vertical direction The output feature components of each pixel location after one-dimensional global average pooling; Indicates the first The first channel in the width direction The output feature components after pooling at each pixel location; and These represent the input feature map in the corresponding channel and spatial coordinates, respectively. Pixel values ​​at the location Pixel value at location; horizontal coordinate variable and vertical coordinate variables These are the spatial indices for the integral summation; For channel indexing; Step S1-2, Cross-dimensional feature reduction and fusion: and Perform transpose and concatenation operations along the spatial dimension to obtain the fused feature map. Input the fused feature map to The convolutional layer performs channel dimensionality reduction with a compression ratio of r, followed by a batch normalization layer and a non-linear activation function. Generate intermediate feature maps containing cross-dimensional interaction information. The calculation formula is: ， in, This represents an intermediate feature tensor that contains cross-dimensional interaction information; This represents the feature tensor obtained by concatenating the feature vectors in the vertical and horizontal directions along the spatial dimensions. express Convolution operation; This indicates a batch normalization operation; Represents a nonlinear activation function; Steps S1-3, Spatial orientation decoupling and attention activation: The intermediate feature map... Re-separate along the spatial dimension, decouple into dimensionality-reduced features in the vertical direction. Dimensionality reduction features in the horizontal direction ; Utilizing independent Convolutional layers respectively and Perform channel dimensionality upgrade to restore the original number of channels. And respectively mapped to the Sigmoid activation function. Interval, obtain the vertical attention weight tensor With horizontal attention weight tensor ; Steps S1-4, orthogonal aggregation of multi-channel spatial heatmaps: attention weight tensor in the vertical direction With horizontal attention weight tensor The process performs channel-level orthogonal outer product operations and mean pooling aggregation along the channel dimensions to generate a two-dimensional spatial target prior heatmap. Space target prior heatmap G spatial coordinates Pixel response value at The expression is: ， in, This represents the generated two-dimensional spatial target prior heatmap without channel dimensions in physical space coordinates. scalar response value at; and They represent the first The vertical coordinates of each channel Attention weight value and horizontal coordinate at the location Attention weight values ​​at the location; the two-dimensional spatial target prior heatmap It aggregates the cross probability distributions of all feature channels on the vertical and horizontal coordinates to explicitly extract and lock the sub-pixel center geometry of infrared point targets. Steps S1-5, Feature Space Attention Weighting: Utilizing the vertical attention weight tensor With horizontal attention weight tensor For the input original feature map Perform channel-wise and pixel-wise multiplicative transform modulation to output attention-enhanced multi-scale spatial feature maps. Multi-scale spatial feature map In the passage Spatial coordinates Mathematical expression at the location : ， in, This indicates that the attention-enhanced multi-scale spatial feature map is at the 1st... Each channel, spatial coordinates Precise pixel-level feature values ​​at the location; This represents the initial pixel-level feature response value of the original input feature map at the corresponding position.

3. The method according to claim 2, characterized in that, In step S2, the spatially guided frequency domain joint gated filtering specifically includes the following steps: Step S2a-1, Dual-path spatial frequency forward mapping: Obtain the multi-scale spatial feature map output from step S1. and two-dimensional spatial target prior heat map ; Multi-scale spatial feature maps along the spatial dimension Prior heatmaps of each channel and space target Performing a two-dimensional discrete Fourier transform maps the multi-scale spatial feature map E and the prior heatmap of spatial targets G from the spatial domain to the complex frequency domain, obtaining the feature spectrum tensors respectively. and heat map spectrum In frequency domain coordinates The positive mapping expression at is: ， ， in, The imaginary unit; It is the characteristic spectrum tensor X in the frequency domain coordinates Positive mapping at the location, It is a heat map spectrum In frequency domain coordinates A positive mapping at; e is the natural constant; Step S2a-2, Prior extraction of frequency domain structure: Extracting the spectrum from the heatmap Extract the response amplitude information from low to high frequencies in space to construct a frequency domain structure prior that serves as a spatial geometric constraint. : ， in and These represent the real and imaginary parts of a complex number, respectively. Step S2a-3, Spatially Guided Joint Gated Modulation: Construction and Characteristic Spectral Tensor Complex weight tensors of uniform size that are globally learnable during network backpropagation ; Prioritize the frequency domain structure After cross-channel broadcasting, element-wise joint gating modulation is performed on the characteristic spectrum tensor X to obtain the enhanced modulation spectrum tensor. ; Step S2a-4, Inverse reconstruction of space-frequency cooperative features: For the modulation spectrum tensor Perform a two-dimensional discrete Fourier inverse transform, extract the real part, and reconstruct the space-frequency co-location feature map mapped back to the spatial domain. .

4. The tracking method according to claim 3, characterized in that, In step S2a-3, the following calculation formula is used: ， in, The enhanced spectral tensor output after joint gate modulation is represented in the frequency domain coordinates. Complex components at the location; These are the complex components of the original feature spectrum obtained by forward mapping; This represents the dynamic weight components of the globally learnable complex weight matrix at the corresponding channel and frequency domain positions; Representing the spatial prior heatmap in the frequency domain coordinates The structural prior amplitude at the location; It represents the Hadamardi (or Hadama) stack; An adaptive balance coefficient for controlling the spatial prior intervention intensity.

5. The tracking method according to claim 4, characterized in that, In step S2a-4, the following calculation formula is used: ， in, This indicates that the space-frequency cooperative feature map reconstructed back to the spatial domain through inverse transformation is in the th... Each channel, spatial coordinates Pixel-level feature values ​​at the location; This represents the mathematical operation of extracting the real part of a complex number. These are the modulated characteristic spectral components; It is a complex exponential basis function of the two-dimensional discrete Fourier inverse transform.

6. The tracking method according to claim 5, characterized in that, In step S2, the high-frequency energy proportion factor is calculated using the following method: Step S2b-1, Spectrum Zero-Frequency Shift and Centering: Obtain the Enhanced Modulation Spectrum Tensor ; for modulation spectrum tensor Each channel performs a two-dimensional zero-frequency shift operation in the spatial frequency plane, translating the zero-frequency DC component at the origin to the geometric center of the frequency spectrum plane. At this point, the centered spectral tensor is obtained. ; Step S2b-2, High-frequency energy boundary threshold division: based on the frequency domain center point Using the origin as the reference point, and based on the preset low-pass cutoff frequency threshold... In the two-dimensional frequency domain plane, a set of high-frequency region masks is delineated to represent the sharp edge response of the target. The set of high-frequency regions The elements satisfy the following two-dimensional Euclidean distance condition: ， in, These are the two-dimensional pixel coordinates in the frequency domain after zero-frequency shift; Step S2b-3, Cross-channel three-dimensional spectral energy aggregation: Based on centralized spectral tensor Calculate in the high frequency region respectively Sum of three-dimensional spectral energy across all channels And the total spectral energy across the entire frequency band. The spectral energy value is calculated from the square of the complex amplitude: ， ， in, Represents the mask in the high-frequency region. Within the range, the sum of the integrals of the squares of the complex magnitudes of the centered spectral tensors spanning all C channels; It represents the sum of the three-dimensional spectral energy across all channels throughout the entire frequency band; Represents the coordinates of the centered spectral tensor in the c-th channel after shifting. Complex frequency response at; This indicates the calculation of the square of the magnitude of a complex number; Step S2b-4, Quantification of thermal crossover environmental factors: Based on the sum of the three-dimensional spectral energy, the high-frequency energy proportion factor is calculated. : 。 7. The method according to claim 6, characterized in that, Step S3 includes: Step S3-1, Basic distance metric for bounding boxes: Obtain the predicted bounding boxes output by the regression branch of the object detection prediction head and the ground truth bounding boxes; calculate the intersection-union ratio (IoU) between the predicted and ground truth bounding boxes, and convert the geometric distance error between the predicted and ground truth bounding boxes into the basic regression loss using existing techniques. The specific formula for calculating the basic regression loss is as follows: Simultaneously, the basic outlier metric parameters of the intelligent intersection-union comparison mechanism are extracted. The basic outlier metric parameter Defined as the ratio of the basic regression loss of the current bounding box to the historical momentum moving average loss of the basic regression loss of all predicted bounding boxes in the current batch, it is used to initially characterize the relative regression quality of the current predicted bounding box; Step S3-2, Cross-domain penalty and outlier reconstruction of spectral energy: The high-frequency energy proportion factor is... Introduced as a cross-domain feedback parameter into the outlier calculation, the thermal cross-sensing outlier is reconstructed. : ， in, This is the parameter for exponential smoothing adjustment; Step S3-3, Generate non-monotonic dynamic focusing coefficients: This involves reconstructing the thermal cross-sensing outlier coefficients. Substituting into the non-monotonic focusing distribution function, the non-monotonic focusing coefficient based on spectral energy sensing is calculated. : ， in, To control the shape hyperparameter of the peak position of the focus distribution curve; To control the shape hyperparameter of the peak width of the focused distribution curve; Step S3-4, Joint loss weighting and network iterative optimization: utilizing non-monotonic focusing coefficients Basic regression loss Dynamic weighting is applied to construct the final spectral energy-sensing regression loss. The spectrum energy perception regression loss and classification prediction loss are combined as the overall optimization objective of the multi-task prediction head. The backpropagation algorithm is executed to calculate the gradient of the parameters of each layer of the network and update the weights. The converged network model is optimized iteratively, and the precise positioning coordinates and continuous tracking trajectory of the infrared UAV target are output during the inference stage.

8. An infrared unmanned aerial vehicle (UAV) tracking system based on space-frequency coordination and spectral focusing for implementing the method of any one of claims 1 to 7, characterized in that, The system is applied to a twin network architecture and includes: The multi-scale spatial feature and target coordinate prior extraction module is used to acquire template frames and search frames from the infrared video stream, and synchronously input the template frames and search frames into a backbone feature extraction network with shared weights to extract hierarchical original features and input them into a bidirectional weighted feature pyramid. In the lateral connection stage of the bidirectional weighted feature pyramid, the embedded coordinate attention mechanism is used to sequentially perform feature concatenation, channel dimensionality reduction, directional decoupling, and channel dimensionality increase processing on the horizontal and vertical pooled encoded features to obtain the horizontal and vertical attention weight tensors, respectively. Channel-level outer product and channel aggregation processing are performed on the horizontal and vertical attention weight tensors to generate a two-dimensional spatial target prior heatmap for characterizing the spatial position of infrared micro-targets. At the same time, the horizontal and vertical attention weight tensors are used to perform channel-wise and pixel-wise multiplicative modulation on the original features to output the attention-enhanced multi-scale spatial feature map. The space-guided frequency-domain joint gated filtering and energy assessment module is used to synchronously map the multi-scale spatial feature map and the spatial target prior heat map to the complex frequency domain space through two-dimensional discrete Fourier transform; extract the spectral amplitude of the heat map to construct the frequency domain structure prior, and use the frequency domain structure prior to perform space-guided joint gated modulation on the globally learnable complex weight matrix in the frequency domain gated network; subsequently, the real part is extracted through two-dimensional discrete Fourier transform to reconstruct the space-frequency cooperative feature map; in addition, the space-guided frequency-domain joint gated filtering and energy assessment module is also used to perform two-dimensional zero-frequency shift on the modulated feature spectrum, and calculate the ratio of the three-dimensional spectral energy of the high-frequency region across all feature channels to the total spectral energy of the entire frequency band, outputting the high-frequency energy proportion factor as a quantization parameter of the thermal crossover environment; The non-monotonic dynamic focusing regression and prediction module for thermal cross-sensing is used to input the space-frequency collaborative feature map into the multi-task prediction head, extract the basic intersection-union ratio loss and basic outlier parameters of the predicted bounding box and the ground truth bounding box; during the backpropagation stage of network training, the high-frequency energy proportion factor is introduced as a cross-domain feedback parameter into the outlier calculation, an exponential penalty term is constructed to reconstruct the thermal cross-sensing outlier, and the thermal cross-sensing outlier is substituted into the non-monotonic focusing distribution function to generate dynamic focusing coefficients; finally, the basic regression loss of the bounding box is adaptively weighted using the dynamic focusing coefficients to complete the iterative optimization of network parameters and output the accurate UAV position and tracking trajectory during the inference stage.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 7.

Images