Multi-spectral target tracking method based on weight preprocessing and token updating

Abstract

The application discloses a multispectral target tracking method based on weight preprocessing and token updating. A first frame multispectral template image and a frame multispectral search image are determined; spatial channel attention calculation is performed on and to obtain a spatial channel attention weighted search image; patch embedding operation is performed on and to obtain an un-updated template token and an un-updated search token after flattening and splicing; calculation is performed on and to obtain an updated spectral token and a Transformer encoding layer token; interaction is performed on and to obtain an updated search token; addition is performed on and to obtain a second-updated search token; and finally, a prediction head is inputted to obtain a tracking result of the frame. t ​

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a multispectral target tracking method based on weight preprocessing and token update. Background Technology

[0002] Unmanned aerial vehicles (UAVs) have become an important remote sensing platform for monitoring dynamic targets on the Earth's surface, supporting a wide range of applications such as environmental monitoring, precision agriculture, and security monitoring. Thanks to their high mobility and flexible flight deployment, UAVs can acquire high spatiotemporal resolution images and continuously observe moving targets in complex environments.

[0003] Most existing UAV-based tracking methods are developed for RGB images, relying primarily on appearance similarity to locate targets across frames. However, aerial tracking often faces significant challenges, including drastic viewpoint changes, scale variations, motion blur, and cluttered backgrounds, which frequently lead to tracking drift or failure, thus limiting the development of UAV-based target tracking. These limitations are largely attributed to the limited spectral representation of RGB images, which provides limited discriminative information for distinguishing targets from visually similar backgrounds. Multispectral images, by capturing rich spectral reflectance across multiple bands, enable greater discriminative differentiation of object materials and features. Mainstream tracking methods can be broadly categorized into two types: methods that primarily rely on handcrafted features and integrated correlation filters, and deep learning-based methods.

[0004] However, the complex motion patterns of targets and the hidden high-level semantic information are difficult to represent with handcrafted features, which limits the performance of trackers in complex and dynamic contexts.

[0005] Due to the successful application of deep learning in target tracking tasks, mainstream methods have now shifted towards deep learning-based approaches. Trans-DAT, which uses a Transformer architecture to automatically extract the spectral-spatial features of targets through end-to-end learning, has received widespread attention. SENSE uses a Siamese network framework and improves tracking robustness by fusing material and motion cues. TBR-Net uses deep convolutional neural networks, along with transform-based frequency band reorganization and feature refinement networks, to improve model performance in complex dynamic environments. Compared to the limited feature representation capabilities of handcrafted features, using a Transformer architecture or deep convolutional neural networks to extract multispectral features and integrating feature enhancement modules can effectively improve the ability to handle complex nonlinear patterns.

[0006] Overall, current multispectral video tracking still faces two key challenges: first, the attention weights for the multispectral search region are not updated in a timely manner, making it difficult to fully utilize the spatial and spectral information of the target; second, the global context information and historical frame information of the target are not fully utilized, which limits its robustness in tracking in complex dynamic environments. Summary of the Invention

[0007] Therefore, the main objective of this invention is to provide a multispectral target tracking method based on weight preprocessing and token update.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows: This invention provides a multispectral target tracking method based on weight preprocessing and token updating. The method is as follows: Step 1: Load the multispectral image sequence and determine the template image of the first frame of the multispectral image sequence. and the Search image for frame multispectral images ,in, This indicates the number of frames in the input multispectral image sequence. It is an integer greater than or equal to 1; Step 2, the above and Spatial attention is calculated to obtain spatial attention. , will the and Perform channel attention calculation to obtain the channel attention. ,use and right Weight preprocessing is performed to obtain a spatial channel weighted search image. ; Step 3, the above Perform the Patchembedding operation and flatten and stitch together to obtain the unupdated template token. , will the Perform the Patchembedding operation and flatten the stitches to obtain the unupdated search token. ; Step 4: Place the above The updated template token is obtained through calculation. , will the and Perform Transformer encoding to obtain the Transformer search token. , will the The original spectral cue token is obtained through calculation. ; Step 5: Place the above Calculations are performed to obtain an updated spectral cue token. ; Step Six: [The sentence is incomplete and requires more context to be translated accurately.] and Perform multi-head attention interaction to obtain an updated search token. ; Step 7: Place the aforementioned and The sum is the search token updated for the second time. ; Step 8: Place the aforementioned The prediction head is fed in to obtain the prediction box position, completing the first step. Target tracking of multispectral images involves loading each frame of the multispectral image sequence sequentially.

[0009] Optionally, step two is implemented through the following steps: (201) According to the following formula, for the... Spatial attention is calculated. : (1) in, This represents a 1×1 convolution operation. This represents a 3×3 convolution operation; (202) According to the following formula, for the... Channel attention is calculated. : (2) (3) (4) in, This indicates the average pooling operation. This represents the template image features after average pooling, where, Represents the query matrix. Represents the key matrix, Represents a value matrix, Indicates the number of spectral channels. This indicates the transpose operation. This indicates a single-head self-attention operation. This represents the activation function. This indicates a normalization operation; (203) According to the following formula, for the... Weight preprocessing is performed to obtain a spatial channel weighted search image. : (5) in, This indicates a cosmetic procedure. This indicates an element-wise multiplication operation.

[0010] Optionally, step three specifically includes: [the following steps are described in the original text] Perform a Patchembedding operation, dividing the template into 16×16 pixel patches and flattening them into one-dimensional vectors. Then, concatenate the patches to obtain the unupdated template token. Regarding the above Perform a Patchembedding operation to divide the data into 16×16 pixel patches, flatten them into one-dimensional vectors, and then concatenate them to obtain the unupdated search token. .

[0011] Optionally, step four is implemented through the following steps: (401) According to the following formula, for The updated template token is obtained by performing a global selective context update. : (6) (7) in, This indicates that a global selective context update is being performed. The state equation is... The coefficient matrix represents the hidden state equation. For the hidden state equation, This represents the coefficient matrix of the input tokens. To enter a token, To output the equation, The coefficient matrix representing the state equation; (402) According to the following formula, for and Perform Transformer encoding to obtain the Transformer search token. : (8) in, This indicates Transformer encoding processing; (403) According to the following formula, for Perform global average pooling to obtain the original spectral cue token. : (9) in, This indicates that a global average pooling operation is being performed. Optionally, step five is implemented through the following steps: (501) According to the following formula, using and Calculate the intermediate token : (10) in, This indicates a compression operation. This indicates the introduction of an incentive operation. express GeLU Activation function; (502) According to the following formula, using Calculations are performed to obtain updated spectral cue tokens. : (11) in, This represents a multi-layered sensing mechanism.

[0012] Optionally, step six is ​​implemented through the following steps: (601) According to the following formula, for and Calculations are performed to obtain the first Group multi-head attention : (12) in, It is an integer greater than or equal to 1; (602) According to the following formula, for Group Perform a concatenation operation and multiply it with the transformation matrix to obtain the updated search region token. : (13) in, This indicates a serial operation. This represents the transformation matrix.

[0013] Optionally, step seven is implemented through the following steps: (14) in, This indicates an element-wise addition operation.

[0014] Optionally, step eight specifically includes: […]. Input the prediction head, output a classification score map, a center offset map, and a scale change map. Determine the target center position based on the position with the highest classification score, and refine the target center position based on the center offset map and the scale change map, thus completing the first step. Target tracking of multispectral images involves loading each frame of the multispectral image sequence sequentially to complete target tracking of the entire multispectral image sequence.

[0015] Compared with existing technologies, this invention proposes a multispectral target tracking method based on weight preprocessing and token updating. In the weight preprocessing stage, the weights are adaptively updated by calculating attention weights in both spatial and channel dimensions. Spatial attention and channel attention suppress the interference of spatial background and enhance the weights of important channels, thereby strengthening the spectral differences between the target and the background. Then, in the token updating stage, a Mamba layer is introduced, which uses its adaptive parameter adjustment mechanism to selectively compress global context information and continuously update the feature representation in the time dimension, thereby effectively modeling long-term dependencies and improving feature representation capabilities. The spectral cue layer continuously updates the spectral cue tokens using historical frame information, thereby refining the spectral information. The token updating stage enables the model to achieve dynamic updating and fine modeling of spectral information while maintaining computational efficiency. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention; Figure 2 In an embodiment of the present invention, a visualization template region image of the first frame of the OV-Truck-001 multispectral image sequence; Figure 3 In an embodiment of the present invention, a visualization image of the weighted preprocessing of the 48th frame OV-Truck-001 multispectral image sequence is shown. Figure 4 This is the tracking result of the 64th frame of the OV-Truck-001 multispectral image sequence in this embodiment of the invention; Figure 5 This is the tracking result of the OV-Truck-001 multispectral image sequence in the 137th frame of this invention; Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] This invention provides a multispectral target tracking method based on weight preprocessing and token updating, such as... Figure 1-5 As shown, the method is as follows: Step 1: Load the multispectral image sequence and determine the template image of the first frame of the multispectral image sequence. and the Search image for frame multispectral images ,in, This indicates the number of frames in the input multispectral image sequence. It is an integer greater than or equal to 1; Specifically, this embodiment uses the MUST dataset as the validation dataset. This is the first large-scale multispectral tracking benchmark specifically designed for UAV platforms. The dataset contains a total of 43,000 frames, corresponding to approximately 8,600 seconds of video duration. It includes 160 sequences for the training set and 90 sequences for the test set. The OV-Truck-001 image sequence is selected. The OV-Truck-001 image sequence contains 157 frames, with each OV-Truck-001 multispectral image including 8 channels. The MUST dataset includes frame-level bounding box annotations and is labeled using 12 challenging attributes: background clutter (BC), camera motion (CM), fast motion (FM), complete occlusion (FOC), illumination variation (IV), low resolution (LR), motion blur (MB), out-of-line-of-sight (OV), partial occlusion (POC), similar colors (SC), similar object background (SOB), and scale variation (SV). These attributes comprehensively reflect the real-world challenges faced by UAV target tracking. A template image for the first frame of the multispectral image is determined using pre-labeled ground truth values. and search area image . Figure 2 This is a visualization template region image of the first frame of the OV-Truck-001 multispectral image sequence. Figure 3 The image is a weighted preprocessing visualization of the 48th frame of the OV-Truck-001 multispectral image sequence, where the template image has a dimension of 192×192×8 pixels and the search region image has a dimension of 384×384×8 pixels.

[0019] Step 2, the above and Spatial attention is calculated to obtain spatial attention. , will the and Perform channel attention calculation to obtain the channel attention. ,use and right Weight preprocessing is performed to obtain a spatial channel weighted search image. ; Specifically, (201) according to the following formula, the... Spatial attention is calculated. : (1) in, This indicates an average operation where the pixel values ​​at corresponding spatial locations are added together. This represents a 1×1 convolution operation. This represents a 3×3 convolution operation; According to equation (1), the template image is first reduced using a 1×1 convolution. We increase the channel dimension to compress template feature information and improve computational efficiency. Then, we pass it through two 3×3 convolutional layers to further learn contextual information. Finally, we use a 1×1 convolution again to output a spatial attention map. Its dimensions are 384×384×8. Its dimensions are 384×384×1.

[0020] (202) According to the following formula, for the... Channel attention is calculated. : (2) (3) (4) in, This indicates the average pooling operation. This represents the template image features after average pooling, where, Represents the query matrix. Represents the key matrix, Represents a value matrix, Indicates the number of spectral channels. This indicates the transpose operation. This indicates a single-head self-attention operation. This represents the activation function. This indicates a normalization operation; According to equation (2), the spatial resolution of the template image is first adjusted by average pooling. Then, according to equation (3), the spatial resolution is adjusted using... The similarity between channels is calculated. Then, according to equation (4), the spatial resolution of the template image is further adjusted through average pooling. Finally, channel attention is generated through normalization. The query matrix... The dimensions are 384×384×8, key matrix The dimensions are 384×384×8, and the value matrix is... Its dimensions are 384×384×8. The dimensions are 1×1×8.

[0021] (203) According to the following formula, for the... Weight preprocessing is performed to obtain a spatial channel weighted search image. : (5) in, This indicates a cosmetic procedure. This represents the element-wise multiplication operation; According to equation (5), the shaping operation will transform the channel attention of dimension 1×1×8. Weighted spatial attention in a dimension of 384×384×1 Each pixel. This ensures that the final attention dimension conforms to... The dimension of the element-wise multiplication is determined. Finally, the spatially channel-weighted search image, preprocessed with target space and channel attention weights, is obtained through element-wise multiplication. Weight preprocessing enhances the difference between the target and the background, improving robustness in tracking under complex dynamic environments. Among these, Its dimensions are 384×384×8, and The dimensions remain consistent at 384×384×8.

[0022] Step 3, the above Perform the Patchembedding operation and flatten and stitch together to obtain the unupdated template token. , will the Perform the Patchembedding operation and flatten the stitches to obtain the unupdated search token. ; Specifically, regarding the aforementioned Perform a Patchembedding operation, dividing the template into 16×16 pixel patches and flattening them into one-dimensional vectors. Then, concatenate the patches to obtain the unupdated template token. Regarding the above Perform a Patchembedding operation to divide the data into 16×16 pixel patches, flatten them into one-dimensional vectors, and then concatenate them to obtain the unupdated search token. ; For a template image of 192×192×8 pixels Search image with spatial channel attention weighting of 384×384×8 pixels Perform the Patchembedding operation separately to obtain the updated template token. and outdated search tokens The patch size for Patchembedding is 16×16 pixels, therefore and The data is divided into 16×16 pixel patches. Each patch is then linearly mapped and embedded into a high-dimensional feature space, resulting in a final feature space with dimensions of 512×144×1. and dimensions of 512×576×1 .

[0023] Step 4: Place the above The updated template token is obtained through calculation. , will the and Perform Transformer encoding to obtain the Transformer search token. , will the The original spectral cue token is obtained through calculation. ; Specifically, (401) according to the following formula, for The updated template token is obtained by performing a global selective context update. : (6) (7) in, This indicates that a global selective context update is being performed. The state equation is... The coefficient matrix represents the hidden state equation. For the hidden state equation, This represents the coefficient matrix of the input tokens. To enter a token, To output the equation, The coefficient matrix representing the state equation; According to Equation (6), the Mamba layer utilizes its adaptive parameter adjustment mechanism to selectively compress global context information and continuously update the feature representation over time, thereby effectively modeling long-term dependencies and improving feature representation capabilities. According to Equation (7), the SSM is an important component of the Mamba layer that selectively focuses on historical hidden states, primarily through selective updates using a dynamic coefficient matrix. Input The coefficient matrix is ​​determined using the projection function. , and Therefore, the coefficient matrix is ​​not constant, but changes with the input. Dynamically updated. Among them... Its dimensions are 512×144×1, and The dimensions are kept consistent at 512×144×1.

[0024] (402) According to the following formula, for and Perform Transformer encoding to obtain the Transformer search token. : (8) in, This indicates Transformer encoding processing; According to Equation (8), Transformer encoding is a prior art technique. In this invention, ViT-B / 16 is used as the backbone network, with a 12-layer Transformer encoder structure and 12 attention heads. ViT-B / 16 can extract contextual information from the fused token, thereby improving the feature representation effect of the token.

[0025] (403) According to the following formula, The original spectral hint token is obtained through a global average pooling operation. : (9) in, This indicates that global average pooling is performed along the channel dimension; The dimension is 1×1×512.

[0026] Step 5: Place the above Calculations are performed to obtain an updated spectral cue token. ; Specifically, (501) according to the following formula, using and Calculate the intermediate token : (10) in, This indicates a compression operation. This indicates the introduction of an incentive operation. express GeLU Activation function; According to formula (10), in subsequent tokens In the calculation, this invention uses the original spectral cue token. and current template token By combining these elements, a band-based recalibration comparison and fusion is performed. This involves inputting historically stable spectral priors and the spectral features of the current template, ultimately outputting a set of adaptive spectral weights for updating. This suppresses drift, refines spectral information, and amplifies the discriminative band weights. (First layer) The compression operation primarily aims to reduce some less important spectral information, thereby enabling more efficient extraction of key spectral information. The second layer... It is responsible for introducing incentives to introduce nonlinear factors, enabling the network to fit complex functions and achieve effective differentiation and decision-making in the spectral dimension.

[0027] (502) According to the following formula, using Calculations are performed to obtain updated spectral cue tokens. : (11) in, This represents a multi-layered sensing mechanism; Multi-layer sensing mechanisms are existing technologies. Among them, Its dimensions are 1×1×512, and The dimensions are kept consistent at 1×1×512.

[0028] Step Six: [The sentence is incomplete and requires more context to be translated accurately.] and Perform multi-head attention interaction to obtain an updated search token. ; Specifically, (601) according to the following formula, for and Calculations are performed to obtain the first Group multi-head attention : (12) in, It is an integer greater than or equal to 1; According to equation (12), multi-head attention is calculated in order to... The updated spectral information for the target is passed to the search token. To enhance Robustness of target tracking over long time series. Among them, the query matrix... The dimensions are 384×384×8, key matrix The dimensions are 384×384×8, and the value matrix is... Its dimensions are 384×384×8.

[0029] (602) According to the following formula, for Group Perform a concatenation operation and multiply it with the transformation matrix to obtain the updated search region token. : (13) in, This indicates a serial operation. Represents the transformation matrix; According to equation (13), the chain operation is to calculate the obtained The results of the multi-head attention group are concatenated along the projection direction. Then, the concatenated results are transformed using a transformation matrix. The dimension of the multi-head attention results is adjusted to the token dimension. Among them, The dimensions of are 512×576×1, and The dimensions are kept consistent at 512×576×1.

[0030] Step 7: Place the aforementioned and The sum is the search token updated for the second time. ; Specifically, it is achieved through the following formula. and The sum is the search token updated for the second time. : (14) in, This indicates an element-wise addition operation.

[0031] The dimensions of are 512×576×1, and The dimensions are kept consistent at 512×576×1.

[0032] Step 8: Place the aforementioned The prediction head is fed in to obtain the prediction box position, completing the first step. Target tracking of multispectral images involves loading each frame of the multispectral image sequence sequentially to complete target tracking of the entire multispectral image sequence.

[0033] Specifically, Input the prediction head, output a classification score map, a center offset map, and a scale change map. Determine the target center position based on the position with the highest classification score, and refine the target center position based on the center offset map and the scale change map, thus completing the first step. Target tracking of multispectral images involves loading each frame of the multispectral image sequence sequentially to complete target tracking of the entire multispectral image sequence.

[0034] This invention employs a center-based prediction head to predict and generate three score maps over a search region. Specifically, the prediction head generates a classification score map, a center offset map, and a scale change map. The classification score map encodes the probability that the target center is located at each spatial location. The center offset map provides sub-pixel refinement for the predicted center coordinates, while the scale change map re-estimates the normalized adjustments to the target width and height to account for scaling variations. All values ​​in the center offset map and the scale change map are normalized to the range [0, 1]. During inference, the spatial location with the highest score in the classification score map is selected as the predicted target center, and the center location is refined using the corresponding offset from the center offset map. The final bounding box is then obtained by adjusting the base size according to the prediction scaling factor in the scale change map.

[0035] Figure 4The tracking results of the 64th frame of the OV-Truck-001 multispectral image sequence in this embodiment of the invention demonstrate that the method can still stably track the target under the challenge of multiple similar targets. Figure 5 In this embodiment of the invention, the tracking results of the OV-Truck-001 multispectral image sequence in frame 137 demonstrate that the method can still stably track the target even when the target loses its field of view.

[0036] This invention first designs a spatial channel attention network that adaptively updates attention in both spatial and channel dimensions. By suppressing spatial background interference and enhancing the weights of important channels through the attention network, it strengthens the spectral differences between the target and the background. Second, a token update network is designed to eliminate the need for additional feature enhancement modules. By introducing a Mamba layer, its adaptive parameter adjustment mechanism is used to selectively compress global context information and continuously update feature representations in the time dimension, thereby effectively modeling long-term dependencies and improving feature representation capabilities. The spectral cue layer continuously updates spectral cue tokens using historical frame information, thereby refining spectral information. The token update network enables the model to achieve dynamic updating and refined modeling of spectral information while maintaining computational efficiency.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A multispectral target tracking method based on weight preprocessing and token update, characterized in that, The method is as follows: Step 1: Load the multispectral image sequence and determine the template image of the first frame of the multispectral image sequence. and the Search image for frame multispectral images ,in, This indicates the number of frames in the input multispectral image sequence. It is an integer greater than or equal to 1; Step 2, the above and Spatial attention is calculated to obtain spatial attention. , will the and Perform channel attention calculation to obtain the channel attention. ,use and right Weight preprocessing is performed to obtain a spatial channel weighted search image. ; Step 3, the above Perform the Patchembedding operation and flatten and stitch together to obtain the unupdated template token. , will the Perform the Patchembedding operation and flatten the stitches to obtain the unupdated search token. ; Step 4: Place the above The updated template token is obtained through calculation. , will the and Perform Transformer encoding to obtain the Transformer search token. , will the The original spectral cue token is obtained through calculation. ; Step 5: Place the above Calculations are performed to obtain the updated spectral cue token. ; Step Six: [The sentence is incomplete and requires more context to be translated accurately.] and Perform multi-head attention interaction to obtain an updated search token. ; Step 7: Place the aforementioned and The sum is the search token updated for the second time. ; Step 8: Place the aforementioned The prediction head is fed in to obtain the prediction box position, completing the first step. Target tracking in multispectral images.

2. The multispectral target tracking method based on weight preprocessing and token update according to claim 1, characterized in that, Step two is implemented through the following steps: (201) According to the following formula, for the... Spatial attention is calculated. : (1) in, This represents a 1×1 convolution operation. This represents a 3×3 convolution operation; (202) According to the following formula, for the... Channel attention is calculated. : (2) (3) (4) in, This indicates the average pooling operation. This represents the template image features after average pooling, where, Represents the query matrix. Represents the key matrix, Represents a value matrix, Indicates the number of spectral channels. This indicates the transpose operation. This indicates a single-head self-attention operation. This represents the activation function. This indicates a normalization operation; (203) According to the following formula, for the... Weight preprocessing is performed to obtain a spatial channel weighted search image. : (5) in, This indicates a cosmetic procedure. This indicates an element-wise multiplication operation.

3. The multispectral target tracking method based on weight preprocessing and token update according to claim 1 or 2, characterized in that, Step three specifically involves: [The sentence is incomplete and requires more context to be translated accurately.] Perform a Patchembedding operation, dividing the template into 16×16 pixel patches and flattening them into one-dimensional vectors. Then, concatenate the patches to obtain the unupdated template token. Regarding the above Perform a Patchembedding operation to divide the data into 16×16 pixel patches, flatten them into one-dimensional vectors, and then concatenate them to obtain the unupdated search token. .

4. The multispectral target tracking method based on weight preprocessing and token update according to claim 3, characterized in that, Step four is implemented through the following steps: (401) According to the following formula, for The updated template token is obtained by performing a global selective context update. : (6) (7) in, This indicates that a global selective context update is being performed. The state equation is... The coefficient matrix represents the hidden state equation. For the hidden state equation, This represents the coefficient matrix of the input tokens. To enter a token, To output the equation, The coefficient matrix representing the state equation; (402) According to the following formula, for and Perform Transformer encoding to obtain the Transformer search token. : (8) in, This indicates Transformer encoding processing; (403) According to the following formula, for Perform global average pooling to obtain the original spectral cue token. : (9) in, This indicates that a global average pooling operation is being performed.

5. The multispectral target tracking method based on weight preprocessing and token update according to claim 4, characterized in that, Step five is implemented through the following steps: (501) According to the following formula, using and Calculate the intermediate token : (10) in, This indicates a compression operation. This indicates the introduction of an incentive operation. express GeLU Activation function; (502) According to the following formula, using Calculations are performed to obtain updated spectral cue tokens. : (11) in, This represents a multi-layered sensing mechanism.

6. The multispectral target tracking method based on weight preprocessing and token update according to claim 5, characterized in that, Step six is ​​implemented through the following steps: (601) According to the following formula, for and Calculations are performed to obtain the first Group multi-head attention : (12) in, It is an integer greater than or equal to 1; (602) According to the following formula, for Group Perform a concatenation operation and multiply it with the transformation matrix to obtain the updated search region token. : (13) in, Indicates a serial operation. This represents the transformation matrix.

7. The multispectral target tracking method based on weight preprocessing and token update according to claim 6, characterized in that, Step seven is specifically implemented using the following formula. and The sum is the search token updated for the second time. : (14) in, This indicates an element-wise addition operation.

8. The multispectral target tracking method based on weight preprocessing and token update according to claim 7, characterized in that, Step eight specifically involves: Input the prediction head, output a classification score map, a center offset map, and a scale change map. Determine the target center position based on the position with the highest classification score, and refine the target center position based on the center offset map and the scale change map, thus completing the first step. Target tracking of multispectral images involves loading each frame of the multispectral image sequence sequentially to complete target tracking of the entire multispectral image sequence.

Images